├── utils ├── __init__.py └── image_util.py ├── pororo ├── tasks │ ├── utils │ │ ├── __init__.py │ │ ├── tokenizer.py │ │ ├── config.py │ │ ├── base.py │ │ └── download_utils.py │ ├── __init__.py │ └── optical_character_recognition.py ├── __version__.py ├── models │ └── brainOCR │ │ ├── modules │ │ ├── __init__.py │ │ ├── sequence_modeling.py │ │ ├── basenet.py │ │ ├── prediction.py │ │ ├── transformation.py │ │ └── feature_extraction.py │ │ ├── __init__.py │ │ ├── _dataset.py │ │ ├── imgproc.py │ │ ├── detection.py │ │ ├── craft.py │ │ ├── model.py │ │ ├── recognition.py │ │ ├── brainocr.py │ │ ├── craft_utils.py │ │ ├── _modules.py │ │ └── utils.py ├── __init__.py ├── utils.py └── pororo.py ├── assets └── images │ ├── test_image_1.jpg │ ├── test_image_2.jpg │ └── test_image_3.jpg ├── requirements.txt ├── README.md ├── main.py └── LICENSE /utils/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /pororo/tasks/utils/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /pororo/__version__.py: -------------------------------------------------------------------------------- 1 | version = "0.4.1" 2 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/__init__.py: -------------------------------------------------------------------------------- 1 | from .brainocr import Reader # noqa 2 | -------------------------------------------------------------------------------- /assets/images/test_image_1.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yunwoong7/korean_ocr_using_pororo/HEAD/assets/images/test_image_1.jpg -------------------------------------------------------------------------------- /assets/images/test_image_2.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yunwoong7/korean_ocr_using_pororo/HEAD/assets/images/test_image_2.jpg -------------------------------------------------------------------------------- /assets/images/test_image_3.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/yunwoong7/korean_ocr_using_pororo/HEAD/assets/images/test_image_3.jpg -------------------------------------------------------------------------------- /pororo/__init__.py: -------------------------------------------------------------------------------- 1 | from pororo.__version__ import version as __version__ # noqa 2 | from pororo.pororo import Pororo # noqa 3 | -------------------------------------------------------------------------------- /pororo/tasks/__init__.py: -------------------------------------------------------------------------------- 1 | # flake8: noqa 2 | """ 3 | __init__.py for import child .py files 4 | 5 | isort:skip_file 6 | """ 7 | 8 | # Utility classes & functions 9 | import pororo.tasks.utils 10 | from pororo.tasks.utils.download_utils import download_or_load 11 | from pororo.tasks.utils.base import ( 12 | PororoBiencoderBase, 13 | PororoFactoryBase, 14 | PororoGenerationBase, 15 | PororoSimpleBase, 16 | PororoTaskGenerationBase, 17 | ) 18 | 19 | # Factory classes 20 | from pororo.tasks.optical_character_recognition import PororoOcrFactory 21 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/sequence_modeling.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | 4 | class BidirectionalLSTM(nn.Module): 5 | 6 | def __init__(self, input_size: int, hidden_size: int, output_size: int): 7 | super(BidirectionalLSTM, self).__init__() 8 | self.rnn = nn.LSTM(input_size, 9 | hidden_size, 10 | bidirectional=True, 11 | batch_first=True) 12 | self.linear = nn.Linear(hidden_size * 2, output_size) 13 | 14 | def forward(self, x): 15 | """ 16 | x : visual feature [batch_size x T=24 x input_size=512] 17 | output : contextual feature [batch_size x T x output_size] 18 | """ 19 | self.rnn.flatten_parameters() 20 | recurrent, _ = self.rnn( 21 | x 22 | ) # batch_size x T x input_size -> batch_size x T x (2*hidden_size) 23 | output = self.linear(recurrent) # batch_size x T x output_size 24 | return output 25 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | appnope==0.1.3 2 | backcall==0.2.0 3 | certifi==2022.12.7 4 | charset-normalizer==3.0.1 5 | cycler==0.11.0 6 | debugpy==1.6.6 7 | decorator==5.1.1 8 | entrypoints==0.4 9 | fonttools==4.38.0 10 | idna==3.4 11 | imageio==2.25.0 12 | ipykernel==6.16.2 13 | ipython==7.34.0 14 | jedi==0.18.2 15 | jupyter_client==7.4.9 16 | jupyter_core==4.12.0 17 | kiwisolver==1.4.4 18 | matplotlib==3.5.3 19 | matplotlib-inline==0.1.6 20 | nest-asyncio==1.5.6 21 | networkx==2.6.3 22 | numpy==1.21.6 23 | opencv-python==4.7.0.68 24 | packaging==23.0 25 | parso==0.8.3 26 | pexpect==4.8.0 27 | pickleshare==0.7.5 28 | Pillow==9.4.0 29 | pip==22.3.1 30 | prompt-toolkit==3.0.36 31 | psutil==5.9.4 32 | ptyprocess==0.7.0 33 | Pygments==2.14.0 34 | pyparsing==3.0.9 35 | python-dateutil==2.8.2 36 | PyWavelets==1.3.0 37 | pyzmq==25.0.0 38 | requests==2.28.2 39 | scikit-image==0.19.3 40 | scipy==1.7.3 41 | setuptools==65.6.3 42 | six==1.16.0 43 | tifffile==2021.11.2 44 | torch==1.13.1 45 | torchvision==0.14.1 46 | tornado==6.2 47 | traitlets==5.9.0 48 | typing_extensions==4.4.0 49 | urllib3==1.26.14 50 | wcwidth==0.2.6 51 | wget==3.2 52 | wheel==0.37.1 53 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/_dataset.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | from natsort import natsorted 4 | from PIL import Image 5 | from torch.utils.data import Dataset 6 | 7 | 8 | class RawDataset(Dataset): 9 | 10 | def __init__(self, root, imgW, imgH): 11 | self.imgW = imgW 12 | self.imgH = imgH 13 | self.image_path_list = [] 14 | for dirpath, _, filenames in os.walk(root): 15 | for name in filenames: 16 | _, ext = os.path.splitext(name) 17 | ext = ext.lower() 18 | if ext in (".jpg", ".jpeg", ".png"): 19 | self.image_path_list.append(os.path.join(dirpath, name)) 20 | 21 | self.image_path_list = natsorted(self.image_path_list) 22 | self.nSamples = len(self.image_path_list) 23 | 24 | def __len__(self): 25 | return self.nSamples 26 | 27 | def __getitem__(self, index): 28 | try: 29 | img = Image.open(self.image_path_list[index]).convert("L") 30 | 31 | except IOError: 32 | print(f"Corrupted image for {index}") 33 | img = Image.new("L", (self.imgW, self.imgH)) 34 | 35 | return img, self.image_path_list[index] 36 | -------------------------------------------------------------------------------- /utils/image_util.py: -------------------------------------------------------------------------------- 1 | import cv2 2 | import numpy as np 3 | import platform 4 | from PIL import ImageFont, ImageDraw, Image 5 | from matplotlib import pyplot as plt 6 | 7 | 8 | def plt_imshow(title='image', img=None, figsize=(8, 5)): 9 | plt.figure(figsize=figsize) 10 | 11 | if type(img) is str: 12 | img = cv2.imread(img) 13 | 14 | if type(img) == list: 15 | if type(title) == list: 16 | titles = title 17 | else: 18 | titles = [] 19 | 20 | for i in range(len(img)): 21 | titles.append(title) 22 | 23 | for i in range(len(img)): 24 | if len(img[i].shape) <= 2: 25 | rgbImg = cv2.cvtColor(img[i], cv2.COLOR_GRAY2RGB) 26 | else: 27 | rgbImg = cv2.cvtColor(img[i], cv2.COLOR_BGR2RGB) 28 | 29 | plt.subplot(1, len(img), i + 1), plt.imshow(rgbImg) 30 | plt.title(titles[i]) 31 | plt.xticks([]), plt.yticks([]) 32 | 33 | plt.show() 34 | else: 35 | if len(img.shape) < 3: 36 | rgbImg = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) 37 | else: 38 | rgbImg = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 39 | 40 | plt.imshow(rgbImg) 41 | plt.title(title) 42 | plt.xticks([]), plt.yticks([]) 43 | plt.show() 44 | 45 | 46 | def put_text(image, text, x, y, color=(0, 255, 0), font_size=22): 47 | if type(image) == np.ndarray: 48 | color_coverted = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) 49 | image = Image.fromarray(color_coverted) 50 | 51 | if platform.system() == 'Darwin': 52 | font = 'AppleGothic.ttf' 53 | elif platform.system() == 'Windows': 54 | font = 'malgun.ttf' 55 | 56 | image_font = ImageFont.truetype(font, font_size) 57 | font = ImageFont.load_default() 58 | draw = ImageDraw.Draw(image) 59 | 60 | draw.text((x, y), text, font=image_font, fill=color) 61 | 62 | numpy_image = np.array(image) 63 | opencv_image = cv2.cvtColor(numpy_image, cv2.COLOR_RGB2BGR) 64 | 65 | return opencv_image -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 |

2 | Korean OCR using pororo 3 |

4 | 5 |
6 | 7 | 8 | 9 | 10 |
11 | 12 | This is a Korean OCR Python code using the Pororo library. 13 | 14 |
15 | 16 |
17 | 18 | ## Requirements 19 | 20 | - torch 21 | - torchvision 22 | - opencv-python 23 | 24 | You can install it from PyPI: 25 | 26 | ```sh 27 | pip install torch 28 | pip install torchvision 29 | pip install opencv-python 30 | ``` 31 | 32 | ## PORORO: Platform Of neuRal mOdels for natuRal language prOcessing 33 | 34 | [pororo](https://github.com/kakaobrain/pororo) is a library developed by KakaoBrain for performing natural language processing and speech-related tasks. 35 | 36 | This repository is configured to only include the OCR functionality from the pororo library. If you wish to use other pororo features such as natural language processing, please install pororo through `pip install pororo`. 37 | 38 | ## Usage 39 | 40 | ```python 41 | from pororo import Pororo 42 | 43 | ocr = PororoOcr() 44 | image_path = input("Enter image path: ") 45 | text = ocr.run_ocr(image_path, debug=True) 46 | print('Result :', text) 47 | ``` 48 | 49 | Output: 50 | 51 | ```sh 52 | ['메이크업존 MAKEUP ZONE', '드레스 피팅룸 DRESS FITTING ROOM', '포토존 PHOTO ZONE'] 53 | ``` 54 | 55 | ------ 56 | 57 | 58 | 59 |
60 | 61 |
62 | 63 | ```sh 64 | ["Life is ot a spectator sport. If you're going to spend your whole life in the grandstand just watching what goes on, in my apinion you're wasting your life.", 65 | "인생은 구경거리가 아니다. 무슨 일이 일어나는지 보기만 하는 것은 인생을 낭비하고 있는 것이다.", 66 | 'Jackie Robinson'] 67 | ``` -------------------------------------------------------------------------------- /pororo/utils.py: -------------------------------------------------------------------------------- 1 | from contextlib import contextmanager 2 | from tempfile import NamedTemporaryFile 3 | 4 | from requests import get 5 | 6 | 7 | def postprocess_span(tagger, text: str) -> str: 8 | """ 9 | Postprocess NOUN span to remove unnecessary character 10 | 11 | Args: 12 | text (str): NOUN span to be processed 13 | 14 | Returns: 15 | (str): post-processed NOUN span 16 | 17 | Examples: 18 | >>> postprocess_span("강감찬 장군은") 19 | '강감찬 장군' 20 | >>> postprocess_span("그녀에게") 21 | '그녀' 22 | 23 | """ 24 | 25 | # First, strip punctuations 26 | text = text.strip("""!"\#$&'()*+,\-./:;<=>?@\^_‘{|}~《》""") 27 | 28 | # Complete imbalanced parentheses pair 29 | if text.count("(") == text.count(")") + 1: 30 | text += ")" 31 | elif text.count("(") + 1 == text.count(")"): 32 | text = "(" + text 33 | 34 | # Preserve beginning tokens since we only want to extract noun phrase of the last eojeol 35 | noun_phrase = " ".join(text.rsplit(" ", 1)[:-1]) 36 | tokens = text.split(" ") 37 | eojeols = list() 38 | for token in tokens: 39 | eojeols.append(tagger.pos(token)) 40 | last_eojeol = eojeols[-1] 41 | 42 | # Iterate backwardly to remove unnecessary postfixes 43 | i = 0 44 | for i, token in enumerate(last_eojeol[::-1]): 45 | _, pos = token 46 | # 1. The loop breaks when you meet a noun 47 | # 2. The loop also breaks when you meet a XSN (e.g. 8/SN+일/NNB LG/SL 전/XSN) 48 | if (pos[0] in ("N", "S")) or pos.startswith("XSN"): 49 | break 50 | idx = len(last_eojeol) - i 51 | 52 | # Extract noun span from last eojeol and postpend it to beginning tokens 53 | ext_last_eojeol = "".join(morph for morph, _ in last_eojeol[:idx]) 54 | noun_phrase += " " + ext_last_eojeol 55 | return noun_phrase.strip() 56 | 57 | 58 | @contextmanager 59 | def control_temp(file_path: str): 60 | """ 61 | Download temporary file from web, then remove it after some context 62 | 63 | Args: 64 | file_path (str): web file path 65 | 66 | """ 67 | # yapf: disable 68 | assert file_path.startswith("http"), "File path should contain `http` prefix !" 69 | # yapf: enable 70 | 71 | ext = file_path[file_path.rfind("."):] 72 | 73 | with NamedTemporaryFile("wb", suffix=ext, delete=True) as f: 74 | response = get(file_path, allow_redirects=True) 75 | f.write(response.content) 76 | yield f.name 77 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/imgproc.py: -------------------------------------------------------------------------------- 1 | """ 2 | This is adapted from https://github.com/clovaai/CRAFT-pytorch/blob/master/imgproc.py 3 | Copyright (c) 2019-present NAVER Corp. 4 | MIT License 5 | """ 6 | 7 | import cv2 8 | import numpy as np 9 | from skimage import io 10 | 11 | 12 | def load_image(img_file): 13 | img = io.imread(img_file) # RGB order 14 | if img.shape[0] == 2: 15 | img = img[0] 16 | if len(img.shape) == 2: 17 | img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) 18 | if img.shape[2] == 4: 19 | img = img[:, :, :3] 20 | img = np.array(img) 21 | 22 | return img 23 | 24 | 25 | def normalize_mean_variance( 26 | in_img, 27 | mean=(0.485, 0.456, 0.406), 28 | variance=(0.229, 0.224, 0.225), 29 | ): 30 | # should be RGB order 31 | img = in_img.copy().astype(np.float32) 32 | 33 | img -= np.array([mean[0] * 255.0, mean[1] * 255.0, mean[2] * 255.0], 34 | dtype=np.float32) 35 | img /= np.array( 36 | [variance[0] * 255.0, variance[1] * 255.0, variance[2] * 255.0], 37 | dtype=np.float32, 38 | ) 39 | return img 40 | 41 | 42 | def denormalize_mean_variance( 43 | in_img, 44 | mean=(0.485, 0.456, 0.406), 45 | variance=(0.229, 0.224, 0.225), 46 | ): 47 | # should be RGB order 48 | img = in_img.copy() 49 | img *= variance 50 | img += mean 51 | img *= 255.0 52 | img = np.clip(img, 0, 255).astype(np.uint8) 53 | return img 54 | 55 | 56 | def resize_aspect_ratio( 57 | img: np.ndarray, 58 | square_size: int, 59 | interpolation: int, 60 | mag_ratio: float = 1.0, 61 | ): 62 | height, width, channel = img.shape 63 | 64 | # magnify image size 65 | target_size = mag_ratio * max(height, width) 66 | 67 | # set original image size 68 | if target_size > square_size: 69 | target_size = square_size 70 | 71 | ratio = target_size / max(height, width) 72 | 73 | target_h, target_w = int(height * ratio), int(width * ratio) 74 | proc = cv2.resize(img, (target_w, target_h), interpolation=interpolation) 75 | 76 | # make canvas and paste image 77 | target_h32, target_w32 = target_h, target_w 78 | if target_h % 32 != 0: 79 | target_h32 = target_h + (32 - target_h % 32) 80 | if target_w % 32 != 0: 81 | target_w32 = target_w + (32 - target_w % 32) 82 | resized = np.zeros((target_h32, target_w32, channel), dtype=np.float32) 83 | resized[0:target_h, 0:target_w, :] = proc 84 | target_h, target_w = target_h32, target_w32 85 | 86 | size_heatmap = (int(target_w / 2), int(target_h / 2)) 87 | 88 | return resized, ratio, size_heatmap 89 | 90 | 91 | def cvt2heatmap_img(img): 92 | img = (np.clip(img, 0, 1) * 255).astype(np.uint8) 93 | img = cv2.applyColorMap(img, cv2.COLORMAP_JET) 94 | return img 95 | -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import cv2 2 | from pororo import Pororo 3 | from pororo.pororo import SUPPORTED_TASKS 4 | from utils.image_util import plt_imshow, put_text 5 | import warnings 6 | 7 | warnings.filterwarnings('ignore') 8 | 9 | 10 | class PororoOcr: 11 | def __init__(self, model: str = "brainocr", lang: str = "ko", **kwargs): 12 | self.model = model 13 | self.lang = lang 14 | self._ocr = Pororo(task="ocr", lang=lang, model=model, **kwargs) 15 | self.img_path = None 16 | self.ocr_result = {} 17 | 18 | def run_ocr(self, img_path: str, debug: bool = False): 19 | self.img_path = img_path 20 | self.ocr_result = self._ocr(img_path, detail=True) 21 | 22 | if self.ocr_result['description']: 23 | ocr_text = self.ocr_result["description"] 24 | else: 25 | ocr_text = "No text detected." 26 | 27 | if debug: 28 | self.show_img_with_ocr() 29 | 30 | return ocr_text 31 | 32 | @staticmethod 33 | def get_available_langs(): 34 | return SUPPORTED_TASKS["ocr"].get_available_langs() 35 | 36 | @staticmethod 37 | def get_available_models(): 38 | return SUPPORTED_TASKS["ocr"].get_available_models() 39 | 40 | def get_ocr_result(self): 41 | return self.ocr_result 42 | 43 | def get_img_path(self): 44 | return self.img_path 45 | 46 | def show_img(self): 47 | plt_imshow(img=self.img_path) 48 | 49 | def show_img_with_ocr(self): 50 | img = cv2.imread(self.img_path) 51 | roi_img = img.copy() 52 | 53 | for text_result in self.ocr_result['bounding_poly']: 54 | text = text_result['description'] 55 | tlX = text_result['vertices'][0]['x'] 56 | tlY = text_result['vertices'][0]['y'] 57 | trX = text_result['vertices'][1]['x'] 58 | trY = text_result['vertices'][1]['y'] 59 | brX = text_result['vertices'][2]['x'] 60 | brY = text_result['vertices'][2]['y'] 61 | blX = text_result['vertices'][3]['x'] 62 | blY = text_result['vertices'][3]['y'] 63 | 64 | pts = ((tlX, tlY), (trX, trY), (brX, brY), (blX, blY)) 65 | 66 | topLeft = pts[0] 67 | topRight = pts[1] 68 | bottomRight = pts[2] 69 | bottomLeft = pts[3] 70 | 71 | cv2.line(roi_img, topLeft, topRight, (0, 255, 0), 2) 72 | cv2.line(roi_img, topRight, bottomRight, (0, 255, 0), 2) 73 | cv2.line(roi_img, bottomRight, bottomLeft, (0, 255, 0), 2) 74 | cv2.line(roi_img, bottomLeft, topLeft, (0, 255, 0), 2) 75 | roi_img = put_text(roi_img, text, topLeft[0], topLeft[1] - 20, font_size=15) 76 | 77 | # print(text) 78 | 79 | plt_imshow(["Original", "ROI"], [img, roi_img], figsize=(16, 10)) 80 | 81 | 82 | if __name__ == "__main__": 83 | ocr = PororoOcr() 84 | image_path = input("Enter image path: ") 85 | text = ocr.run_ocr(image_path, debug=True) 86 | print('Result :', text) 87 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/detection.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from https://github.com/JaidedAI/EasyOCR/blob/master/easyocr/detection.py 3 | """ 4 | 5 | from collections import OrderedDict 6 | 7 | import cv2 8 | import numpy as np 9 | import torch 10 | import torch.backends.cudnn as cudnn 11 | from torch.autograd import Variable 12 | 13 | from .craft import CRAFT 14 | from .craft_utils import adjust_result_coordinates, get_det_boxes 15 | from .imgproc import normalize_mean_variance, resize_aspect_ratio 16 | 17 | 18 | def copy_state_dict(state_dict): 19 | if list(state_dict.keys())[0].startswith("module"): 20 | start_idx = 1 21 | else: 22 | start_idx = 0 23 | new_state_dict = OrderedDict() 24 | for k, v in state_dict.items(): 25 | name = ".".join(k.split(".")[start_idx:]) 26 | new_state_dict[name] = v 27 | return new_state_dict 28 | 29 | 30 | def test_net(image: np.ndarray, net, opt2val: dict): 31 | canvas_size = opt2val["canvas_size"] 32 | mag_ratio = opt2val["mag_ratio"] 33 | text_threshold = opt2val["text_threshold"] 34 | link_threshold = opt2val["link_threshold"] 35 | low_text = opt2val["low_text"] 36 | device = opt2val["device"] 37 | 38 | # resize 39 | img_resized, target_ratio, size_heatmap = resize_aspect_ratio( 40 | image, canvas_size, interpolation=cv2.INTER_LINEAR, mag_ratio=mag_ratio) 41 | ratio_h = ratio_w = 1 / target_ratio 42 | 43 | # preprocessing 44 | x = normalize_mean_variance(img_resized) 45 | x = torch.from_numpy(x).permute(2, 0, 1) # [h, w, c] to [c, h, w] 46 | x = Variable(x.unsqueeze(0)) # [c, h, w] to [b, c, h, w] 47 | x = x.to(device) 48 | 49 | # forward pass 50 | with torch.no_grad(): 51 | y, feature = net(x) 52 | 53 | # make score and link map 54 | score_text = y[0, :, :, 0].cpu().data.numpy() 55 | score_link = y[0, :, :, 1].cpu().data.numpy() 56 | 57 | # Post-processing 58 | boxes, polys = get_det_boxes( 59 | score_text, 60 | score_link, 61 | text_threshold, 62 | link_threshold, 63 | low_text, 64 | ) 65 | 66 | # coordinate adjustment 67 | boxes = adjust_result_coordinates(boxes, ratio_w, ratio_h) 68 | polys = adjust_result_coordinates(polys, ratio_w, ratio_h) 69 | for k in range(len(polys)): 70 | if polys[k] is None: 71 | polys[k] = boxes[k] 72 | 73 | return boxes, polys 74 | 75 | 76 | def get_detector(det_model_ckpt_fp: str, device: str = "cpu"): 77 | net = CRAFT() 78 | 79 | net.load_state_dict( 80 | copy_state_dict(torch.load(det_model_ckpt_fp, map_location=device))) 81 | if device == "cuda": 82 | net = torch.nn.DataParallel(net).to(device) 83 | cudnn.benchmark = False 84 | 85 | net.eval() 86 | return net 87 | 88 | 89 | def get_textbox(detector, image: np.ndarray, opt2val: dict): 90 | bboxes, polys = test_net(image, detector, opt2val) 91 | result = [] 92 | for i, box in enumerate(polys): 93 | poly = np.array(box).astype(np.int32).reshape((-1)) 94 | result.append(poly) 95 | 96 | return result 97 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/basenet.py: -------------------------------------------------------------------------------- 1 | from collections import namedtuple 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.init as init 6 | from torchvision import models 7 | from torchvision.models.vgg import model_urls 8 | 9 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 10 | 11 | 12 | def init_weights(modules): 13 | for m in modules: 14 | if isinstance(m, nn.Conv2d): 15 | init.xavier_uniform_(m.weight.data) 16 | if m.bias is not None: 17 | m.bias.data.zero_() 18 | elif isinstance(m, nn.BatchNorm2d): 19 | m.weight.data.fill_(1) 20 | m.bias.data.zero_() 21 | elif isinstance(m, nn.Linear): 22 | m.weight.data.normal_(0, 0.01) 23 | m.bias.data.zero_() 24 | 25 | 26 | class Vgg16BN(torch.nn.Module): 27 | 28 | def __init__(self, pretrained: bool = True, freeze: bool = True): 29 | super(Vgg16BN, self).__init__() 30 | model_urls["vgg16_bn"] = model_urls["vgg16_bn"].replace( 31 | "https://", "http://") 32 | vgg_pretrained_features = models.vgg16_bn( 33 | pretrained=pretrained).features 34 | self.slice1 = torch.nn.Sequential() 35 | self.slice2 = torch.nn.Sequential() 36 | self.slice3 = torch.nn.Sequential() 37 | self.slice4 = torch.nn.Sequential() 38 | self.slice5 = torch.nn.Sequential() 39 | for x in range(12): # conv2_2 40 | self.slice1.add_module(str(x), vgg_pretrained_features[x]) 41 | for x in range(12, 19): # conv3_3 42 | self.slice2.add_module(str(x), vgg_pretrained_features[x]) 43 | for x in range(19, 29): # conv4_3 44 | self.slice3.add_module(str(x), vgg_pretrained_features[x]) 45 | for x in range(29, 39): # conv5_3 46 | self.slice4.add_module(str(x), vgg_pretrained_features[x]) 47 | 48 | # fc6, fc7 without atrous conv 49 | self.slice5 = torch.nn.Sequential( 50 | nn.MaxPool2d(kernel_size=3, stride=1, padding=1), 51 | nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6), 52 | nn.Conv2d(1024, 1024, kernel_size=1), 53 | ) 54 | 55 | if not pretrained: 56 | init_weights(self.slice1.modules()) 57 | init_weights(self.slice2.modules()) 58 | init_weights(self.slice3.modules()) 59 | init_weights(self.slice4.modules()) 60 | 61 | init_weights( 62 | self.slice5.modules()) # no pretrained model for fc6 and fc7 63 | 64 | if freeze: 65 | for param in self.slice1.parameters(): # only first conv 66 | param.requires_grad = False 67 | 68 | def forward(self, x): 69 | h = self.slice1(x) 70 | h_relu2_2 = h 71 | h = self.slice2(h) 72 | h_relu3_2 = h 73 | h = self.slice3(h) 74 | h_relu4_3 = h 75 | h = self.slice4(h) 76 | h_relu5_3 = h 77 | h = self.slice5(h) 78 | h_fc7 = h 79 | vgg_outputs = namedtuple( 80 | "VggOutputs", ["fc7", "relu5_3", "relu4_3", "relu3_2", "relu2_2"]) 81 | out = vgg_outputs(h_fc7, h_relu5_3, h_relu4_3, h_relu3_2, h_relu2_2) 82 | return out 83 | -------------------------------------------------------------------------------- /pororo/tasks/utils/tokenizer.py: -------------------------------------------------------------------------------- 1 | from typing import List, Optional, Union 2 | 3 | from tokenizers import Tokenizer, decoders, pre_tokenizers 4 | from tokenizers.implementations import BaseTokenizer 5 | from tokenizers.models import BPE, Unigram 6 | from tokenizers.normalizers import NFKC 7 | 8 | 9 | class CustomTokenizer(BaseTokenizer): 10 | 11 | def __init__( 12 | self, 13 | vocab: Union[str, List], 14 | merges: Union[str, None], 15 | unk_token: str = "", 16 | replacement: str = "▁", 17 | add_prefix_space: bool = True, 18 | dropout: Optional[float] = None, 19 | normalize: bool = True, 20 | ): 21 | if merges: 22 | n_model = "BPE" 23 | tokenizer = Tokenizer( 24 | BPE( 25 | vocab, # type: ignore 26 | merges, 27 | unk_token=unk_token, 28 | fuse_unk=True, 29 | )) 30 | else: 31 | n_model = "Unigram" 32 | tokenizer = Tokenizer(Unigram(vocab, 1)) # type: ignore 33 | 34 | if normalize: 35 | tokenizer.normalizer = NFKC() 36 | 37 | tokenizer.pre_tokenizer = pre_tokenizers.Metaspace( 38 | replacement=replacement, 39 | add_prefix_space=add_prefix_space, 40 | ) 41 | 42 | tokenizer.decoder = decoders.Metaspace( 43 | replacement=replacement, 44 | add_prefix_space=add_prefix_space, 45 | ) 46 | 47 | parameters = { 48 | "model": f"SentencePiece{n_model}", 49 | "unk_token": unk_token, 50 | "replacement": replacement, 51 | "add_prefix_space": add_prefix_space, 52 | "dropout": dropout, 53 | } 54 | super().__init__(tokenizer, parameters) 55 | 56 | @staticmethod 57 | def from_file( 58 | vocab_filename: str, 59 | merges_filename: Union[str, None], 60 | **kwargs, 61 | ): 62 | # BPE 63 | if merges_filename: 64 | vocab, merges = BPE.read_file(vocab_filename, merges_filename) 65 | 66 | # Unigram 67 | else: 68 | vocab = [] 69 | merges = None 70 | with open(vocab_filename, "r") as f_in: 71 | for line in f_in.readlines(): 72 | token, score = line.strip().split("\t") 73 | vocab.append((token, float(score))) 74 | 75 | return CustomTokenizer(vocab, merges, **kwargs) 76 | 77 | def segment(self, text: str) -> List[str]: 78 | """ 79 | Segment text into subword list 80 | 81 | Args: 82 | text (str): input text to be segmented 83 | 84 | Returns: 85 | List[str]: segmented subword list 86 | 87 | """ 88 | encoding = self.encode(text) 89 | 90 | offsets = encoding.offsets 91 | tokens = encoding.tokens 92 | 93 | result = [] 94 | for offset, token in zip(offsets, tokens): 95 | if token != "": 96 | result.append(token) 97 | continue 98 | s, e = offset 99 | result.append(text[s:e]) 100 | return result 101 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/craft.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from https://github.com/clovaai/CRAFT-pytorch/blob/master/craft.py. 3 | Copyright (c) 2019-present NAVER Corp. 4 | MIT License 5 | """ 6 | 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | from torch import Tensor 11 | 12 | from ._modules import Vgg16BN, init_weights 13 | 14 | 15 | class DoubleConv(nn.Module): 16 | 17 | def __init__(self, in_ch: int, mid_ch: int, out_ch: int) -> None: 18 | super(DoubleConv, self).__init__() 19 | self.conv = nn.Sequential( 20 | nn.Conv2d(in_ch + mid_ch, mid_ch, kernel_size=1), 21 | nn.BatchNorm2d(mid_ch), 22 | nn.ReLU(inplace=True), 23 | nn.Conv2d(mid_ch, out_ch, kernel_size=3, padding=1), 24 | nn.BatchNorm2d(out_ch), 25 | nn.ReLU(inplace=True), 26 | ) 27 | 28 | def forward(self, x: Tensor): 29 | x = self.conv(x) 30 | return x 31 | 32 | 33 | class CRAFT(nn.Module): 34 | 35 | def __init__(self, pretrained: bool = False, freeze: bool = False) -> None: 36 | super(CRAFT, self).__init__() 37 | 38 | # Base network 39 | self.basenet = Vgg16BN(pretrained, freeze) 40 | 41 | # U network 42 | self.upconv1 = DoubleConv(1024, 512, 256) 43 | self.upconv2 = DoubleConv(512, 256, 128) 44 | self.upconv3 = DoubleConv(256, 128, 64) 45 | self.upconv4 = DoubleConv(128, 64, 32) 46 | 47 | num_class = 2 48 | self.conv_cls = nn.Sequential( 49 | nn.Conv2d(32, 32, kernel_size=3, padding=1), 50 | nn.ReLU(inplace=True), 51 | nn.Conv2d(32, 32, kernel_size=3, padding=1), 52 | nn.ReLU(inplace=True), 53 | nn.Conv2d(32, 16, kernel_size=3, padding=1), 54 | nn.ReLU(inplace=True), 55 | nn.Conv2d(16, 16, kernel_size=1), 56 | nn.ReLU(inplace=True), 57 | nn.Conv2d(16, num_class, kernel_size=1), 58 | ) 59 | 60 | init_weights(self.upconv1.modules()) 61 | init_weights(self.upconv2.modules()) 62 | init_weights(self.upconv3.modules()) 63 | init_weights(self.upconv4.modules()) 64 | init_weights(self.conv_cls.modules()) 65 | 66 | def forward(self, x: Tensor): 67 | # Base network 68 | sources = self.basenet(x) 69 | 70 | # U network 71 | y = torch.cat([sources[0], sources[1]], dim=1) 72 | y = self.upconv1(y) 73 | 74 | y = F.interpolate( 75 | y, 76 | size=sources[2].size()[2:], 77 | mode="bilinear", 78 | align_corners=False, 79 | ) 80 | y = torch.cat([y, sources[2]], dim=1) 81 | y = self.upconv2(y) 82 | 83 | y = F.interpolate( 84 | y, 85 | size=sources[3].size()[2:], 86 | mode="bilinear", 87 | align_corners=False, 88 | ) 89 | y = torch.cat([y, sources[3]], dim=1) 90 | y = self.upconv3(y) 91 | 92 | y = F.interpolate( 93 | y, 94 | size=sources[4].size()[2:], 95 | mode="bilinear", 96 | align_corners=False, 97 | ) 98 | y = torch.cat([y, sources[4]], dim=1) 99 | feature = self.upconv4(y) 100 | 101 | y = self.conv_cls(feature) 102 | 103 | return y.permute(0, 2, 3, 1), feature 104 | -------------------------------------------------------------------------------- /pororo/pororo.py: -------------------------------------------------------------------------------- 1 | """ 2 | Pororo task-specific factory class 3 | 4 | isort:skip_file 5 | 6 | """ 7 | 8 | import logging 9 | from typing import Optional 10 | from pororo.tasks.utils.base import PororoTaskBase 11 | 12 | import torch 13 | 14 | from pororo.tasks import ( 15 | PororoOcrFactory, 16 | ) 17 | 18 | SUPPORTED_TASKS = { 19 | "ocr": PororoOcrFactory, 20 | } 21 | 22 | LANG_ALIASES = { 23 | "english": "en", 24 | "eng": "en", 25 | "korean": "ko", 26 | "kor": "ko", 27 | "kr": "ko", 28 | "chinese": "zh", 29 | "chn": "zh", 30 | "cn": "zh", 31 | "japanese": "ja", 32 | "jap": "ja", 33 | "jp": "ja", 34 | "jejueo": "je", 35 | "jje": "je", 36 | } 37 | 38 | logging.getLogger("transformers").setLevel(logging.WARN) 39 | logging.getLogger("fairseq").setLevel(logging.WARN) 40 | logging.getLogger("sentence_transformers").setLevel(logging.WARN) 41 | logging.getLogger("youtube_dl").setLevel(logging.WARN) 42 | logging.getLogger("pydub").setLevel(logging.WARN) 43 | logging.getLogger("librosa").setLevel(logging.WARN) 44 | 45 | 46 | class Pororo: 47 | r""" 48 | This is a generic class that will return one of the task-specific model classes of the library 49 | when created with the `__new__()` method 50 | 51 | """ 52 | 53 | def __new__( 54 | cls, 55 | task: str, 56 | lang: str = "en", 57 | model: Optional[str] = None, 58 | **kwargs, 59 | ) -> PororoTaskBase: 60 | if task not in SUPPORTED_TASKS: 61 | raise KeyError("Unknown task {}, available tasks are {}".format( 62 | task, 63 | list(SUPPORTED_TASKS.keys()), 64 | )) 65 | 66 | lang = lang.lower() 67 | lang = LANG_ALIASES[lang] if lang in LANG_ALIASES else lang 68 | 69 | # Get device information from torch API 70 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 71 | 72 | # Instantiate task-specific pipeline module, if possible 73 | task_module = SUPPORTED_TASKS[task]( 74 | task, 75 | lang, 76 | model, 77 | **kwargs, 78 | ).load(device) 79 | 80 | return task_module 81 | 82 | @staticmethod 83 | def available_tasks() -> str: 84 | """ 85 | Returns available tasks in Pororo project 86 | 87 | Returns: 88 | str: Supported task names 89 | 90 | """ 91 | return "Available tasks are {}".format(list(SUPPORTED_TASKS.keys())) 92 | 93 | @staticmethod 94 | def available_models(task: str) -> str: 95 | """ 96 | Returns available model names correponding to the user-input task 97 | 98 | Args: 99 | task (str): user-input task name 100 | 101 | Returns: 102 | str: Supported model names corresponding to the user-input task 103 | 104 | Raises: 105 | KeyError: When user-input task is not supported 106 | 107 | """ 108 | if task not in SUPPORTED_TASKS: 109 | raise KeyError( 110 | "Unknown task {} ! Please check available models via `available_tasks()`" 111 | .format(task)) 112 | 113 | langs = SUPPORTED_TASKS[task].get_available_models() 114 | output = f"Available models for {task} are " 115 | for lang in langs: 116 | output += f"([lang]: {lang}, [model]: {', '.join(langs[lang])}), " 117 | return output[:-2] 118 | -------------------------------------------------------------------------------- /pororo/tasks/utils/config.py: -------------------------------------------------------------------------------- 1 | from dataclasses import dataclass 2 | from typing import Union 3 | 4 | 5 | @dataclass 6 | class TransformerConfig: 7 | src_dict: Union[str, None] 8 | tgt_dict: Union[str, None] 9 | src_tok: Union[str, None] 10 | tgt_tok: Union[str, None] 11 | 12 | 13 | CONFIGS = { 14 | "transformer.base.ko.const": 15 | TransformerConfig( 16 | "dict.transformer.base.ko.const", 17 | "dict.transformer.base.ko.const", 18 | None, 19 | None, 20 | ), 21 | "transformer.base.ko.pg": 22 | TransformerConfig( 23 | "dict.transformer.base.ko.mt", 24 | "dict.transformer.base.ko.mt", 25 | "bpe8k.ko", 26 | None, 27 | ), 28 | "transformer.base.ko.pg_long": 29 | TransformerConfig( 30 | "dict.transformer.base.ko.mt", 31 | "dict.transformer.base.ko.mt", 32 | "bpe8k.ko", 33 | None, 34 | ), 35 | "transformer.base.en.gec": 36 | TransformerConfig( 37 | "dict.transformer.base.en.mt", 38 | "dict.transformer.base.en.mt", 39 | "bpe32k.en", 40 | None, 41 | ), 42 | "transformer.base.zh.pg": 43 | TransformerConfig( 44 | "dict.transformer.base.zh.mt", 45 | "dict.transformer.base.zh.mt", 46 | None, 47 | None, 48 | ), 49 | "transformer.base.ja.pg": 50 | TransformerConfig( 51 | "dict.transformer.base.ja.mt", 52 | "dict.transformer.base.ja.mt", 53 | "bpe8k.ja", 54 | None, 55 | ), 56 | "transformer.base.zh.const": 57 | TransformerConfig( 58 | "dict.transformer.base.zh.const", 59 | "dict.transformer.base.zh.const", 60 | None, 61 | None, 62 | ), 63 | "transformer.base.en.const": 64 | TransformerConfig( 65 | "dict.transformer.base.en.const", 66 | "dict.transformer.base.en.const", 67 | None, 68 | None, 69 | ), 70 | "transformer.base.en.pg": 71 | TransformerConfig( 72 | "dict.transformer.base.en.mt", 73 | "dict.transformer.base.en.mt", 74 | "bpe32k.en", 75 | None, 76 | ), 77 | "transformer.base.ko.gec": 78 | TransformerConfig( 79 | "dict.transformer.base.ko.gec", 80 | "dict.transformer.base.ko.gec", 81 | "bpe8k.ko", 82 | None, 83 | ), 84 | "transformer.base.en.char_gec": 85 | TransformerConfig( 86 | "dict.transformer.base.en.char_gec", 87 | "dict.transformer.base.en.char_gec", 88 | None, 89 | None, 90 | ), 91 | "transformer.base.en.caption": 92 | TransformerConfig( 93 | None, 94 | None, 95 | None, 96 | None, 97 | ), 98 | "transformer.base.ja.p2g": 99 | TransformerConfig( 100 | "dict.transformer.base.ja.p2g", 101 | "dict.transformer.base.ja.p2g", 102 | None, 103 | None, 104 | ), 105 | "transformer.large.multi.mtpg": 106 | TransformerConfig( 107 | "dict.transformer.large.multi.mtpg", 108 | "dict.transformer.large.multi.mtpg", 109 | "bpe32k.en", 110 | None, 111 | ), 112 | "transformer.large.multi.fast.mtpg": 113 | TransformerConfig( 114 | "dict.transformer.large.multi.mtpg", 115 | "dict.transformer.large.multi.mtpg", 116 | "bpe32k.en", 117 | None, 118 | ), 119 | "transformer.large.ko.wsd": 120 | TransformerConfig( 121 | "dict.transformer.large.ko.wsd", 122 | "dict.transformer.large.ko.wsd", 123 | None, 124 | None, 125 | ), 126 | } 127 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/model.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from 3 | https://github.com/clovaai/deep-text-recognition-benchmark/blob/master/model.py 4 | """ 5 | 6 | import torch.nn as nn 7 | from torch import Tensor 8 | 9 | from .modules.feature_extraction import ( 10 | ResNetFeatureExtractor, 11 | VGGFeatureExtractor, 12 | ) 13 | from .modules.prediction import Attention 14 | from .modules.sequence_modeling import BidirectionalLSTM 15 | from .modules.transformation import TpsSpatialTransformerNetwork 16 | 17 | 18 | class Model(nn.Module): 19 | 20 | def __init__(self, opt2val: dict): 21 | super(Model, self).__init__() 22 | 23 | input_channel = opt2val["input_channel"] 24 | output_channel = opt2val["output_channel"] 25 | hidden_size = opt2val["hidden_size"] 26 | vocab_size = opt2val["vocab_size"] 27 | num_fiducial = opt2val["num_fiducial"] 28 | imgH = opt2val["imgH"] 29 | imgW = opt2val["imgW"] 30 | FeatureExtraction = opt2val["FeatureExtraction"] 31 | Transformation = opt2val["Transformation"] 32 | SequenceModeling = opt2val["SequenceModeling"] 33 | Prediction = opt2val["Prediction"] 34 | 35 | # Transformation 36 | if Transformation == "TPS": 37 | self.Transformation = TpsSpatialTransformerNetwork( 38 | F=num_fiducial, 39 | I_size=(imgH, imgW), 40 | I_r_size=(imgH, imgW), 41 | I_channel_num=input_channel, 42 | ) 43 | else: 44 | print("No Transformation module specified") 45 | 46 | # FeatureExtraction 47 | if FeatureExtraction == "VGG": 48 | extractor = VGGFeatureExtractor 49 | else: # ResNet 50 | extractor = ResNetFeatureExtractor 51 | self.FeatureExtraction = extractor( 52 | input_channel, 53 | output_channel, 54 | opt2val, 55 | ) 56 | self.FeatureExtraction_output = output_channel # int(imgH/16-1) * 512 57 | self.AdaptiveAvgPool = nn.AdaptiveAvgPool2d( 58 | (None, 1)) # Transform final (imgH/16-1) -> 1 59 | 60 | # Sequence modeling 61 | if SequenceModeling == "BiLSTM": 62 | self.SequenceModeling = nn.Sequential( 63 | BidirectionalLSTM( 64 | self.FeatureExtraction_output, 65 | hidden_size, 66 | hidden_size, 67 | ), 68 | BidirectionalLSTM(hidden_size, hidden_size, hidden_size), 69 | ) 70 | self.SequenceModeling_output = hidden_size 71 | else: 72 | print("No SequenceModeling module specified") 73 | self.SequenceModeling_output = self.FeatureExtraction_output 74 | 75 | # Prediction 76 | if Prediction == "CTC": 77 | self.Prediction = nn.Linear( 78 | self.SequenceModeling_output, 79 | vocab_size, 80 | ) 81 | elif Prediction == "Attn": 82 | self.Prediction = Attention( 83 | self.SequenceModeling_output, 84 | hidden_size, 85 | vocab_size, 86 | ) 87 | elif Prediction == "Transformer": # TODO 88 | pass 89 | else: 90 | raise Exception("Prediction is neither CTC or Attn") 91 | 92 | def forward(self, x: Tensor): 93 | """ 94 | :param x: (batch, input_channel, height, width) 95 | :return: 96 | """ 97 | # Transformation stage 98 | x = self.Transformation(x) 99 | 100 | # Feature extraction stage 101 | visual_feature = self.FeatureExtraction( 102 | x) # (b, output_channel=512, h=3, w) 103 | visual_feature = self.AdaptiveAvgPool(visual_feature.permute( 104 | 0, 3, 1, 2)) # (b, w, channel=512, h=1) 105 | visual_feature = visual_feature.squeeze(3) # (b, w, channel=512) 106 | 107 | # Sequence modeling stage 108 | self.SequenceModeling.eval() 109 | contextual_feature = self.SequenceModeling(visual_feature) 110 | 111 | # Prediction stage 112 | prediction = self.Prediction( 113 | contextual_feature.contiguous()) # (b, T, num_classes) 114 | 115 | return prediction 116 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/prediction.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 6 | 7 | 8 | class Attention(nn.Module): 9 | 10 | def __init__(self, input_size, hidden_size, num_classes): 11 | super(Attention, self).__init__() 12 | self.attention_cell = AttentionCell(input_size, hidden_size, 13 | num_classes) 14 | self.hidden_size = hidden_size 15 | self.num_classes = num_classes 16 | self.generator = nn.Linear(hidden_size, num_classes) 17 | 18 | def _char_to_onehot(self, input_char, onehot_dim=38): 19 | input_char = input_char.unsqueeze(1) 20 | batch_size = input_char.size(0) 21 | one_hot = torch.FloatTensor(batch_size, onehot_dim).zero_().to(device) 22 | one_hot = one_hot.scatter_(1, input_char, 1) 23 | return one_hot 24 | 25 | def forward(self, batch_H, text, is_train=True, batch_max_length=25): 26 | """ 27 | input: 28 | batch_H : contextual_feature H = hidden state of encoder. [batch_size x num_steps x contextual_feature_channels] 29 | text : the text-index of each image. [batch_size x (max_length+1)]. +1 for [GO] token. text[:, 0] = [GO]. 30 | output: probability distribution at each step [batch_size x num_steps x num_classes] 31 | """ 32 | batch_size = batch_H.size(0) 33 | num_steps = batch_max_length + 1 # +1 for [s] at end of sentence. 34 | 35 | output_hiddens = (torch.FloatTensor( 36 | batch_size, num_steps, self.hidden_size).fill_(0).to(device)) 37 | hidden = ( 38 | torch.FloatTensor(batch_size, self.hidden_size).fill_(0).to(device), 39 | torch.FloatTensor(batch_size, self.hidden_size).fill_(0).to(device), 40 | ) 41 | 42 | if is_train: 43 | for i in range(num_steps): 44 | # one-hot vectors for a i-th char. in a batch 45 | char_onehots = self._char_to_onehot(text[:, i], 46 | onehot_dim=self.num_classes) 47 | # hidden : decoder's hidden s_{t-1}, batch_H : encoder's hidden H, char_onehots : one-hot(y_{t-1}) 48 | hidden, alpha = self.attention_cell(hidden, batch_H, 49 | char_onehots) 50 | output_hiddens[:, i, :] = hidden[ 51 | 0] # LSTM hidden index (0: hidden, 1: Cell) 52 | probs = self.generator(output_hiddens) 53 | 54 | else: 55 | targets = torch.LongTensor(batch_size).fill_(0).to( 56 | device) # [GO] token 57 | probs = (torch.FloatTensor(batch_size, num_steps, 58 | self.num_classes).fill_(0).to(device)) 59 | 60 | for i in range(num_steps): 61 | char_onehots = self._char_to_onehot(targets, 62 | onehot_dim=self.num_classes) 63 | hidden, alpha = self.attention_cell(hidden, batch_H, 64 | char_onehots) 65 | probs_step = self.generator(hidden[0]) 66 | probs[:, i, :] = probs_step 67 | _, next_input = probs_step.max(1) 68 | targets = next_input 69 | 70 | return probs # batch_size x num_steps x num_classes 71 | 72 | 73 | class AttentionCell(nn.Module): 74 | 75 | def __init__(self, input_size, hidden_size, num_embeddings): 76 | super(AttentionCell, self).__init__() 77 | self.i2h = nn.Linear(input_size, hidden_size, bias=False) 78 | self.h2h = nn.Linear(hidden_size, 79 | hidden_size) # either i2i or h2h should have bias 80 | self.score = nn.Linear(hidden_size, 1, bias=False) 81 | self.rnn = nn.LSTMCell(input_size + num_embeddings, hidden_size) 82 | self.hidden_size = hidden_size 83 | 84 | def forward(self, prev_hidden, batch_H, char_onehots): 85 | # [batch_size x num_encoder_step x num_channel] -> [batch_size x num_encoder_step x hidden_size] 86 | batch_H_proj = self.i2h(batch_H) 87 | prev_hidden_proj = self.h2h(prev_hidden[0]).unsqueeze(1) 88 | e = self.score( 89 | torch.tanh(batch_H_proj + 90 | prev_hidden_proj)) # batch_size x num_encoder_step * 1 91 | 92 | alpha = F.softmax(e, dim=1) 93 | context = torch.bmm(alpha.permute(0, 2, 1), 94 | batch_H).squeeze(1) # batch_size x num_channel 95 | concat_context = torch.cat( 96 | [context, char_onehots], 97 | 1) # batch_size x (num_channel + num_embedding) 98 | cur_hidden = self.rnn(concat_context, prev_hidden) 99 | return cur_hidden, alpha 100 | -------------------------------------------------------------------------------- /pororo/tasks/utils/base.py: -------------------------------------------------------------------------------- 1 | import re 2 | import unicodedata 3 | from abc import abstractmethod 4 | from dataclasses import dataclass 5 | from typing import List, Mapping, Optional, Union 6 | 7 | 8 | @dataclass 9 | class TaskConfig: 10 | task: str 11 | lang: str 12 | n_model: str 13 | 14 | 15 | class PororoTaskBase: 16 | r"""Task base class that implements basic functions for prediction""" 17 | 18 | def __init__(self, config: TaskConfig): 19 | self.config = config 20 | 21 | @property 22 | def n_model(self): 23 | return self.config.n_model 24 | 25 | @property 26 | def lang(self): 27 | return self.config.lang 28 | 29 | @abstractmethod 30 | def predict( 31 | self, 32 | text: Union[str, List[str]], 33 | **kwargs, 34 | ): 35 | raise NotImplementedError( 36 | "`predict()` function is not implemented properly!") 37 | 38 | def __call__(self): 39 | raise NotImplementedError( 40 | "`call()` function is not implemented properly!") 41 | 42 | def __repr__(self): 43 | return f"[TASK]: {self.config.task.upper()}\n[LANG]: {self.config.lang.upper()}\n[MODEL]: {self.config.n_model}" 44 | 45 | def _normalize(self, text: str): 46 | """Unicode normalization and whitespace removal (often needed for contexts)""" 47 | text = unicodedata.normalize("NFKC", text) 48 | text = re.sub(r"\s+", " ", text).strip() 49 | return text 50 | 51 | 52 | class PororoFactoryBase(object): 53 | r"""This is a factory base class that construct task-specific module""" 54 | 55 | def __init__( 56 | self, 57 | task: str, 58 | lang: str, 59 | model: Optional[str] = None, 60 | ): 61 | self._available_langs = self.get_available_langs() 62 | self._available_models = self.get_available_models() 63 | self._model2lang = { 64 | v: k for k, vs in self._available_models.items() for v in vs 65 | } 66 | 67 | # Set default language as very first supported language 68 | assert ( 69 | lang in self._available_langs 70 | ), f"Following langs are supported for this task: {self._available_langs}" 71 | 72 | if lang is None: 73 | lang = self._available_langs[0] 74 | 75 | # Change language option if model is defined by user 76 | if model is not None: 77 | lang = self._model2lang[model] 78 | 79 | # Set default model 80 | if model is None: 81 | model = self.get_default_model(lang) 82 | 83 | # yapf: disable 84 | assert (model in self._available_models[lang]), f"{model} is NOT supported for {lang}" 85 | # yapf: enable 86 | 87 | self.config = TaskConfig(task, lang, model) 88 | 89 | @abstractmethod 90 | def get_available_langs(self) -> List[str]: 91 | raise NotImplementedError( 92 | "`get_available_langs()` is not implemented properly!") 93 | 94 | @abstractmethod 95 | def get_available_models(self) -> Mapping[str, List[str]]: 96 | raise NotImplementedError( 97 | "`get_available_models()` is not implemented properly!") 98 | 99 | @abstractmethod 100 | def get_default_model(self, lang: str) -> str: 101 | return self._available_models[lang][0] 102 | 103 | @classmethod 104 | def load(cls) -> PororoTaskBase: 105 | raise NotImplementedError( 106 | "Model load function is not implemented properly!") 107 | 108 | 109 | class PororoSimpleBase(PororoTaskBase): 110 | r"""Simple task base wrapper class""" 111 | 112 | def __call__(self, text: str, **kwargs): 113 | return self.predict(text, **kwargs) 114 | 115 | 116 | class PororoBiencoderBase(PororoTaskBase): 117 | r"""Bi-Encoder base wrapper class""" 118 | 119 | def __call__( 120 | self, 121 | sent_a: str, 122 | sent_b: Union[str, List[str]], 123 | **kwargs, 124 | ): 125 | assert isinstance(sent_a, str), "sent_a should be string type" 126 | assert isinstance(sent_b, str) or isinstance( 127 | sent_b, list), "sent_b should be string or list of string type" 128 | 129 | sent_a = self._normalize(sent_a) 130 | 131 | # For "Find Similar Sentence" task 132 | if isinstance(sent_b, list): 133 | sent_b = [self._normalize(t) for t in sent_b] 134 | else: 135 | sent_b = self._normalize(sent_b) 136 | 137 | return self.predict(sent_a, sent_b, **kwargs) 138 | 139 | 140 | class PororoGenerationBase(PororoTaskBase): 141 | r"""Generation task wrapper class using various generation tricks""" 142 | 143 | def __call__( 144 | self, 145 | text: str, 146 | beam: int = 5, 147 | temperature: float = 1.0, 148 | top_k: int = -1, 149 | top_p: float = -1, 150 | no_repeat_ngram_size: int = 4, 151 | len_penalty: float = 1.0, 152 | **kwargs, 153 | ): 154 | assert isinstance(text, str), "Input text should be string type" 155 | 156 | return self.predict( 157 | text, 158 | beam=beam, 159 | temperature=temperature, 160 | top_k=top_k, 161 | top_p=top_p, 162 | no_repeat_ngram_size=no_repeat_ngram_size, 163 | len_penalty=len_penalty, 164 | **kwargs, 165 | ) 166 | 167 | 168 | class PororoTaskGenerationBase(PororoTaskBase): 169 | r"""Generation task wrapper class using only beam search""" 170 | 171 | def __call__(self, text: str, beam: int = 1, **kwargs): 172 | assert isinstance(text, str), "Input text should be string type" 173 | 174 | text = self._normalize(text) 175 | 176 | return self.predict(text, beam=beam, **kwargs) 177 | -------------------------------------------------------------------------------- /pororo/tasks/optical_character_recognition.py: -------------------------------------------------------------------------------- 1 | """OCR related modeling class""" 2 | 3 | from typing import Optional 4 | 5 | from pororo.tasks import download_or_load 6 | from pororo.tasks.utils.base import PororoFactoryBase, PororoSimpleBase 7 | 8 | 9 | class PororoOcrFactory(PororoFactoryBase): 10 | """ 11 | Recognize optical characters in image file 12 | Currently support Korean language 13 | 14 | English + Korean (`brainocr`) 15 | 16 | - dataset: Internal data + AI hub Font Image dataset 17 | - metric: TBU 18 | - ref: https://www.aihub.or.kr/aidata/133 19 | 20 | Examples: 21 | >>> ocr = Pororo(task="ocr", lang="ko") 22 | >>> ocr(IMAGE_PATH) 23 | ["사이렌'(' 신마'", "내가 말했잖아 속지열라고 이 손을 잡는 너는 위협해질 거라고"] 24 | 25 | >>> ocr = Pororo(task="ocr", lang="ko") 26 | >>> ocr(IMAGE_PATH, detail=True) 27 | { 28 | 'description': ["사이렌'(' 신마', "내가 말했잖아 속지열라고 이 손을 잡는 너는 위협해질 거라고"], 29 | 'bounding_poly': [ 30 | { 31 | 'description': "사이렌'(' 신마'", 32 | 'vertices': [ 33 | {'x': 93, 'y': 7}, 34 | {'x': 164, 'y': 7}, 35 | {'x': 164, 'y': 21}, 36 | {'x': 93, 'y': 21} 37 | ] 38 | }, 39 | { 40 | 'description': "내가 말했잖아 속지열라고 이 손을 잡는 너는 위협해질 거라고", 41 | 'vertices': [ 42 | {'x': 0, 'y': 30}, 43 | {'x': 259, 'y': 30}, 44 | {'x': 259, 'y': 194}, 45 | {'x': 0, 'y': 194}]} 46 | ] 47 | } 48 | } 49 | """ 50 | 51 | def __init__(self, task: str, lang: str, model: Optional[str]): 52 | super().__init__(task, lang, model) 53 | self.detect_model = "craft" 54 | self.ocr_opt = "ocr-opt" 55 | 56 | @staticmethod 57 | def get_available_langs(): 58 | return ["en", "ko"] 59 | 60 | @staticmethod 61 | def get_available_models(): 62 | return { 63 | "en": ["brainocr"], 64 | "ko": ["brainocr"], 65 | } 66 | 67 | def load(self, device: str): 68 | """ 69 | Load user-selected task-specific model 70 | 71 | Args: 72 | device (str): device information 73 | 74 | Returns: 75 | object: User-selected task-specific model 76 | 77 | """ 78 | if self.config.n_model == "brainocr": 79 | from pororo.models.brainOCR import brainocr 80 | 81 | if self.config.lang not in self.get_available_langs(): 82 | raise ValueError( 83 | f"Unsupported Language : {self.config.lang}", 84 | 'Support Languages : ["en", "ko"]', 85 | ) 86 | 87 | det_model_path = download_or_load( 88 | f"misc/{self.detect_model}.pt", 89 | self.config.lang, 90 | ) 91 | rec_model_path = download_or_load( 92 | f"misc/{self.config.n_model}.pt", 93 | self.config.lang, 94 | ) 95 | opt_fp = download_or_load( 96 | f"misc/{self.ocr_opt}.txt", 97 | self.config.lang, 98 | ) 99 | model = brainocr.Reader( 100 | self.config.lang, 101 | det_model_ckpt_fp=det_model_path, 102 | rec_model_ckpt_fp=rec_model_path, 103 | opt_fp=opt_fp, 104 | device=device, 105 | ) 106 | model.detector.to(device) 107 | model.recognizer.to(device) 108 | return PororoOCR(model, self.config) 109 | 110 | 111 | class PororoOCR(PororoSimpleBase): 112 | 113 | def __init__(self, model, config): 114 | super().__init__(config) 115 | self._model = model 116 | 117 | def _postprocess(self, ocr_results, detail: bool = False): 118 | """ 119 | Post-process for OCR result 120 | 121 | Args: 122 | ocr_results (list): list contains result of OCR 123 | detail (bool): if True, returned to include details. (bounding poly, vertices, etc) 124 | 125 | """ 126 | sorted_ocr_results = sorted( 127 | ocr_results, 128 | key=lambda x: ( 129 | x[0][0][1], 130 | x[0][0][0], 131 | ), 132 | ) 133 | 134 | if not detail: 135 | return [ 136 | sorted_ocr_results[i][-1] 137 | for i in range(len(sorted_ocr_results)) 138 | ] 139 | 140 | result_dict = { 141 | "description": list(), 142 | "bounding_poly": list(), 143 | } 144 | 145 | for ocr_result in sorted_ocr_results: 146 | vertices = list() 147 | 148 | for vertice in ocr_result[0]: 149 | vertices.append({ 150 | "x": vertice[0], 151 | "y": vertice[1], 152 | }) 153 | 154 | result_dict["description"].append(ocr_result[1]) 155 | result_dict["bounding_poly"].append({ 156 | "description": ocr_result[1], 157 | "vertices": vertices 158 | }) 159 | 160 | return result_dict 161 | 162 | def predict(self, image_path: str, **kwargs): 163 | """ 164 | Conduct Optical Character Recognition (OCR) 165 | 166 | Args: 167 | image_path (str): the image file path 168 | detail (bool): if True, returned to include details. (bounding poly, vertices, etc) 169 | 170 | """ 171 | detail = kwargs.get("detail", False) 172 | 173 | return self._postprocess( 174 | self._model( 175 | image_path, 176 | skip_details=False, 177 | batch_size=1, 178 | paragraph=True, 179 | ), 180 | detail, 181 | ) 182 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/recognition.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from https://github.com/JaidedAI/EasyOCR/blob/8af936ba1b2f3c230968dc1022d0cd3e9ca1efbb/easyocr/recognition.py 3 | """ 4 | 5 | import math 6 | 7 | import numpy as np 8 | import torch 9 | import torch.nn.functional as F 10 | import torch.utils.data 11 | import torchvision.transforms as transforms 12 | from PIL import Image 13 | 14 | from .model import Model 15 | from .utils import CTCLabelConverter 16 | 17 | 18 | def contrast_grey(img): 19 | high = np.percentile(img, 90) 20 | low = np.percentile(img, 10) 21 | return (high - low) / np.maximum(10, high + low), high, low 22 | 23 | 24 | def adjust_contrast_grey(img, target: float = 0.4): 25 | contrast, high, low = contrast_grey(img) 26 | if contrast < target: 27 | img = img.astype(int) 28 | ratio = 200.0 / np.maximum(10, high - low) 29 | img = (img - low + 25) * ratio 30 | img = np.maximum( 31 | np.full(img.shape, 0), 32 | np.minimum( 33 | np.full(img.shape, 255), 34 | img, 35 | ), 36 | ).astype(np.uint8) 37 | return img 38 | 39 | 40 | class NormalizePAD(object): 41 | 42 | def __init__(self, max_size, PAD_type: str = "right"): 43 | self.toTensor = transforms.ToTensor() 44 | self.max_size = max_size 45 | self.max_width_half = math.floor(max_size[2] / 2) 46 | self.PAD_type = PAD_type 47 | 48 | def __call__(self, img): 49 | img = self.toTensor(img) 50 | img.sub_(0.5).div_(0.5) 51 | c, h, w = img.size() 52 | Pad_img = torch.FloatTensor(*self.max_size).fill_(0) 53 | Pad_img[:, :, :w] = img # right pad 54 | if self.max_size[2] != w: # add border Pad 55 | Pad_img[:, :, w:] = (img[:, :, w - 1].unsqueeze(2).expand( 56 | c, 57 | h, 58 | self.max_size[2] - w, 59 | )) 60 | 61 | return Pad_img 62 | 63 | 64 | class ListDataset(torch.utils.data.Dataset): 65 | 66 | def __init__(self, image_list: list): 67 | self.image_list = image_list 68 | self.nSamples = len(image_list) 69 | 70 | def __len__(self): 71 | return self.nSamples 72 | 73 | def __getitem__(self, index): 74 | img = self.image_list[index] 75 | return Image.fromarray(img, "L") 76 | 77 | 78 | class AlignCollate(object): 79 | 80 | def __init__(self, imgH: int, imgW: int, adjust_contrast: float): 81 | self.imgH = imgH 82 | self.imgW = imgW 83 | self.keep_ratio_with_pad = True # Do Not Change 84 | self.adjust_contrast = adjust_contrast 85 | 86 | def __call__(self, batch): 87 | batch = filter(lambda x: x is not None, batch) 88 | images = batch 89 | 90 | resized_max_w = self.imgW 91 | input_channel = 1 92 | transform = NormalizePAD((input_channel, self.imgH, resized_max_w)) 93 | 94 | resized_images = [] 95 | for image in images: 96 | w, h = image.size 97 | # augmentation here - change contrast 98 | if self.adjust_contrast > 0: 99 | image = np.array(image.convert("L")) 100 | image = adjust_contrast_grey(image, target=self.adjust_contrast) 101 | image = Image.fromarray(image, "L") 102 | 103 | ratio = w / float(h) 104 | if math.ceil(self.imgH * ratio) > self.imgW: 105 | resized_w = self.imgW 106 | else: 107 | resized_w = math.ceil(self.imgH * ratio) 108 | 109 | resized_image = image.resize((resized_w, self.imgH), Image.BICUBIC) 110 | resized_images.append(transform(resized_image)) 111 | 112 | image_tensors = torch.cat([t.unsqueeze(0) for t in resized_images], 0) 113 | return image_tensors 114 | 115 | 116 | def recognizer_predict(model, converter, test_loader, opt2val: dict): 117 | device = opt2val["device"] 118 | 119 | model.eval() 120 | result = [] 121 | with torch.no_grad(): 122 | for image_tensors in test_loader: 123 | batch_size = image_tensors.size(0) 124 | inputs = image_tensors.to(device) 125 | preds = model(inputs) # (N, length, num_classes) 126 | 127 | # rebalance 128 | preds_prob = F.softmax(preds, dim=2) 129 | preds_prob = preds_prob.cpu().detach().numpy() 130 | pred_norm = preds_prob.sum(axis=2) 131 | preds_prob = preds_prob / np.expand_dims(pred_norm, axis=-1) 132 | preds_prob = torch.from_numpy(preds_prob).float().to(device) 133 | 134 | # Select max probabilty (greedy decoding), then decode index to character 135 | preds_lengths = torch.IntTensor([preds.size(1)] * 136 | batch_size) # (N,) 137 | _, preds_indices = preds_prob.max(2) # (N, length) 138 | preds_indices = preds_indices.view(-1) # (N*length) 139 | preds_str = converter.decode_greedy(preds_indices, preds_lengths) 140 | 141 | preds_max_prob, _ = preds_prob.max(dim=2) 142 | 143 | for pred, pred_max_prob in zip(preds_str, preds_max_prob): 144 | confidence_score = pred_max_prob.cumprod(dim=0)[-1] 145 | result.append([pred, confidence_score.item()]) 146 | 147 | return result 148 | 149 | 150 | def get_recognizer(opt2val: dict): 151 | """ 152 | :return: 153 | recognizer: recognition net 154 | converter: CTCLabelConverter 155 | """ 156 | # converter 157 | vocab = opt2val["vocab"] 158 | converter = CTCLabelConverter(vocab) 159 | 160 | # recognizer 161 | recognizer = Model(opt2val) 162 | 163 | # state_dict 164 | rec_model_ckpt_fp = opt2val["rec_model_ckpt_fp"] 165 | device = opt2val["device"] 166 | state_dict = torch.load(rec_model_ckpt_fp, map_location=device) 167 | 168 | if device == "cuda": 169 | recognizer = torch.nn.DataParallel(recognizer).to(device) 170 | else: 171 | # TODO temporary: multigpu 학습한 뒤 ckpt loading 문제 172 | from collections import OrderedDict 173 | 174 | def _sync_tensor_name(state_dict): 175 | state_dict_ = OrderedDict() 176 | for name, val in state_dict.items(): 177 | name = name.replace("module.", "") 178 | state_dict_[name] = val 179 | return state_dict_ 180 | 181 | state_dict = _sync_tensor_name(state_dict) 182 | 183 | recognizer.load_state_dict(state_dict) 184 | 185 | return recognizer, converter 186 | 187 | 188 | def get_text(image_list, recognizer, converter, opt2val: dict): 189 | imgW = opt2val["imgW"] 190 | imgH = opt2val["imgH"] 191 | adjust_contrast = opt2val["adjust_contrast"] 192 | batch_size = opt2val["batch_size"] 193 | n_workers = opt2val["n_workers"] 194 | contrast_ths = opt2val["contrast_ths"] 195 | 196 | # TODO: figure out what is this for 197 | # batch_max_length = int(imgW / 10) 198 | 199 | coord = [item[0] for item in image_list] 200 | img_list = [item[1] for item in image_list] 201 | AlignCollate_normal = AlignCollate(imgH, imgW, adjust_contrast) 202 | test_data = ListDataset(img_list) 203 | test_loader = torch.utils.data.DataLoader( 204 | test_data, 205 | batch_size=batch_size, 206 | shuffle=False, 207 | num_workers=n_workers, 208 | collate_fn=AlignCollate_normal, 209 | pin_memory=True, 210 | ) 211 | 212 | # predict first round 213 | result1 = recognizer_predict(recognizer, converter, test_loader, opt2val) 214 | 215 | # predict second round 216 | low_confident_idx = [ 217 | i for i, item in enumerate(result1) if (item[1] < contrast_ths) 218 | ] 219 | if len(low_confident_idx) > 0: 220 | img_list2 = [img_list[i] for i in low_confident_idx] 221 | AlignCollate_contrast = AlignCollate(imgH, imgW, adjust_contrast) 222 | test_data = ListDataset(img_list2) 223 | test_loader = torch.utils.data.DataLoader( 224 | test_data, 225 | batch_size=batch_size, 226 | shuffle=False, 227 | num_workers=n_workers, 228 | collate_fn=AlignCollate_contrast, 229 | pin_memory=True, 230 | ) 231 | result2 = recognizer_predict(recognizer, converter, test_loader, 232 | opt2val) 233 | 234 | result = [] 235 | for i, zipped in enumerate(zip(coord, result1)): 236 | box, pred1 = zipped 237 | if i in low_confident_idx: 238 | pred2 = result2[low_confident_idx.index(i)] 239 | if pred1[1] > pred2[1]: 240 | result.append((box, pred1[0], pred1[1])) 241 | else: 242 | result.append((box, pred2[0], pred2[1])) 243 | else: 244 | result.append((box, pred1[0], pred1[1])) 245 | 246 | return result 247 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/brainocr.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is primarily based on the following: 3 | https://github.com/JaidedAI/EasyOCR/blob/8af936ba1b2f3c230968dc1022d0cd3e9ca1efbb/easyocr/easyocr.py 4 | 5 | Basic usage: 6 | >>> from pororo import Pororo 7 | >>> ocr = Pororo(task="ocr", lang="ko") 8 | >>> ocr("IMAGE_FILE") 9 | """ 10 | 11 | import ast 12 | from logging import getLogger 13 | from typing import List 14 | 15 | import cv2 16 | import numpy as np 17 | from PIL import Image 18 | 19 | from .detection import get_detector, get_textbox 20 | from .recognition import get_recognizer, get_text 21 | from .utils import ( 22 | diff, 23 | get_image_list, 24 | get_paragraph, 25 | group_text_box, 26 | reformat_input, 27 | ) 28 | 29 | LOGGER = getLogger(__name__) 30 | 31 | 32 | class Reader(object): 33 | 34 | def __init__( 35 | self, 36 | lang: str, 37 | det_model_ckpt_fp: str, 38 | rec_model_ckpt_fp: str, 39 | opt_fp: str, 40 | device: str, 41 | ) -> None: 42 | """ 43 | TODO @karter: modify this such that you download the pretrained checkpoint files 44 | Parameters: 45 | lang: language code. e.g, "en" or "ko" 46 | det_model_ckpt_fp: Detection model's checkpoint path e.g., 'craft_mlt_25k.pth' 47 | rec_model_ckpt_fp: Recognition model's checkpoint path 48 | opt_fp: option file path 49 | """ 50 | # Plug options in the dictionary 51 | opt2val = self.parse_options(opt_fp) # e.g., {"imgH": 64, ...} 52 | opt2val["vocab"] = self.build_vocab(opt2val["character"]) 53 | opt2val["vocab_size"] = len(opt2val["vocab"]) 54 | opt2val["device"] = device 55 | opt2val["lang"] = lang 56 | opt2val["det_model_ckpt_fp"] = det_model_ckpt_fp 57 | opt2val["rec_model_ckpt_fp"] = rec_model_ckpt_fp 58 | 59 | # Get model objects 60 | self.detector = get_detector(det_model_ckpt_fp, opt2val["device"]) 61 | self.recognizer, self.converter = get_recognizer(opt2val) 62 | self.opt2val = opt2val 63 | 64 | @staticmethod 65 | def parse_options(opt_fp: str) -> dict: 66 | opt2val = dict() 67 | for line in open(opt_fp, "r", encoding="utf8"): 68 | line = line.strip() 69 | if ": " in line: 70 | opt, val = line.split(": ", 1) 71 | try: 72 | opt2val[opt] = ast.literal_eval(val) 73 | except: 74 | opt2val[opt] = val 75 | 76 | return opt2val 77 | 78 | @staticmethod 79 | def build_vocab(character: str) -> List[str]: 80 | """Returns vocabulary (=list of characters)""" 81 | vocab = ["[blank]"] + list( 82 | character) # dummy '[blank]' token for CTCLoss (index 0) 83 | return vocab 84 | 85 | def detect(self, img: np.ndarray, opt2val: dict): 86 | """ 87 | :return: 88 | horizontal_list (list): e.g., [[613, 1496, 51, 190], [136, 1544, 134, 508]] 89 | free_list (list): e.g., [] 90 | """ 91 | text_box = get_textbox(self.detector, img, opt2val) 92 | horizontal_list, free_list = group_text_box( 93 | text_box, 94 | opt2val["slope_ths"], 95 | opt2val["ycenter_ths"], 96 | opt2val["height_ths"], 97 | opt2val["width_ths"], 98 | opt2val["add_margin"], 99 | ) 100 | 101 | min_size = opt2val["min_size"] 102 | if min_size: 103 | horizontal_list = [ 104 | i for i in horizontal_list 105 | if max(i[1] - i[0], i[3] - i[2]) > min_size 106 | ] 107 | free_list = [ 108 | i for i in free_list 109 | if max(diff([c[0] for c in i]), diff([c[1] 110 | for c in i])) > min_size 111 | ] 112 | 113 | return horizontal_list, free_list 114 | 115 | def recognize( 116 | self, 117 | img_cv_grey: np.ndarray, 118 | horizontal_list: list, 119 | free_list: list, 120 | opt2val: dict, 121 | ): 122 | """ 123 | Read text in the image 124 | :return: 125 | result (list): bounding box, text and confident score 126 | e.g., [([[189, 75], [469, 75], [469, 165], [189, 165]], '愚园路', 0.3754989504814148), 127 | ([[86, 80], [134, 80], [134, 128], [86, 128]], '西', 0.40452659130096436), 128 | ([[517, 81], [565, 81], [565, 123], [517, 123]], '东', 0.9989598989486694), 129 | ([[78, 126], [136, 126], [136, 156], [78, 156]], '315', 0.8125889301300049), 130 | ([[514, 126], [574, 126], [574, 156], [514, 156]], '309', 0.4971577227115631), 131 | ([[226, 170], [414, 170], [414, 220], [226, 220]], 'Yuyuan Rd.', 0.8261902332305908), 132 | ([[79, 173], [125, 173], [125, 213], [79, 213]], 'W', 0.9848111271858215), 133 | ([[529, 173], [569, 173], [569, 213], [529, 213]], 'E', 0.8405593633651733)] 134 | or list of texts (if skip_details is True) 135 | e.g., ['愚园路', '西', '东', '315', '309', 'Yuyuan Rd.', 'W', 'E'] 136 | """ 137 | imgH = opt2val["imgH"] 138 | paragraph = opt2val["paragraph"] 139 | skip_details = opt2val["skip_details"] 140 | 141 | if (horizontal_list is None) and (free_list is None): 142 | y_max, x_max = img_cv_grey.shape 143 | ratio = x_max / y_max 144 | max_width = int(imgH * ratio) 145 | crop_img = cv2.resize( 146 | img_cv_grey, 147 | (max_width, imgH), 148 | interpolation=Image.ANTIALIAS, 149 | ) 150 | image_list = [([[0, 0], [x_max, 0], [x_max, y_max], 151 | [0, y_max]], crop_img)] 152 | else: 153 | image_list, max_width = get_image_list( 154 | horizontal_list, 155 | free_list, 156 | img_cv_grey, 157 | model_height=imgH, 158 | ) 159 | 160 | result = get_text(image_list, self.recognizer, self.converter, opt2val) 161 | 162 | if paragraph: 163 | result = get_paragraph(result, mode="ltr") 164 | 165 | if skip_details: # texts only 166 | return [item[1] for item in result] 167 | else: # full outputs: bounding box, text and confident score 168 | return result 169 | 170 | def __call__( 171 | self, 172 | image, 173 | batch_size: int = 1, 174 | n_workers: int = 0, 175 | skip_details: bool = False, 176 | paragraph: bool = False, 177 | min_size: int = 20, 178 | contrast_ths: float = 0.1, 179 | adjust_contrast: float = 0.5, 180 | filter_ths: float = 0.003, 181 | text_threshold: float = 0.7, 182 | low_text: float = 0.4, 183 | link_threshold: float = 0.4, 184 | canvas_size: int = 2560, 185 | mag_ratio: float = 1.0, 186 | slope_ths: float = 0.1, 187 | ycenter_ths: float = 0.5, 188 | height_ths: float = 0.5, 189 | width_ths: float = 0.5, 190 | add_margin: float = 0.1, 191 | ): 192 | """ 193 | Detect text in the image and then recognize it. 194 | :param image: file path or numpy-array or a byte stream object 195 | :param batch_size: 196 | :param n_workers: 197 | :param skip_details: 198 | :param paragraph: 199 | :param min_size: 200 | :param contrast_ths: 201 | :param adjust_contrast: 202 | :param filter_ths: 203 | :param text_threshold: 204 | :param low_text: 205 | :param link_threshold: 206 | :param canvas_size: 207 | :param mag_ratio: 208 | :param slope_ths: 209 | :param ycenter_ths: 210 | :param height_ths: 211 | :param width_ths: 212 | :param add_margin: 213 | :return: 214 | """ 215 | # update `opt2val` 216 | self.opt2val["batch_size"] = batch_size 217 | self.opt2val["n_workers"] = n_workers 218 | self.opt2val["skip_details"] = skip_details 219 | self.opt2val["paragraph"] = paragraph 220 | self.opt2val["min_size"] = min_size 221 | self.opt2val["contrast_ths"] = contrast_ths 222 | self.opt2val["adjust_contrast"] = adjust_contrast 223 | self.opt2val["filter_ths"] = filter_ths 224 | self.opt2val["text_threshold"] = text_threshold 225 | self.opt2val["low_text"] = low_text 226 | self.opt2val["link_threshold"] = link_threshold 227 | self.opt2val["canvas_size"] = canvas_size 228 | self.opt2val["mag_ratio"] = mag_ratio 229 | self.opt2val["slope_ths"] = slope_ths 230 | self.opt2val["ycenter_ths"] = ycenter_ths 231 | self.opt2val["height_ths"] = height_ths 232 | self.opt2val["width_ths"] = width_ths 233 | self.opt2val["add_margin"] = add_margin 234 | 235 | img, img_cv_grey = reformat_input(image) # img, img_cv_grey: array 236 | 237 | horizontal_list, free_list = self.detect(img, self.opt2val) 238 | result = self.recognize( 239 | img_cv_grey, 240 | horizontal_list, 241 | free_list, 242 | self.opt2val, 243 | ) 244 | 245 | return result 246 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/transformation.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | 6 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 7 | 8 | 9 | class TpsSpatialTransformerNetwork(nn.Module): 10 | """ Rectification Network of RARE, namely TPS based STN """ 11 | 12 | def __init__(self, F, I_size, I_r_size, I_channel_num: int = 1): 13 | """Based on RARE TPS 14 | input: 15 | batch_I: Batch Input Image [batch_size x I_channel_num x I_height x I_width] 16 | I_size : (height, width) of the input image I 17 | I_r_size : (height, width) of the rectified image I_r 18 | I_channel_num : the number of channels of the input image I 19 | output: 20 | batch_I_r: rectified image [batch_size x I_channel_num x I_r_height x I_r_width] 21 | """ 22 | super(TpsSpatialTransformerNetwork, self).__init__() 23 | self.F = F 24 | self.I_size = I_size 25 | self.I_r_size = I_r_size # = (I_r_height, I_r_width) 26 | self.I_channel_num = I_channel_num 27 | self.LocalizationNetwork = LocalizationNetwork(self.F, 28 | self.I_channel_num) 29 | self.GridGenerator = GridGenerator(self.F, self.I_r_size) 30 | 31 | def forward(self, batch_I): 32 | batch_C_prime = self.LocalizationNetwork(batch_I) # batch_size x K x 2 33 | build_P_prime = self.GridGenerator.build_P_prime( 34 | batch_C_prime) # batch_size x n (= I_r_width x I_r_height) x 2 35 | build_P_prime_reshape = build_P_prime.reshape( 36 | [build_P_prime.size(0), self.I_r_size[0], self.I_r_size[1], 2]) 37 | 38 | # if torch.__version__ > "1.2.0": 39 | # batch_I_r = F.grid_sample(batch_I, build_P_prime_reshape, padding_mode='border', align_corners=True) 40 | # else: 41 | batch_I_r = F.grid_sample(batch_I, 42 | build_P_prime_reshape, 43 | padding_mode="border") 44 | 45 | return batch_I_r 46 | 47 | 48 | class LocalizationNetwork(nn.Module): 49 | """ Localization Network of RARE, which predicts C' (K x 2) from I (I_width x I_height) """ 50 | 51 | def __init__(self, F, I_channel_num: int): 52 | super(LocalizationNetwork, self).__init__() 53 | self.F = F 54 | self.I_channel_num = I_channel_num 55 | self.conv = nn.Sequential( 56 | nn.Conv2d( 57 | in_channels=self.I_channel_num, 58 | out_channels=64, 59 | kernel_size=3, 60 | stride=1, 61 | padding=1, 62 | bias=False, 63 | ), 64 | nn.BatchNorm2d(64), 65 | nn.ReLU(True), 66 | nn.MaxPool2d(2, 2), # batch_size x 64 x I_height/2 x I_width/2 67 | nn.Conv2d(64, 128, 3, 1, 1, bias=False), 68 | nn.BatchNorm2d(128), 69 | nn.ReLU(True), 70 | nn.MaxPool2d(2, 2), # batch_size x 128 x I_height/4 x I_width/4 71 | nn.Conv2d(128, 256, 3, 1, 1, bias=False), 72 | nn.BatchNorm2d(256), 73 | nn.ReLU(True), 74 | nn.MaxPool2d(2, 2), # batch_size x 256 x I_height/8 x I_width/8 75 | nn.Conv2d(256, 512, 3, 1, 1, bias=False), 76 | nn.BatchNorm2d(512), 77 | nn.ReLU(True), 78 | nn.AdaptiveAvgPool2d(1), # batch_size x 512 79 | ) 80 | 81 | self.localization_fc1 = nn.Sequential(nn.Linear(512, 256), 82 | nn.ReLU(True)) 83 | self.localization_fc2 = nn.Linear(256, self.F * 2) 84 | 85 | # Init fc2 in LocalizationNetwork 86 | self.localization_fc2.weight.data.fill_(0) 87 | 88 | # see RARE paper Fig. 6 (a) 89 | ctrl_pts_x = np.linspace(-1.0, 1.0, int(F / 2)) 90 | ctrl_pts_y_top = np.linspace(0.0, -1.0, num=int(F / 2)) 91 | ctrl_pts_y_bottom = np.linspace(1.0, 0.0, num=int(F / 2)) 92 | ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1) 93 | ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1) 94 | initial_bias = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0) 95 | self.localization_fc2.bias.data = ( 96 | torch.from_numpy(initial_bias).float().view(-1)) 97 | 98 | def forward(self, batch_I): 99 | """ 100 | :param batch_I : Batch Input Image [batch_size x I_channel_num x I_height x I_width] 101 | :return: batch_C_prime : Predicted coordinates of fiducial points for input batch [batch_size x F x 2] 102 | """ 103 | batch_size = batch_I.size(0) 104 | features = self.conv(batch_I).view(batch_size, -1) 105 | batch_C_prime = self.localization_fc2( 106 | self.localization_fc1(features)).view(batch_size, self.F, 2) 107 | return batch_C_prime 108 | 109 | 110 | class GridGenerator(nn.Module): 111 | """ Grid Generator of RARE, which produces P_prime by multipling T with P """ 112 | 113 | def __init__(self, F, I_r_size): 114 | """ Generate P_hat and inv_delta_C for later """ 115 | super(GridGenerator, self).__init__() 116 | self.eps = 1e-6 117 | self.I_r_height, self.I_r_width = I_r_size 118 | self.F = F 119 | self.C = self._build_C(self.F) # F x 2 120 | self.P = self._build_P(self.I_r_width, self.I_r_height) 121 | 122 | # for multi-gpu, you need register buffer 123 | self.register_buffer( 124 | "inv_delta_C", 125 | torch.tensor(self._build_inv_delta_C(self.F, 126 | self.C)).float()) # F+3 x F+3 127 | self.register_buffer("P_hat", 128 | torch.tensor( 129 | self._build_P_hat(self.F, self.C, 130 | self.P)).float()) # n x F+3 131 | 132 | def _build_C(self, F): 133 | """ Return coordinates of fiducial points in I_r; C """ 134 | ctrl_pts_x = np.linspace(-1.0, 1.0, int(F / 2)) 135 | ctrl_pts_y_top = -1 * np.ones(int(F / 2)) 136 | ctrl_pts_y_bottom = np.ones(int(F / 2)) 137 | ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1) 138 | ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1) 139 | C = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0) 140 | return C # F x 2 141 | 142 | def _build_inv_delta_C(self, F, C): 143 | """ Return inv_delta_C which is needed to calculate T """ 144 | hat_C = np.zeros((F, F), dtype=float) # F x F 145 | for i in range(0, F): 146 | for j in range(i, F): 147 | r = np.linalg.norm(C[i] - C[j]) 148 | hat_C[i, j] = r 149 | hat_C[j, i] = r 150 | np.fill_diagonal(hat_C, 1) 151 | hat_C = (hat_C**2) * np.log(hat_C) 152 | # print(C.shape, hat_C.shape) 153 | delta_C = np.concatenate( # F+3 x F+3 154 | [ 155 | np.concatenate([np.ones((F, 1)), C, hat_C], axis=1), # F x F+3 156 | np.concatenate([np.zeros( 157 | (2, 3)), np.transpose(C)], axis=1), # 2 x F+3 158 | np.concatenate([np.zeros( 159 | (1, 3)), np.ones((1, F))], axis=1), # 1 x F+3 160 | ], 161 | axis=0, 162 | ) 163 | inv_delta_C = np.linalg.inv(delta_C) 164 | return inv_delta_C # F+3 x F+3 165 | 166 | def _build_P(self, I_r_width, I_r_height): 167 | I_r_grid_x = (np.arange(-I_r_width, I_r_width, 2) + 168 | 1.0) / I_r_width # self.I_r_width 169 | I_r_grid_y = (np.arange(-I_r_height, I_r_height, 2) + 170 | 1.0) / I_r_height # self.I_r_height 171 | P = np.stack( # self.I_r_width x self.I_r_height x 2 172 | np.meshgrid(I_r_grid_x, I_r_grid_y), 173 | axis=2) 174 | return P.reshape([-1, 2]) # n (= self.I_r_width x self.I_r_height) x 2 175 | 176 | def _build_P_hat(self, F, C, P): 177 | n = P.shape[0] # n (= self.I_r_width x self.I_r_height) 178 | P_tile = np.tile(np.expand_dims(P, axis=1), 179 | (1, F, 1)) # n x 2 -> n x 1 x 2 -> n x F x 2 180 | C_tile = np.expand_dims(C, axis=0) # 1 x F x 2 181 | P_diff = P_tile - C_tile # n x F x 2 182 | rbf_norm = np.linalg.norm(P_diff, ord=2, axis=2, 183 | keepdims=False) # n x F 184 | rbf = np.multiply(np.square(rbf_norm), 185 | np.log(rbf_norm + self.eps)) # n x F 186 | P_hat = np.concatenate([np.ones((n, 1)), P, rbf], axis=1) 187 | return P_hat # n x F+3 188 | 189 | def build_P_prime(self, batch_C_prime): 190 | """ Generate Grid from batch_C_prime [batch_size x F x 2] """ 191 | batch_size = batch_C_prime.size(0) 192 | batch_inv_delta_C = self.inv_delta_C.repeat(batch_size, 1, 1) 193 | batch_P_hat = self.P_hat.repeat(batch_size, 1, 1) 194 | batch_C_prime_with_zeros = torch.cat( 195 | (batch_C_prime, torch.zeros(batch_size, 3, 2).float().to(device)), 196 | dim=1) # batch_size x F+3 x 2 197 | batch_T = torch.bmm(batch_inv_delta_C, 198 | batch_C_prime_with_zeros) # batch_size x F+3 x 2 199 | batch_P_prime = torch.bmm(batch_P_hat, batch_T) # batch_size x n x 2 200 | return batch_P_prime # batch_size x n x 2 201 | -------------------------------------------------------------------------------- /pororo/tasks/utils/download_utils.py: -------------------------------------------------------------------------------- 1 | """Module download related function from. Tenth""" 2 | 3 | import logging 4 | import os 5 | import platform 6 | import sys 7 | import zipfile 8 | from dataclasses import dataclass 9 | from typing import Tuple, Union 10 | 11 | import wget 12 | 13 | from pororo.tasks.utils.config import CONFIGS 14 | 15 | DEFAULT_PREFIX = { 16 | "model": "https://twg.kakaocdn.net/pororo/{lang}/models", 17 | "dict": "https://twg.kakaocdn.net/pororo/{lang}/dicts", 18 | } 19 | 20 | 21 | @dataclass 22 | class TransformerInfo: 23 | r"Dataclass for transformer-based model" 24 | path: str 25 | dict_path: str 26 | src_dict: str 27 | tgt_dict: str 28 | src_tok: Union[str, None] 29 | tgt_tok: Union[str, None] 30 | 31 | 32 | @dataclass 33 | class DownloadInfo: 34 | r"Download information such as defined directory, language and model name" 35 | n_model: str 36 | lang: str 37 | root_dir: str 38 | 39 | 40 | def get_save_dir(save_dir: str = None) -> str: 41 | """ 42 | Get default save directory 43 | 44 | Args: 45 | savd_dir(str): User-defined save directory 46 | 47 | Returns: 48 | str: Set save directory 49 | 50 | """ 51 | # If user wants to manually define save directory 52 | if save_dir: 53 | os.makedirs(save_dir, exist_ok=True) 54 | return save_dir 55 | 56 | pf = platform.system() 57 | 58 | if pf == "Windows": 59 | save_dir = "C:\\pororo" 60 | else: 61 | home_dir = os.path.expanduser("~") 62 | save_dir = os.path.join(home_dir, ".pororo") 63 | 64 | if not os.path.exists(save_dir): 65 | os.makedirs(save_dir) 66 | 67 | return save_dir 68 | 69 | 70 | def get_download_url(n_model: str, key: str, lang: str) -> str: 71 | """ 72 | Get download url using default prefix 73 | 74 | Args: 75 | n_model (str): model name 76 | key (str): key name either `model` or `dict` 77 | lang (str): language name 78 | 79 | Returns: 80 | str: generated download url 81 | 82 | """ 83 | default_prefix = DEFAULT_PREFIX[key].format(lang=lang) 84 | return f"{default_prefix}/{n_model}" 85 | 86 | 87 | def download_or_load_bert(info: DownloadInfo) -> str: 88 | """ 89 | Download fine-tuned BrainBert & BrainSBert model and dict 90 | 91 | Args: 92 | info (DownloadInfo): download information 93 | 94 | Returns: 95 | str: downloaded bert & sbert path 96 | 97 | """ 98 | model_path = os.path.join(info.root_dir, info.n_model) 99 | 100 | if not os.path.exists(model_path): 101 | info.n_model += ".zip" 102 | zip_path = os.path.join(info.root_dir, info.n_model) 103 | 104 | type_dir = download_from_url( 105 | info.n_model, 106 | zip_path, 107 | key="model", 108 | lang=info.lang, 109 | ) 110 | 111 | zip_file = zipfile.ZipFile(zip_path) 112 | zip_file.extractall(type_dir) 113 | zip_file.close() 114 | 115 | return model_path 116 | 117 | 118 | def download_or_load_transformer(info: DownloadInfo) -> TransformerInfo: 119 | """ 120 | Download pre-trained Transformer model and corresponding dict 121 | 122 | Args: 123 | info (DownloadInfo): download information 124 | 125 | Returns: 126 | TransformerInfo: information dataclass for transformer construction 127 | 128 | """ 129 | config = CONFIGS[info.n_model.split("/")[-1]] 130 | 131 | src_dict_in = config.src_dict 132 | tgt_dict_in = config.tgt_dict 133 | src_tok = config.src_tok 134 | tgt_tok = config.tgt_tok 135 | 136 | info.n_model += ".pt" 137 | model_path = os.path.join(info.root_dir, info.n_model) 138 | 139 | # Download or load Transformer model 140 | model_type_dir = "/".join(model_path.split("/")[:-1]) 141 | if not os.path.exists(model_path): 142 | model_type_dir = download_from_url( 143 | info.n_model, 144 | model_path, 145 | key="model", 146 | lang=info.lang, 147 | ) 148 | 149 | dict_type_dir = str() 150 | src_dict, tgt_dict = str(), str() 151 | 152 | # Download or load corresponding dictionary 153 | if src_dict_in: 154 | src_dict = f"{src_dict_in}.txt" 155 | src_dict_path = os.path.join(info.root_dir, f"dicts/{src_dict}") 156 | dict_type_dir = "/".join(src_dict_path.split("/")[:-1]) 157 | if not os.path.exists(src_dict_path): 158 | dict_type_dir = download_from_url( 159 | src_dict, 160 | src_dict_path, 161 | key="dict", 162 | lang=info.lang, 163 | ) 164 | 165 | if tgt_dict_in: 166 | tgt_dict = f"{tgt_dict_in}.txt" 167 | tgt_dict_path = os.path.join(info.root_dir, f"dicts/{tgt_dict}") 168 | if not os.path.exists(tgt_dict_path): 169 | download_from_url( 170 | tgt_dict, 171 | tgt_dict_path, 172 | key="dict", 173 | lang=info.lang, 174 | ) 175 | 176 | # Download or load corresponding tokenizer 177 | src_tok_path, tgt_tok_path = None, None 178 | if src_tok: 179 | src_tok_path = download_or_load( 180 | f"tokenizers/{src_tok}.zip", 181 | lang=info.lang, 182 | ) 183 | if tgt_tok: 184 | tgt_tok_path = download_or_load( 185 | f"tokenizers/{tgt_tok}.zip", 186 | lang=info.lang, 187 | ) 188 | 189 | return TransformerInfo( 190 | path=model_type_dir, 191 | dict_path=dict_type_dir, 192 | # Drop prefix "dict." and postfix ".txt" 193 | src_dict=".".join(src_dict.split(".")[1:-1]), 194 | # to follow fairseq's dictionary load process 195 | tgt_dict=".".join(tgt_dict.split(".")[1:-1]), 196 | src_tok=src_tok_path, 197 | tgt_tok=tgt_tok_path, 198 | ) 199 | 200 | 201 | def download_or_load_misc(info: DownloadInfo) -> str: 202 | """ 203 | Download (pre-trained) miscellaneous model 204 | 205 | Args: 206 | info (DownloadInfo): download information 207 | 208 | Returns: 209 | str: miscellaneous model path 210 | 211 | """ 212 | # Add postfix <.model> for sentencepiece 213 | if "sentencepiece" in info.n_model: 214 | info.n_model += ".model" 215 | 216 | # Generate target model path using root directory 217 | model_path = os.path.join(info.root_dir, info.n_model) 218 | if not os.path.exists(model_path): 219 | type_dir = download_from_url( 220 | info.n_model, 221 | model_path, 222 | key="model", 223 | lang=info.lang, 224 | ) 225 | 226 | if ".zip" in info.n_model: 227 | zip_file = zipfile.ZipFile(model_path) 228 | zip_file.extractall(type_dir) 229 | zip_file.close() 230 | 231 | if ".zip" in info.n_model: 232 | model_path = model_path[:model_path.rfind(".zip")] 233 | return model_path 234 | 235 | 236 | def download_or_load_bart(info: DownloadInfo) -> Union[str, Tuple[str, str]]: 237 | """ 238 | Download BART model 239 | 240 | Args: 241 | info (DownloadInfo): download information 242 | 243 | Returns: 244 | Union[str, Tuple[str, str]]: BART model path (with. corresponding SentencePiece) 245 | 246 | """ 247 | info.n_model += ".pt" 248 | 249 | model_path = os.path.join(info.root_dir, info.n_model) 250 | if not os.path.exists(model_path): 251 | download_from_url( 252 | info.n_model, 253 | model_path, 254 | key="model", 255 | lang=info.lang, 256 | ) 257 | 258 | return model_path 259 | 260 | 261 | def download_from_url( 262 | n_model: str, 263 | model_path: str, 264 | key: str, 265 | lang: str, 266 | ) -> str: 267 | """ 268 | Download specified model from Tenth 269 | 270 | Args: 271 | n_model (str): model name 272 | model_path (str): pre-defined model path 273 | key (str): type key (either model or dict) 274 | lang (str): language name 275 | 276 | Returns: 277 | str: default type directory 278 | 279 | """ 280 | # Get default type dir path 281 | type_dir = "/".join(model_path.split("/")[:-1]) 282 | os.makedirs(type_dir, exist_ok=True) 283 | 284 | # Get download tenth url 285 | url = get_download_url(n_model, key=key, lang=lang) 286 | 287 | logging.info("Downloading user-selected model...") 288 | wget.download(url, type_dir) 289 | sys.stderr.write("\n") 290 | sys.stderr.flush() 291 | 292 | return type_dir 293 | 294 | 295 | def download_or_load( 296 | n_model: str, 297 | lang: str, 298 | custom_save_dir: str = None, 299 | ) -> Union[TransformerInfo, str, Tuple[str, str]]: 300 | """ 301 | Download or load model based on model information 302 | 303 | Args: 304 | n_model (str): model name 305 | lang (str): language information 306 | custom_save_dir (str, optional): user-defined save directory path. defaults to None. 307 | 308 | Returns: 309 | Union[TransformerInfo, str, Tuple[str, str]] 310 | 311 | """ 312 | root_dir = get_save_dir(save_dir=custom_save_dir) 313 | info = DownloadInfo(n_model, lang, root_dir) 314 | 315 | if "transformer" in n_model: 316 | return download_or_load_transformer(info) 317 | if "bert" in n_model: 318 | return download_or_load_bert(info) 319 | if "bart" in n_model and "bpe" not in n_model: 320 | return download_or_load_bart(info) 321 | 322 | return download_or_load_misc(info) 323 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/craft_utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from avhttps://github.com/clovaai/CRAFT-pytorch/blob/master/craft_utils.py 3 | MIT License 4 | """ 5 | 6 | import math 7 | 8 | import cv2 9 | import numpy as np 10 | 11 | 12 | def warp_coord(Minv, pt): 13 | """auxilary functions: unwarp corodinates: """ 14 | out = np.matmul(Minv, (pt[0], pt[1], 1)) 15 | return np.array([out[0] / out[2], out[1] / out[2]]) 16 | 17 | 18 | def get_det_boxes_core(textmap, linkmap, text_threshold, link_threshold, 19 | low_text): 20 | # prepare data 21 | linkmap = linkmap.copy() 22 | textmap = textmap.copy() 23 | img_h, img_w = textmap.shape 24 | 25 | # labeling method 26 | ret, text_score = cv2.threshold(textmap, low_text, 1, 0) 27 | ret, link_score = cv2.threshold(linkmap, link_threshold, 1, 0) 28 | 29 | text_score_comb = np.clip(text_score + link_score, 0, 1) 30 | nLabels, labels, stats, centroids = cv2.connectedComponentsWithStats( 31 | text_score_comb.astype(np.uint8), connectivity=4) 32 | 33 | det = [] 34 | mapper = [] 35 | for k in range(1, nLabels): 36 | # size filtering 37 | size = stats[k, cv2.CC_STAT_AREA] 38 | if size < 10: 39 | continue 40 | 41 | # thresholding 42 | if np.max(textmap[labels == k]) < text_threshold: 43 | continue 44 | 45 | # make segmentation map 46 | segmap = np.zeros(textmap.shape, dtype=np.uint8) 47 | segmap[labels == k] = 255 48 | segmap[np.logical_and(link_score == 1, 49 | text_score == 0)] = 0 # remove link area 50 | x, y = stats[k, cv2.CC_STAT_LEFT], stats[k, cv2.CC_STAT_TOP] 51 | w, h = stats[k, cv2.CC_STAT_WIDTH], stats[k, cv2.CC_STAT_HEIGHT] 52 | niter = int(math.sqrt(size * min(w, h) / (w * h)) * 2) 53 | sx, ex, sy, ey = x - niter, x + w + niter + 1, y - niter, y + h + niter + 1 54 | # boundary check 55 | if sx < 0: 56 | sx = 0 57 | if sy < 0: 58 | sy = 0 59 | if ex >= img_w: 60 | ex = img_w 61 | if ey >= img_h: 62 | ey = img_h 63 | kernel = cv2.getStructuringElement( 64 | cv2.MORPH_RECT, 65 | (1 + niter, 1 + niter), 66 | ) 67 | segmap[sy:ey, sx:ex] = cv2.dilate(segmap[sy:ey, sx:ex], kernel) 68 | 69 | # make box 70 | np_contours = (np.roll(np.array(np.where(segmap != 0)), 1, 71 | axis=0).transpose().reshape(-1, 2)) 72 | rectangle = cv2.minAreaRect(np_contours) 73 | box = cv2.boxPoints(rectangle) 74 | 75 | # align diamond-shape 76 | w, h = np.linalg.norm(box[0] - box[1]), np.linalg.norm(box[1] - box[2]) 77 | box_ratio = max(w, h) / (min(w, h) + 1e-5) 78 | if abs(1 - box_ratio) <= 0.1: 79 | l, r = min(np_contours[:, 0]), max(np_contours[:, 0]) 80 | t, b = min(np_contours[:, 1]), max(np_contours[:, 1]) 81 | box = np.array([[l, t], [r, t], [r, b], [l, b]], dtype=np.float32) 82 | 83 | # make clock-wise order 84 | startidx = box.sum(axis=1).argmin() 85 | box = np.roll(box, 4 - startidx, 0) 86 | box = np.array(box) 87 | 88 | det.append(box) 89 | mapper.append(k) 90 | 91 | return det, labels, mapper 92 | 93 | 94 | def get_poly_core(boxes, labels, mapper, linkmap): 95 | # configs 96 | num_cp = 5 97 | max_len_ratio = 0.7 98 | expand_ratio = 1.45 99 | max_r = 2.0 100 | step_r = 0.2 101 | 102 | polys = [] 103 | for k, box in enumerate(boxes): 104 | # size filter for small instance 105 | w, h = int(np.linalg.norm(box[0] - box[1]) + 106 | 1), int(np.linalg.norm(box[1] - box[2]) + 1) 107 | if w < 10 or h < 10: 108 | polys.append(None) 109 | continue 110 | 111 | # warp image 112 | tar = np.float32([[0, 0], [w, 0], [w, h], [0, h]]) 113 | M = cv2.getPerspectiveTransform(box, tar) 114 | word_label = cv2.warpPerspective( 115 | labels, 116 | M, 117 | (w, h), 118 | flags=cv2.INTER_NEAREST, 119 | ) 120 | try: 121 | Minv = np.linalg.inv(M) 122 | except: 123 | polys.append(None) 124 | continue 125 | 126 | # binarization for selected label 127 | cur_label = mapper[k] 128 | word_label[word_label != cur_label] = 0 129 | word_label[word_label > 0] = 1 130 | 131 | # Polygon generation: find top/bottom contours 132 | cp = [] 133 | max_len = -1 134 | for i in range(w): 135 | region = np.where(word_label[:, i] != 0)[0] 136 | if len(region) < 2: 137 | continue 138 | cp.append((i, region[0], region[-1])) 139 | length = region[-1] - region[0] + 1 140 | if length > max_len: 141 | max_len = length 142 | 143 | # pass if max_len is similar to h 144 | if h * max_len_ratio < max_len: 145 | polys.append(None) 146 | continue 147 | 148 | # get pivot points with fixed length 149 | tot_seg = num_cp * 2 + 1 150 | seg_w = w / tot_seg # segment width 151 | pp = [None] * num_cp # init pivot points 152 | cp_section = [[0, 0]] * tot_seg 153 | seg_height = [0] * num_cp 154 | seg_num = 0 155 | num_sec = 0 156 | prev_h = -1 157 | for i in range(0, len(cp)): 158 | (x, sy, ey) = cp[i] 159 | if (seg_num + 1) * seg_w <= x and seg_num <= tot_seg: 160 | # average previous segment 161 | if num_sec == 0: 162 | break 163 | cp_section[seg_num] = [ 164 | cp_section[seg_num][0] / num_sec, 165 | cp_section[seg_num][1] / num_sec, 166 | ] 167 | num_sec = 0 168 | 169 | # reset variables 170 | seg_num += 1 171 | prev_h = -1 172 | 173 | # accumulate center points 174 | cy = (sy + ey) * 0.5 175 | cur_h = ey - sy + 1 176 | cp_section[seg_num] = [ 177 | cp_section[seg_num][0] + x, 178 | cp_section[seg_num][1] + cy, 179 | ] 180 | num_sec += 1 181 | 182 | if seg_num % 2 == 0: 183 | continue # No polygon area 184 | 185 | if prev_h < cur_h: 186 | pp[int((seg_num - 1) / 2)] = (x, cy) 187 | seg_height[int((seg_num - 1) / 2)] = cur_h 188 | prev_h = cur_h 189 | 190 | # processing last segment 191 | if num_sec != 0: 192 | cp_section[-1] = [ 193 | cp_section[-1][0] / num_sec, cp_section[-1][1] / num_sec 194 | ] 195 | 196 | # pass if num of pivots is not sufficient or segment width is smaller than character height 197 | if None in pp or seg_w < np.max(seg_height) * 0.25: 198 | polys.append(None) 199 | continue 200 | 201 | # calc median maximum of pivot points 202 | half_char_h = np.median(seg_height) * expand_ratio / 2 203 | 204 | # calc gradiant and apply to make horizontal pivots 205 | new_pp = [] 206 | for i, (x, cy) in enumerate(pp): 207 | dx = cp_section[i * 2 + 2][0] - cp_section[i * 2][0] 208 | dy = cp_section[i * 2 + 2][1] - cp_section[i * 2][1] 209 | if dx == 0: # gradient if zero 210 | new_pp.append([x, cy - half_char_h, x, cy + half_char_h]) 211 | continue 212 | rad = -math.atan2(dy, dx) 213 | c, s = half_char_h * math.cos(rad), half_char_h * math.sin(rad) 214 | new_pp.append([x - s, cy - c, x + s, cy + c]) 215 | 216 | # get edge points to cover character heatmaps 217 | isSppFound, isEppFound = False, False 218 | grad_s = (pp[1][1] - pp[0][1]) / (pp[1][0] - pp[0][0]) + ( 219 | pp[2][1] - pp[1][1]) / (pp[2][0] - pp[1][0]) 220 | grad_e = (pp[-2][1] - pp[-1][1]) / (pp[-2][0] - pp[-1][0]) + ( 221 | pp[-3][1] - pp[-2][1]) / (pp[-3][0] - pp[-2][0]) 222 | for r in np.arange(0.5, max_r, step_r): 223 | dx = 2 * half_char_h * r 224 | if not isSppFound: 225 | line_img = np.zeros(word_label.shape, dtype=np.uint8) 226 | dy = grad_s * dx 227 | p = np.array(new_pp[0]) - np.array([dx, dy, dx, dy]) 228 | cv2.line( 229 | line_img, 230 | (int(p[0]), int(p[1])), 231 | (int(p[2]), int(p[3])), 232 | 1, 233 | thickness=1, 234 | ) 235 | if (np.sum(np.logical_and(word_label, line_img)) == 0 or 236 | r + 2 * step_r >= max_r): 237 | spp = p 238 | isSppFound = True 239 | if not isEppFound: 240 | line_img = np.zeros(word_label.shape, dtype=np.uint8) 241 | dy = grad_e * dx 242 | p = np.array(new_pp[-1]) + np.array([dx, dy, dx, dy]) 243 | cv2.line( 244 | line_img, 245 | (int(p[0]), int(p[1])), 246 | (int(p[2]), int(p[3])), 247 | 1, 248 | thickness=1, 249 | ) 250 | if (np.sum(np.logical_and(word_label, line_img)) == 0 or 251 | r + 2 * step_r >= max_r): 252 | epp = p 253 | isEppFound = True 254 | if isSppFound and isEppFound: 255 | break 256 | 257 | # pass if boundary of polygon is not found 258 | if not (isSppFound and isEppFound): 259 | polys.append(None) 260 | continue 261 | 262 | # make final polygon 263 | poly = [] 264 | poly.append(warp_coord(Minv, (spp[0], spp[1]))) 265 | for p in new_pp: 266 | poly.append(warp_coord(Minv, (p[0], p[1]))) 267 | poly.append(warp_coord(Minv, (epp[0], epp[1]))) 268 | poly.append(warp_coord(Minv, (epp[2], epp[3]))) 269 | for p in reversed(new_pp): 270 | poly.append(warp_coord(Minv, (p[2], p[3]))) 271 | poly.append(warp_coord(Minv, (spp[2], spp[3]))) 272 | 273 | # add to final result 274 | polys.append(np.array(poly)) 275 | 276 | return polys 277 | 278 | 279 | def get_det_boxes( 280 | textmap, 281 | linkmap, 282 | text_threshold, 283 | link_threshold, 284 | low_text, 285 | poly=False, 286 | ): 287 | boxes, labels, mapper = get_det_boxes_core( 288 | textmap, 289 | linkmap, 290 | text_threshold, 291 | link_threshold, 292 | low_text, 293 | ) 294 | 295 | if poly: 296 | polys = get_poly_core(boxes, labels, mapper, linkmap) 297 | else: 298 | polys = [None] * len(boxes) 299 | 300 | return boxes, polys 301 | 302 | 303 | def adjust_result_coordinates(polys, ratio_w, ratio_h, ratio_net=2): 304 | if len(polys) > 0: 305 | polys = np.array(polys) 306 | for k in range(len(polys)): 307 | if polys[k] is not None: 308 | polys[k] *= (ratio_w * ratio_net, ratio_h * ratio_net) 309 | return polys 310 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/modules/feature_extraction.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | 4 | class VGGFeatureExtractor(nn.Module): 5 | """ FeatureExtractor of CRNN (https://arxiv.org/pdf/1507.05717.pdf) """ 6 | 7 | def __init__(self, 8 | n_input_channels: int = 1, 9 | n_output_channels: int = 512, 10 | opt2val=None): 11 | super(VGGFeatureExtractor, self).__init__() 12 | 13 | self.output_channel = [ 14 | int(n_output_channels / 8), 15 | int(n_output_channels / 4), 16 | int(n_output_channels / 2), 17 | n_output_channels, 18 | ] # [64, 128, 256, 512] 19 | 20 | rec_model_ckpt_fp = opt2val["rec_model_ckpt_fp"] 21 | if "baseline" in rec_model_ckpt_fp: 22 | self.ConvNet = nn.Sequential( 23 | nn.Conv2d(n_input_channels, self.output_channel[0], 3, 1, 1), 24 | nn.ReLU(True), 25 | nn.MaxPool2d(2, 2), # 64x16x50 26 | nn.Conv2d(self.output_channel[0], self.output_channel[1], 3, 1, 27 | 1), 28 | nn.ReLU(True), 29 | nn.MaxPool2d(2, 2), # 128x8x25 30 | nn.Conv2d(self.output_channel[1], self.output_channel[2], 3, 1, 31 | 1), 32 | nn.ReLU(True), # 256x8x25 33 | nn.Conv2d(self.output_channel[2], self.output_channel[2], 3, 1, 34 | 1), 35 | nn.ReLU(True), 36 | nn.MaxPool2d((2, 1), (2, 1)), # 256x4x25 37 | nn.Conv2d(self.output_channel[2], 38 | self.output_channel[3], 39 | 3, 40 | 1, 41 | 1, 42 | bias=False), 43 | nn.BatchNorm2d(self.output_channel[3]), 44 | nn.ReLU(True), # 512x4x25 45 | nn.Conv2d(self.output_channel[3], 46 | self.output_channel[3], 47 | 3, 48 | 1, 49 | 1, 50 | bias=False), 51 | nn.BatchNorm2d(self.output_channel[3]), 52 | nn.ReLU(True), 53 | nn.MaxPool2d((2, 1), (2, 1)), # 512x2x25 54 | # nn.Conv2d(self.output_channel[3], self.output_channel[3], 2, 1, 0), nn.ReLU(True)) # 512x1x24 55 | nn.ConvTranspose2d(self.output_channel[3], 56 | self.output_channel[3], 2, 2), 57 | nn.ReLU(True), 58 | ) # 512x4x50 59 | else: 60 | self.ConvNet = nn.Sequential( 61 | nn.Conv2d(n_input_channels, self.output_channel[0], 3, 1, 1), 62 | nn.ReLU(True), 63 | nn.MaxPool2d(2, 2), # 64x16x50 64 | nn.Conv2d(self.output_channel[0], self.output_channel[1], 3, 1, 65 | 1), 66 | nn.ReLU(True), 67 | nn.MaxPool2d(2, 2), # 128x8x25 68 | nn.Conv2d(self.output_channel[1], self.output_channel[2], 3, 1, 69 | 1), 70 | nn.ReLU(True), # 256x8x25 71 | nn.Conv2d(self.output_channel[2], self.output_channel[2], 3, 1, 72 | 1), 73 | nn.ReLU(True), 74 | nn.MaxPool2d((2, 1), (2, 1)), # 256x4x25 75 | nn.Conv2d(self.output_channel[2], 76 | self.output_channel[3], 77 | 3, 78 | 1, 79 | 1, 80 | bias=False), 81 | nn.BatchNorm2d(self.output_channel[3]), 82 | nn.ReLU(True), # 512x4x25 83 | nn.Conv2d(self.output_channel[3], 84 | self.output_channel[3], 85 | 3, 86 | 1, 87 | 1, 88 | bias=False), 89 | nn.BatchNorm2d(self.output_channel[3]), 90 | nn.ReLU(True), 91 | nn.MaxPool2d((2, 1), (2, 1)), # 512x2x25 92 | # nn.Conv2d(self.output_channel[3], self.output_channel[3], 2, 1, 0), nn.ReLU(True)) # 512x1x24 93 | nn.ConvTranspose2d(self.output_channel[3], 94 | self.output_channel[3], 2, 2), 95 | nn.ReLU(True), # 512x4x50 96 | nn.ConvTranspose2d(self.output_channel[3], 97 | self.output_channel[3], 2, 2), 98 | nn.ReLU(True), 99 | ) # 512x4x50 100 | 101 | def forward(self, x): 102 | return self.ConvNet(x) 103 | 104 | 105 | class ResNetFeatureExtractor(nn.Module): 106 | """ 107 | FeatureExtractor of FAN 108 | (http://openaccess.thecvf.com/content_ICCV_2017/papers/Cheng_Focusing_Attention_Towards_ICCV_2017_paper.pdf) 109 | """ 110 | 111 | def __init__(self, n_input_channels: int = 1, n_output_channels: int = 512): 112 | super(ResNetFeatureExtractor, self).__init__() 113 | self.ConvNet = ResNet(n_input_channels, n_output_channels, BasicBlock, 114 | [1, 2, 5, 3]) 115 | 116 | def forward(self, inputs): 117 | return self.ConvNet(inputs) 118 | 119 | 120 | class BasicBlock(nn.Module): 121 | expansion = 1 122 | 123 | def __init__(self, 124 | inplanes: int, 125 | planes: int, 126 | stride: int = 1, 127 | downsample=None): 128 | super(BasicBlock, self).__init__() 129 | self.conv1 = self._conv3x3(inplanes, planes) 130 | self.bn1 = nn.BatchNorm2d(planes) 131 | self.conv2 = self._conv3x3(planes, planes) 132 | self.bn2 = nn.BatchNorm2d(planes) 133 | self.relu = nn.ReLU(inplace=True) 134 | self.downsample = downsample 135 | self.stride = stride 136 | 137 | def _conv3x3(self, in_planes, out_planes, stride=1): 138 | "3x3 convolution with padding" 139 | return nn.Conv2d(in_planes, 140 | out_planes, 141 | kernel_size=3, 142 | stride=stride, 143 | padding=1, 144 | bias=False) 145 | 146 | def forward(self, x): 147 | residual = x 148 | 149 | out = self.conv1(x) 150 | out = self.bn1(out) 151 | out = self.relu(out) 152 | 153 | out = self.conv2(out) 154 | out = self.bn2(out) 155 | 156 | if self.downsample is not None: 157 | residual = self.downsample(x) 158 | out += residual 159 | out = self.relu(out) 160 | 161 | return out 162 | 163 | 164 | class ResNet(nn.Module): 165 | 166 | def __init__(self, n_input_channels: int, n_output_channels: int, block, 167 | layers): 168 | """ 169 | :param n_input_channels (int): The number of input channels of the feature extractor 170 | :param n_output_channels (int): The number of output channels of the feature extractor 171 | :param block: 172 | :param layers: 173 | """ 174 | super(ResNet, self).__init__() 175 | 176 | self.output_channel_blocks = [ 177 | int(n_output_channels / 4), 178 | int(n_output_channels / 2), 179 | n_output_channels, 180 | n_output_channels, 181 | ] 182 | 183 | self.inplanes = int(n_output_channels / 8) 184 | self.conv0_1 = nn.Conv2d( 185 | n_input_channels, 186 | int(n_output_channels / 16), 187 | kernel_size=3, 188 | stride=1, 189 | padding=1, 190 | bias=False, 191 | ) 192 | self.bn0_1 = nn.BatchNorm2d(int(n_output_channels / 16)) 193 | self.conv0_2 = nn.Conv2d( 194 | int(n_output_channels / 16), 195 | self.inplanes, 196 | kernel_size=3, 197 | stride=1, 198 | padding=1, 199 | bias=False, 200 | ) 201 | self.bn0_2 = nn.BatchNorm2d(self.inplanes) 202 | self.relu = nn.ReLU(inplace=True) 203 | 204 | self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 205 | self.layer1 = self._make_layer(block, self.output_channel_blocks[0], 206 | layers[0]) 207 | self.conv1 = nn.Conv2d( 208 | self.output_channel_blocks[0], 209 | self.output_channel_blocks[0], 210 | kernel_size=3, 211 | stride=1, 212 | padding=1, 213 | bias=False, 214 | ) 215 | self.bn1 = nn.BatchNorm2d(self.output_channel_blocks[0]) 216 | 217 | self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 218 | self.layer2 = self._make_layer(block, 219 | self.output_channel_blocks[1], 220 | layers[1], 221 | stride=1) 222 | self.conv2 = nn.Conv2d( 223 | self.output_channel_blocks[1], 224 | self.output_channel_blocks[1], 225 | kernel_size=3, 226 | stride=1, 227 | padding=1, 228 | bias=False, 229 | ) 230 | self.bn2 = nn.BatchNorm2d(self.output_channel_blocks[1]) 231 | 232 | self.maxpool3 = nn.MaxPool2d(kernel_size=2, 233 | stride=(2, 1), 234 | padding=(0, 1)) 235 | self.layer3 = self._make_layer(block, 236 | self.output_channel_blocks[2], 237 | layers[2], 238 | stride=1) 239 | self.conv3 = nn.Conv2d( 240 | self.output_channel_blocks[2], 241 | self.output_channel_blocks[2], 242 | kernel_size=3, 243 | stride=1, 244 | padding=1, 245 | bias=False, 246 | ) 247 | self.bn3 = nn.BatchNorm2d(self.output_channel_blocks[2]) 248 | 249 | self.layer4 = self._make_layer(block, 250 | self.output_channel_blocks[3], 251 | layers[3], 252 | stride=1) 253 | self.conv4_1 = nn.Conv2d( 254 | self.output_channel_blocks[3], 255 | self.output_channel_blocks[3], 256 | kernel_size=2, 257 | stride=(2, 1), 258 | padding=(0, 1), 259 | bias=False, 260 | ) 261 | self.bn4_1 = nn.BatchNorm2d(self.output_channel_blocks[3]) 262 | self.conv4_2 = nn.Conv2d( 263 | self.output_channel_blocks[3], 264 | self.output_channel_blocks[3], 265 | kernel_size=2, 266 | stride=1, 267 | padding=0, 268 | bias=False, 269 | ) 270 | self.bn4_2 = nn.BatchNorm2d(self.output_channel_blocks[3]) 271 | 272 | def _make_layer(self, block, planes, blocks, stride=1): 273 | downsample = None 274 | if stride != 1 or self.inplanes != planes * block.expansion: 275 | downsample = nn.Sequential( 276 | nn.Conv2d( 277 | self.inplanes, 278 | planes * block.expansion, 279 | kernel_size=1, 280 | stride=stride, 281 | bias=False, 282 | ), 283 | nn.BatchNorm2d(planes * block.expansion), 284 | ) 285 | 286 | layers = [] 287 | layers.append(block(self.inplanes, planes, stride, downsample)) 288 | self.inplanes = planes * block.expansion 289 | for i in range(1, blocks): 290 | layers.append(block(self.inplanes, planes)) 291 | 292 | return nn.Sequential(*layers) 293 | 294 | def forward(self, x): 295 | x = self.conv0_1(x) 296 | x = self.bn0_1(x) 297 | x = self.relu(x) 298 | x = self.conv0_2(x) 299 | x = self.bn0_2(x) 300 | x = self.relu(x) 301 | 302 | x = self.maxpool1(x) 303 | x = self.layer1(x) 304 | x = self.conv1(x) 305 | x = self.bn1(x) 306 | x = self.relu(x) 307 | 308 | x = self.maxpool2(x) 309 | x = self.layer2(x) 310 | x = self.conv2(x) 311 | x = self.bn2(x) 312 | x = self.relu(x) 313 | 314 | x = self.maxpool3(x) 315 | x = self.layer3(x) 316 | x = self.conv3(x) 317 | x = self.bn3(x) 318 | x = self.relu(x) 319 | 320 | x = self.layer4(x) 321 | x = self.conv4_1(x) 322 | x = self.bn4_1(x) 323 | x = self.relu(x) 324 | x = self.conv4_2(x) 325 | x = self.bn4_2(x) 326 | x = self.relu(x) 327 | 328 | return x 329 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/_modules.py: -------------------------------------------------------------------------------- 1 | from collections import namedtuple 2 | 3 | import numpy as np 4 | import torch 5 | import torch.nn as nn 6 | import torch.nn.functional as F 7 | import torch.nn.init as init 8 | from torchvision import models 9 | from torchvision.models.vgg import model_urls 10 | 11 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 12 | 13 | 14 | def init_weights(modules): 15 | for m in modules: 16 | if isinstance(m, nn.Conv2d): 17 | init.xavier_uniform_(m.weight.data) 18 | if m.bias is not None: 19 | m.bias.data.zero_() 20 | elif isinstance(m, nn.BatchNorm2d): 21 | m.weight.data.fill_(1) 22 | m.bias.data.zero_() 23 | elif isinstance(m, nn.Linear): 24 | m.weight.data.normal_(0, 0.01) 25 | m.bias.data.zero_() 26 | 27 | 28 | class Vgg16BN(torch.nn.Module): 29 | 30 | def __init__(self, pretrained: bool = True, freeze: bool = True): 31 | super(Vgg16BN, self).__init__() 32 | model_urls["vgg16_bn"] = model_urls["vgg16_bn"].replace( 33 | "https://", "http://") 34 | vgg_pretrained_features = models.vgg16_bn( 35 | pretrained=pretrained).features 36 | self.slice1 = torch.nn.Sequential() 37 | self.slice2 = torch.nn.Sequential() 38 | self.slice3 = torch.nn.Sequential() 39 | self.slice4 = torch.nn.Sequential() 40 | self.slice5 = torch.nn.Sequential() 41 | for x in range(12): # conv2_2 42 | self.slice1.add_module(str(x), vgg_pretrained_features[x]) 43 | for x in range(12, 19): # conv3_3 44 | self.slice2.add_module(str(x), vgg_pretrained_features[x]) 45 | for x in range(19, 29): # conv4_3 46 | self.slice3.add_module(str(x), vgg_pretrained_features[x]) 47 | for x in range(29, 39): # conv5_3 48 | self.slice4.add_module(str(x), vgg_pretrained_features[x]) 49 | 50 | # fc6, fc7 without atrous conv 51 | self.slice5 = torch.nn.Sequential( 52 | nn.MaxPool2d(kernel_size=3, stride=1, padding=1), 53 | nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6), 54 | nn.Conv2d(1024, 1024, kernel_size=1), 55 | ) 56 | 57 | if not pretrained: 58 | init_weights(self.slice1.modules()) 59 | init_weights(self.slice2.modules()) 60 | init_weights(self.slice3.modules()) 61 | init_weights(self.slice4.modules()) 62 | 63 | init_weights( 64 | self.slice5.modules()) # no pretrained model for fc6 and fc7 65 | 66 | if freeze: 67 | for param in self.slice1.parameters(): # only first conv 68 | param.requires_grad = False 69 | 70 | def forward(self, x): 71 | h = self.slice1(x) 72 | h_relu2_2 = h 73 | h = self.slice2(h) 74 | h_relu3_2 = h 75 | h = self.slice3(h) 76 | h_relu4_3 = h 77 | h = self.slice4(h) 78 | h_relu5_3 = h 79 | h = self.slice5(h) 80 | h_fc7 = h 81 | vgg_outputs = namedtuple( 82 | "VggOutputs", ["fc7", "relu5_3", "relu4_3", "relu3_2", "relu2_2"]) 83 | out = vgg_outputs(h_fc7, h_relu5_3, h_relu4_3, h_relu3_2, h_relu2_2) 84 | return out 85 | 86 | 87 | class VGGFeatureExtractor(nn.Module): 88 | """ FeatureExtractor of CRNN (https://arxiv.org/pdf/1507.05717.pdf) """ 89 | 90 | def __init__(self, n_input_channels: int = 1, n_output_channels: int = 512): 91 | super(VGGFeatureExtractor, self).__init__() 92 | 93 | self.output_channel = [ 94 | int(n_output_channels / 8), 95 | int(n_output_channels / 4), 96 | int(n_output_channels / 2), 97 | n_output_channels, 98 | ] # [64, 128, 256, 512] 99 | self.ConvNet = nn.Sequential( 100 | nn.Conv2d(n_input_channels, self.output_channel[0], 3, 1, 1), 101 | nn.ReLU(True), 102 | nn.MaxPool2d(2, 2), # 64x16x50 103 | nn.Conv2d(self.output_channel[0], self.output_channel[1], 3, 1, 1), 104 | nn.ReLU(True), 105 | nn.MaxPool2d(2, 2), # 128x8x25 106 | nn.Conv2d(self.output_channel[1], self.output_channel[2], 3, 1, 1), 107 | nn.ReLU(True), # 256x8x25 108 | nn.Conv2d(self.output_channel[2], self.output_channel[2], 3, 1, 1), 109 | nn.ReLU(True), 110 | nn.MaxPool2d((2, 1), (2, 1)), # 256x4x25 111 | nn.Conv2d( 112 | self.output_channel[2], 113 | self.output_channel[3], 114 | 3, 115 | 1, 116 | 1, 117 | bias=False, 118 | ), 119 | nn.BatchNorm2d(self.output_channel[3]), 120 | nn.ReLU(True), # 512x4x25 121 | nn.Conv2d( 122 | self.output_channel[3], 123 | self.output_channel[3], 124 | 3, 125 | 1, 126 | 1, 127 | bias=False, 128 | ), 129 | nn.BatchNorm2d(self.output_channel[3]), 130 | nn.ReLU(True), 131 | nn.MaxPool2d((2, 1), (2, 1)), # 512x2x25 132 | nn.Conv2d(self.output_channel[3], self.output_channel[3], 2, 1, 0), 133 | nn.ReLU(True), 134 | ) # 512x1x24 135 | 136 | def forward(self, x): 137 | return self.ConvNet(x) 138 | 139 | 140 | class BidirectionalLSTM(nn.Module): 141 | 142 | def __init__(self, input_size: int, hidden_size: int, output_size: int): 143 | super(BidirectionalLSTM, self).__init__() 144 | self.rnn = nn.LSTM( 145 | input_size, 146 | hidden_size, 147 | bidirectional=True, 148 | batch_first=True, 149 | ) 150 | self.linear = nn.Linear(hidden_size * 2, output_size) 151 | 152 | def forward(self, x): 153 | """ 154 | x : visual feature [batch_size x T x input_size] 155 | output : contextual feature [batch_size x T x output_size] 156 | """ 157 | self.rnn.flatten_parameters() 158 | recurrent, _ = self.rnn( 159 | x 160 | ) # batch_size x T x input_size -> batch_size x T x (2*hidden_size) 161 | output = self.linear(recurrent) # batch_size x T x output_size 162 | return output 163 | 164 | 165 | class ResNetFeatureExtractor(nn.Module): 166 | """ 167 | FeatureExtractor of FAN 168 | (http://openaccess.thecvf.com/content_ICCV_2017/papers/Cheng_Focusing_Attention_Towards_ICCV_2017_paper.pdf) 169 | 170 | """ 171 | 172 | def __init__(self, n_input_channels: int = 1, n_output_channels: int = 512): 173 | super(ResNetFeatureExtractor, self).__init__() 174 | self.ConvNet = ResNet( 175 | n_input_channels, 176 | n_output_channels, 177 | BasicBlock, 178 | [1, 2, 5, 3], 179 | ) 180 | 181 | def forward(self, inputs): 182 | return self.ConvNet(inputs) 183 | 184 | 185 | class BasicBlock(nn.Module): 186 | expansion = 1 187 | 188 | def __init__(self, 189 | inplanes: int, 190 | planes: int, 191 | stride: int = 1, 192 | downsample=None): 193 | super(BasicBlock, self).__init__() 194 | self.conv1 = self._conv3x3(inplanes, planes) 195 | self.bn1 = nn.BatchNorm2d(planes) 196 | self.conv2 = self._conv3x3(planes, planes) 197 | self.bn2 = nn.BatchNorm2d(planes) 198 | self.relu = nn.ReLU(inplace=True) 199 | self.downsample = downsample 200 | self.stride = stride 201 | 202 | def _conv3x3(self, in_planes, out_planes, stride=1): 203 | "3x3 convolution with padding" 204 | return nn.Conv2d( 205 | in_planes, 206 | out_planes, 207 | kernel_size=3, 208 | stride=stride, 209 | padding=1, 210 | bias=False, 211 | ) 212 | 213 | def forward(self, x): 214 | residual = x 215 | 216 | out = self.conv1(x) 217 | out = self.bn1(out) 218 | out = self.relu(out) 219 | 220 | out = self.conv2(out) 221 | out = self.bn2(out) 222 | 223 | if self.downsample is not None: 224 | residual = self.downsample(x) 225 | out += residual 226 | out = self.relu(out) 227 | 228 | return out 229 | 230 | 231 | class ResNet(nn.Module): 232 | 233 | def __init__( 234 | self, 235 | n_input_channels: int, 236 | n_output_channels: int, 237 | block, 238 | layers, 239 | ): 240 | """ 241 | :param n_input_channels (int): The number of input channels of the feature extractor 242 | :param n_output_channels (int): The number of output channels of the feature extractor 243 | :param block: 244 | :param layers: 245 | """ 246 | super(ResNet, self).__init__() 247 | 248 | self.output_channel_blocks = [ 249 | int(n_output_channels / 4), 250 | int(n_output_channels / 2), 251 | n_output_channels, 252 | n_output_channels, 253 | ] 254 | 255 | self.inplanes = int(n_output_channels / 8) 256 | self.conv0_1 = nn.Conv2d( 257 | n_input_channels, 258 | int(n_output_channels / 16), 259 | kernel_size=3, 260 | stride=1, 261 | padding=1, 262 | bias=False, 263 | ) 264 | self.bn0_1 = nn.BatchNorm2d(int(n_output_channels / 16)) 265 | self.conv0_2 = nn.Conv2d( 266 | int(n_output_channels / 16), 267 | self.inplanes, 268 | kernel_size=3, 269 | stride=1, 270 | padding=1, 271 | bias=False, 272 | ) 273 | self.bn0_2 = nn.BatchNorm2d(self.inplanes) 274 | self.relu = nn.ReLU(inplace=True) 275 | 276 | self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 277 | self.layer1 = self._make_layer( 278 | block, 279 | self.output_channel_blocks[0], 280 | layers[0], 281 | ) 282 | self.conv1 = nn.Conv2d( 283 | self.output_channel_blocks[0], 284 | self.output_channel_blocks[0], 285 | kernel_size=3, 286 | stride=1, 287 | padding=1, 288 | bias=False, 289 | ) 290 | self.bn1 = nn.BatchNorm2d(self.output_channel_blocks[0]) 291 | 292 | self.maxpool2 = nn.MaxPool2d(kernel_size=2, stride=2, padding=0) 293 | self.layer2 = self._make_layer( 294 | block, 295 | self.output_channel_blocks[1], 296 | layers[1], 297 | stride=1, 298 | ) 299 | self.conv2 = nn.Conv2d( 300 | self.output_channel_blocks[1], 301 | self.output_channel_blocks[1], 302 | kernel_size=3, 303 | stride=1, 304 | padding=1, 305 | bias=False, 306 | ) 307 | self.bn2 = nn.BatchNorm2d(self.output_channel_blocks[1]) 308 | 309 | self.maxpool3 = nn.MaxPool2d( 310 | kernel_size=2, 311 | stride=(2, 1), 312 | padding=(0, 1), 313 | ) 314 | self.layer3 = self._make_layer( 315 | block, 316 | self.output_channel_blocks[2], 317 | layers[2], 318 | stride=1, 319 | ) 320 | self.conv3 = nn.Conv2d( 321 | self.output_channel_blocks[2], 322 | self.output_channel_blocks[2], 323 | kernel_size=3, 324 | stride=1, 325 | padding=1, 326 | bias=False, 327 | ) 328 | self.bn3 = nn.BatchNorm2d(self.output_channel_blocks[2]) 329 | 330 | self.layer4 = self._make_layer( 331 | block, 332 | self.output_channel_blocks[3], 333 | layers[3], 334 | stride=1, 335 | ) 336 | self.conv4_1 = nn.Conv2d( 337 | self.output_channel_blocks[3], 338 | self.output_channel_blocks[3], 339 | kernel_size=2, 340 | stride=(2, 1), 341 | padding=(0, 1), 342 | bias=False, 343 | ) 344 | self.bn4_1 = nn.BatchNorm2d(self.output_channel_blocks[3]) 345 | self.conv4_2 = nn.Conv2d( 346 | self.output_channel_blocks[3], 347 | self.output_channel_blocks[3], 348 | kernel_size=2, 349 | stride=1, 350 | padding=0, 351 | bias=False, 352 | ) 353 | self.bn4_2 = nn.BatchNorm2d(self.output_channel_blocks[3]) 354 | 355 | def _make_layer(self, block, planes, blocks, stride=1): 356 | downsample = None 357 | if stride != 1 or self.inplanes != planes * block.expansion: 358 | downsample = nn.Sequential( 359 | nn.Conv2d( 360 | self.inplanes, 361 | planes * block.expansion, 362 | kernel_size=1, 363 | stride=stride, 364 | bias=False, 365 | ), 366 | nn.BatchNorm2d(planes * block.expansion), 367 | ) 368 | 369 | layers = [] 370 | layers.append(block(self.inplanes, planes, stride, downsample)) 371 | self.inplanes = planes * block.expansion 372 | for i in range(1, blocks): 373 | layers.append(block(self.inplanes, planes)) 374 | 375 | return nn.Sequential(*layers) 376 | 377 | def forward(self, x): 378 | x = self.conv0_1(x) 379 | x = self.bn0_1(x) 380 | x = self.relu(x) 381 | x = self.conv0_2(x) 382 | x = self.bn0_2(x) 383 | x = self.relu(x) 384 | 385 | x = self.maxpool1(x) 386 | x = self.layer1(x) 387 | x = self.conv1(x) 388 | x = self.bn1(x) 389 | x = self.relu(x) 390 | 391 | x = self.maxpool2(x) 392 | x = self.layer2(x) 393 | x = self.conv2(x) 394 | x = self.bn2(x) 395 | x = self.relu(x) 396 | 397 | x = self.maxpool3(x) 398 | x = self.layer3(x) 399 | x = self.conv3(x) 400 | x = self.bn3(x) 401 | x = self.relu(x) 402 | 403 | x = self.layer4(x) 404 | x = self.conv4_1(x) 405 | x = self.bn4_1(x) 406 | x = self.relu(x) 407 | x = self.conv4_2(x) 408 | x = self.bn4_2(x) 409 | x = self.relu(x) 410 | 411 | return x 412 | 413 | 414 | class TpsSpatialTransformerNetwork(nn.Module): 415 | """ Rectification Network of RARE, namely TPS based STN """ 416 | 417 | def __init__(self, F, I_size, I_r_size, I_channel_num: int = 1): 418 | """Based on RARE TPS 419 | input: 420 | batch_I: Batch Input Image [batch_size x I_channel_num x I_height x I_width] 421 | I_size : (height, width) of the input image I 422 | I_r_size : (height, width) of the rectified image I_r 423 | I_channel_num : the number of channels of the input image I 424 | output: 425 | batch_I_r: rectified image [batch_size x I_channel_num x I_r_height x I_r_width] 426 | """ 427 | super(TpsSpatialTransformerNetwork, self).__init__() 428 | self.F = F 429 | self.I_size = I_size 430 | self.I_r_size = I_r_size # = (I_r_height, I_r_width) 431 | self.I_channel_num = I_channel_num 432 | self.LocalizationNetwork = LocalizationNetwork( 433 | self.F, 434 | self.I_channel_num, 435 | ) 436 | self.GridGenerator = GridGenerator(self.F, self.I_r_size) 437 | 438 | def forward(self, batch_I): 439 | batch_C_prime = self.LocalizationNetwork(batch_I) # batch_size x K x 2 440 | build_P_prime = self.GridGenerator.build_P_prime( 441 | batch_C_prime) # batch_size x n (= I_r_width x I_r_height) x 2 442 | build_P_prime_reshape = build_P_prime.reshape( 443 | [build_P_prime.size(0), self.I_r_size[0], self.I_r_size[1], 2]) 444 | 445 | batch_I_r = F.grid_sample( 446 | batch_I, 447 | build_P_prime_reshape, 448 | padding_mode="border", 449 | ) 450 | 451 | return batch_I_r 452 | 453 | 454 | class LocalizationNetwork(nn.Module): 455 | """ Localization Network of RARE, which predicts C' (K x 2) from I (I_width x I_height) """ 456 | 457 | def __init__(self, F, I_channel_num: int): 458 | super(LocalizationNetwork, self).__init__() 459 | self.F = F 460 | self.I_channel_num = I_channel_num 461 | self.conv = nn.Sequential( 462 | nn.Conv2d( 463 | in_channels=self.I_channel_num, 464 | out_channels=64, 465 | kernel_size=3, 466 | stride=1, 467 | padding=1, 468 | bias=False, 469 | ), 470 | nn.BatchNorm2d(64), 471 | nn.ReLU(True), 472 | nn.MaxPool2d(2, 2), # batch_size x 64 x I_height/2 x I_width/2 473 | nn.Conv2d(64, 128, 3, 1, 1, bias=False), 474 | nn.BatchNorm2d(128), 475 | nn.ReLU(True), 476 | nn.MaxPool2d(2, 2), # batch_size x 128 x I_height/4 x I_width/4 477 | nn.Conv2d(128, 256, 3, 1, 1, bias=False), 478 | nn.BatchNorm2d(256), 479 | nn.ReLU(True), 480 | nn.MaxPool2d(2, 2), # batch_size x 256 x I_height/8 x I_width/8 481 | nn.Conv2d(256, 512, 3, 1, 1, bias=False), 482 | nn.BatchNorm2d(512), 483 | nn.ReLU(True), 484 | nn.AdaptiveAvgPool2d(1), # batch_size x 512 485 | ) 486 | 487 | self.localization_fc1 = nn.Sequential(nn.Linear(512, 256), 488 | nn.ReLU(True)) 489 | self.localization_fc2 = nn.Linear(256, self.F * 2) 490 | 491 | # Init fc2 in LocalizationNetwork 492 | self.localization_fc2.weight.data.fill_(0) 493 | 494 | # see RARE paper Fig. 6 (a) 495 | ctrl_pts_x = np.linspace(-1.0, 1.0, int(F / 2)) 496 | ctrl_pts_y_top = np.linspace(0.0, -1.0, num=int(F / 2)) 497 | ctrl_pts_y_bottom = np.linspace(1.0, 0.0, num=int(F / 2)) 498 | ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1) 499 | ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1) 500 | initial_bias = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0) 501 | self.localization_fc2.bias.data = ( 502 | torch.from_numpy(initial_bias).float().view(-1)) 503 | 504 | def forward(self, batch_I): 505 | """ 506 | :param batch_I : Batch Input Image [batch_size x I_channel_num x I_height x I_width] 507 | :return: batch_C_prime : Predicted coordinates of fiducial points for input batch [batch_size x F x 2] 508 | """ 509 | batch_size = batch_I.size(0) 510 | features = self.conv(batch_I).view(batch_size, -1) 511 | batch_C_prime = self.localization_fc2( 512 | self.localization_fc1(features)).view(batch_size, self.F, 2) 513 | return batch_C_prime 514 | 515 | 516 | class GridGenerator(nn.Module): 517 | """ Grid Generator of RARE, which produces P_prime by multipling T with P """ 518 | 519 | def __init__(self, F, I_r_size): 520 | """ Generate P_hat and inv_delta_C for later """ 521 | super(GridGenerator, self).__init__() 522 | self.eps = 1e-6 523 | self.I_r_height, self.I_r_width = I_r_size 524 | self.F = F 525 | self.C = self._build_C(self.F) # F x 2 526 | self.P = self._build_P(self.I_r_width, self.I_r_height) 527 | 528 | # for multi-gpu, you need register buffer 529 | self.register_buffer( 530 | "inv_delta_C", 531 | torch.tensor(self._build_inv_delta_C( 532 | self.F, 533 | self.C, 534 | )).float(), 535 | ) # F+3 x F+3 536 | self.register_buffer( 537 | "P_hat", 538 | torch.tensor(self._build_P_hat( 539 | self.F, 540 | self.C, 541 | self.P, 542 | )).float(), 543 | ) # n x F+3 544 | 545 | def _build_C(self, F): 546 | """ Return coordinates of fiducial points in I_r; C """ 547 | ctrl_pts_x = np.linspace(-1.0, 1.0, int(F / 2)) 548 | ctrl_pts_y_top = -1 * np.ones(int(F / 2)) 549 | ctrl_pts_y_bottom = np.ones(int(F / 2)) 550 | ctrl_pts_top = np.stack([ctrl_pts_x, ctrl_pts_y_top], axis=1) 551 | ctrl_pts_bottom = np.stack([ctrl_pts_x, ctrl_pts_y_bottom], axis=1) 552 | C = np.concatenate([ctrl_pts_top, ctrl_pts_bottom], axis=0) 553 | return C # F x 2 554 | 555 | def _build_inv_delta_C(self, F, C): 556 | """ Return inv_delta_C which is needed to calculate T """ 557 | hat_C = np.zeros((F, F), dtype=float) # F x F 558 | for i in range(0, F): 559 | for j in range(i, F): 560 | r = np.linalg.norm(C[i] - C[j]) 561 | hat_C[i, j] = r 562 | hat_C[j, i] = r 563 | np.fill_diagonal(hat_C, 1) 564 | hat_C = (hat_C**2) * np.log(hat_C) 565 | # print(C.shape, hat_C.shape) 566 | delta_C = np.concatenate( # F+3 x F+3 567 | [ 568 | np.concatenate([np.ones((F, 1)), C, hat_C], axis=1), # F x F+3 569 | np.concatenate([np.zeros( 570 | (2, 3)), np.transpose(C)], axis=1), # 2 x F+3 571 | np.concatenate([np.zeros( 572 | (1, 3)), np.ones((1, F))], axis=1), # 1 x F+3 573 | ], 574 | axis=0, 575 | ) 576 | inv_delta_C = np.linalg.inv(delta_C) 577 | return inv_delta_C # F+3 x F+3 578 | 579 | def _build_P(self, I_r_width, I_r_height): 580 | I_r_grid_x = (np.arange(-I_r_width, I_r_width, 2) + 581 | 1.0) / I_r_width # self.I_r_width 582 | I_r_grid_y = (np.arange(-I_r_height, I_r_height, 2) + 583 | 1.0) / I_r_height # self.I_r_height 584 | P = np.stack( # self.I_r_width x self.I_r_height x 2 585 | np.meshgrid(I_r_grid_x, I_r_grid_y), 586 | axis=2) 587 | return P.reshape([-1, 2]) # n (= self.I_r_width x self.I_r_height) x 2 588 | 589 | def _build_P_hat(self, F, C, P): 590 | n = P.shape[0] # n (= self.I_r_width x self.I_r_height) 591 | P_tile = np.tile(np.expand_dims(P, axis=1), 592 | (1, F, 1)) # n x 2 -> n x 1 x 2 -> n x F x 2 593 | C_tile = np.expand_dims(C, axis=0) # 1 x F x 2 594 | P_diff = P_tile - C_tile # n x F x 2 595 | rbf_norm = np.linalg.norm( 596 | P_diff, 597 | ord=2, 598 | axis=2, 599 | keepdims=False, 600 | ) # n x F 601 | rbf = np.multiply( 602 | np.square(rbf_norm), 603 | np.log(rbf_norm + self.eps), 604 | ) # n x F 605 | P_hat = np.concatenate([np.ones((n, 1)), P, rbf], axis=1) 606 | return P_hat # n x F+3 607 | 608 | def build_P_prime(self, batch_C_prime): 609 | """ Generate Grid from batch_C_prime [batch_size x F x 2] """ 610 | batch_size = batch_C_prime.size(0) 611 | batch_inv_delta_C = self.inv_delta_C.repeat(batch_size, 1, 1) 612 | batch_P_hat = self.P_hat.repeat(batch_size, 1, 1) 613 | batch_C_prime_with_zeros = torch.cat( 614 | (batch_C_prime, torch.zeros(batch_size, 3, 2).float().to(device)), 615 | dim=1) # batch_size x F+3 x 2 616 | batch_T = torch.bmm( 617 | batch_inv_delta_C, 618 | batch_C_prime_with_zeros, 619 | ) # batch_size x F+3 x 2 620 | batch_P_prime = torch.bmm(batch_P_hat, batch_T) # batch_size x n x 2 621 | return batch_P_prime # batch_size x n x 2 622 | -------------------------------------------------------------------------------- /pororo/models/brainOCR/utils.py: -------------------------------------------------------------------------------- 1 | """ 2 | This code is adapted from https://github.com/JaidedAI/EasyOCR/blob/8af936ba1b2f3c230968dc1022d0cd3e9ca1efbb/easyocr/utils.py 3 | """ 4 | 5 | import math 6 | import os 7 | from urllib.request import urlretrieve 8 | 9 | import cv2 10 | import numpy as np 11 | import torch 12 | from PIL import Image 13 | from torch import Tensor 14 | 15 | from .imgproc import load_image 16 | 17 | 18 | def consecutive(data, mode: str = "first", stepsize: int = 1): 19 | group = np.split(data, np.where(np.diff(data) != stepsize)[0] + 1) 20 | group = [item for item in group if len(item) > 0] 21 | 22 | if mode == "first": 23 | result = [l[0] for l in group] 24 | elif mode == "last": 25 | result = [l[-1] for l in group] 26 | return result 27 | 28 | 29 | def word_segmentation( 30 | mat, 31 | separator_idx={ 32 | "th": [1, 2], 33 | "en": [3, 4] 34 | }, 35 | separator_idx_list=[1, 2, 3, 4], 36 | ): 37 | result = [] 38 | sep_list = [] 39 | start_idx = 0 40 | sep_lang = "" 41 | for sep_idx in separator_idx_list: 42 | if sep_idx % 2 == 0: 43 | mode = "first" 44 | else: 45 | mode = "last" 46 | a = consecutive(np.argwhere(mat == sep_idx).flatten(), mode) 47 | new_sep = [[item, sep_idx] for item in a] 48 | sep_list += new_sep 49 | sep_list = sorted(sep_list, key=lambda x: x[0]) 50 | 51 | for sep in sep_list: 52 | for lang in separator_idx.keys(): 53 | if sep[1] == separator_idx[lang][0]: # start lang 54 | sep_lang = lang 55 | sep_start_idx = sep[0] 56 | elif sep[1] == separator_idx[lang][1]: # end lang 57 | if sep_lang == lang: # check if last entry if the same start lang 58 | new_sep_pair = [lang, [sep_start_idx + 1, sep[0] - 1]] 59 | if sep_start_idx > start_idx: 60 | result.append(["", [start_idx, sep_start_idx - 1]]) 61 | start_idx = sep[0] + 1 62 | result.append(new_sep_pair) 63 | sep_lang = "" # reset 64 | 65 | if start_idx <= len(mat) - 1: 66 | result.append(["", [start_idx, len(mat) - 1]]) 67 | return result 68 | 69 | 70 | # code is based from https://github.com/githubharald/CTCDecoder/blob/master/src/BeamSearch.py 71 | class BeamEntry: 72 | "information about one single beam at specific time-step" 73 | 74 | def __init__(self): 75 | self.prTotal = 0 # blank and non-blank 76 | self.prNonBlank = 0 # non-blank 77 | self.prBlank = 0 # blank 78 | self.prText = 1 # LM score 79 | self.lmApplied = False # flag if LM was already applied to this beam 80 | self.labeling = () # beam-labeling 81 | 82 | 83 | class BeamState: 84 | "information about the beams at specific time-step" 85 | 86 | def __init__(self): 87 | self.entries = {} 88 | 89 | def norm(self): 90 | "length-normalise LM score" 91 | for (k, _) in self.entries.items(): 92 | labelingLen = len(self.entries[k].labeling) 93 | self.entries[k].prText = self.entries[k].prText**( 94 | 1.0 / (labelingLen if labelingLen else 1.0)) 95 | 96 | def sort(self): 97 | "return beam-labelings, sorted by probability" 98 | beams = [v for (_, v) in self.entries.items()] 99 | sortedBeams = sorted( 100 | beams, 101 | reverse=True, 102 | key=lambda x: x.prTotal * x.prText, 103 | ) 104 | return [x.labeling for x in sortedBeams] 105 | 106 | def wordsearch(self, classes, ignore_idx, maxCandidate, dict_list): 107 | beams = [v for (_, v) in self.entries.items()] 108 | sortedBeams = sorted( 109 | beams, 110 | reverse=True, 111 | key=lambda x: x.prTotal * x.prText, 112 | ) 113 | if len(sortedBeams) > maxCandidate: 114 | sortedBeams = sortedBeams[:maxCandidate] 115 | 116 | for j, candidate in enumerate(sortedBeams): 117 | idx_list = candidate.labeling 118 | text = "" 119 | for i, l in enumerate(idx_list): 120 | if l not in ignore_idx and ( 121 | not (i > 0 and idx_list[i - 1] == idx_list[i])): 122 | text += classes[l] 123 | 124 | if j == 0: 125 | best_text = text 126 | if text in dict_list: 127 | # print('found text: ', text) 128 | best_text = text 129 | break 130 | else: 131 | pass 132 | # print('not in dict: ', text) 133 | return best_text 134 | 135 | 136 | def applyLM(parentBeam, childBeam, classes, lm_model, lm_factor: float = 0.01): 137 | "calculate LM score of child beam by taking score from parent beam and bigram probability of last two chars" 138 | if lm_model is not None and not childBeam.lmApplied: 139 | history = parentBeam.labeling 140 | history = " ".join( 141 | classes[each].replace(" ", "▁") for each in history if each != 0) 142 | 143 | current_char = classes[childBeam.labeling[-1]].replace(" ", "▁") 144 | if current_char == "[blank]": 145 | lmProb = 1 146 | else: 147 | text = history + " " + current_char 148 | lmProb = 10**lm_model.score(text, bos=True) * lm_factor 149 | 150 | childBeam.prText = lmProb # probability of char sequence 151 | childBeam.lmApplied = True # only apply LM once per beam entry 152 | 153 | 154 | def simplify_label(labeling, blankIdx: int = 0): 155 | labeling = np.array(labeling) 156 | 157 | # collapse blank 158 | idx = np.where(~((np.roll(labeling, 1) == labeling) & 159 | (labeling == blankIdx)))[0] 160 | labeling = labeling[idx] 161 | 162 | # get rid of blank between different characters 163 | idx = np.where(~((np.roll(labeling, 1) != np.roll(labeling, -1)) & 164 | (labeling == blankIdx)))[0] 165 | 166 | if len(labeling) > 0: 167 | last_idx = len(labeling) - 1 168 | if last_idx not in idx: 169 | idx = np.append(idx, [last_idx]) 170 | labeling = labeling[idx] 171 | 172 | return tuple(labeling) 173 | 174 | 175 | def addBeam(beamState, labeling): 176 | "add beam if it does not yet exist" 177 | if labeling not in beamState.entries: 178 | beamState.entries[labeling] = BeamEntry() 179 | 180 | 181 | def ctcBeamSearch( 182 | mat, 183 | classes: list, 184 | ignore_idx: int, 185 | lm_model, 186 | lm_factor: float = 0.01, 187 | beam_width: int = 5, 188 | ): 189 | blankIdx = 0 190 | maxT, maxC = mat.shape 191 | 192 | # initialise beam state 193 | last = BeamState() 194 | labeling = () 195 | last.entries[labeling] = BeamEntry() 196 | last.entries[labeling].prBlank = 1 197 | last.entries[labeling].prTotal = 1 198 | 199 | # go over all time-steps 200 | for t in range(maxT): 201 | # print("t=", t) 202 | curr = BeamState() 203 | # get beam-labelings of best beams 204 | bestLabelings = last.sort()[0:beam_width] 205 | # go over best beams 206 | for labeling in bestLabelings: 207 | # print("labeling:", labeling) 208 | # probability of paths ending with a non-blank 209 | prNonBlank = 0 210 | # in case of non-empty beam 211 | if labeling: 212 | # probability of paths with repeated last char at the end 213 | prNonBlank = last.entries[labeling].prNonBlank * mat[ 214 | t, labeling[-1]] 215 | 216 | # probability of paths ending with a blank 217 | prBlank = (last.entries[labeling].prTotal) * mat[t, blankIdx] 218 | 219 | # add beam at current time-step if needed 220 | labeling = simplify_label(labeling, blankIdx) 221 | addBeam(curr, labeling) 222 | 223 | # fill in data 224 | curr.entries[labeling].labeling = labeling 225 | curr.entries[labeling].prNonBlank += prNonBlank 226 | curr.entries[labeling].prBlank += prBlank 227 | curr.entries[labeling].prTotal += prBlank + prNonBlank 228 | curr.entries[labeling].prText = last.entries[labeling].prText 229 | # beam-labeling not changed, therefore also LM score unchanged from 230 | 231 | curr.entries[labeling].lmApplied = ( 232 | True # LM already applied at previous time-step for this beam-labeling 233 | ) 234 | 235 | # extend current beam-labeling 236 | # char_highscore = np.argpartition(mat[t, :], -5)[-5:] # run through 5 highest probability 237 | char_highscore = np.where( 238 | mat[t, :] >= 0.5 / 239 | maxC)[0] # run through all probable characters 240 | for c in char_highscore: 241 | # for c in range(maxC - 1): 242 | # add new char to current beam-labeling 243 | newLabeling = labeling + (c,) 244 | newLabeling = simplify_label(newLabeling, blankIdx) 245 | 246 | # if new labeling contains duplicate char at the end, only consider paths ending with a blank 247 | if labeling and labeling[-1] == c: 248 | prNonBlank = mat[t, c] * last.entries[labeling].prBlank 249 | else: 250 | prNonBlank = mat[t, c] * last.entries[labeling].prTotal 251 | 252 | # add beam at current time-step if needed 253 | addBeam(curr, newLabeling) 254 | 255 | # fill in data 256 | curr.entries[newLabeling].labeling = newLabeling 257 | curr.entries[newLabeling].prNonBlank += prNonBlank 258 | curr.entries[newLabeling].prTotal += prNonBlank 259 | 260 | # apply LM 261 | applyLM( 262 | curr.entries[labeling], 263 | curr.entries[newLabeling], 264 | classes, 265 | lm_model, 266 | lm_factor, 267 | ) 268 | 269 | # set new beam state 270 | 271 | last = curr 272 | 273 | # normalise LM scores according to beam-labeling-length 274 | last.norm() 275 | 276 | bestLabeling = last.sort()[0] # get most probable labeling 277 | res = "" 278 | for i, l in enumerate(bestLabeling): 279 | # removing repeated characters and blank. 280 | if l != ignore_idx and (not (i > 0 and 281 | bestLabeling[i - 1] == bestLabeling[i])): 282 | res += classes[l] 283 | 284 | return res 285 | 286 | 287 | class CTCLabelConverter(object): 288 | """ Convert between text-label and text-index """ 289 | 290 | def __init__(self, vocab: list): 291 | self.char2idx = {char: idx for idx, char in enumerate(vocab)} 292 | self.idx2char = {idx: char for idx, char in enumerate(vocab)} 293 | self.ignored_index = 0 294 | self.vocab = vocab 295 | 296 | def encode(self, texts: list): 297 | """ 298 | Convert input texts into indices 299 | texts (list): text labels of each image. [batch_size] 300 | 301 | Returns 302 | text: concatenated text index for CTCLoss. 303 | [sum(text_lengths)] = [text_index_0 + text_index_1 + ... + text_index_(n - 1)] 304 | length: length of each text. [batch_size] 305 | """ 306 | lengths = [len(text) for text in texts] 307 | concatenated_text = "".join(texts) 308 | indices = [self.char2idx[char] for char in concatenated_text] 309 | 310 | return torch.IntTensor(indices), torch.IntTensor(lengths) 311 | 312 | def decode_greedy(self, indices: Tensor, lengths: Tensor): 313 | """convert text-index into text-label. 314 | 315 | :param indices (1D int32 Tensor): [N*length,] 316 | :param lengths (1D int32 Tensor): [N,] 317 | :return: 318 | """ 319 | texts = [] 320 | index = 0 321 | for length in lengths: 322 | text = indices[index:index + length] 323 | 324 | chars = [] 325 | for i in range(length): 326 | if (text[i] != self.ignored_index) and ( 327 | not (i > 0 and text[i - 1] == text[i]) 328 | ): # removing repeated characters and blank (and separator). 329 | chars.append(self.idx2char[text[i].item()]) 330 | texts.append("".join(chars)) 331 | index += length 332 | return texts 333 | 334 | def decode_beamsearch(self, mat, lm_model, lm_factor, beam_width: int = 5): 335 | texts = [] 336 | for i in range(mat.shape[0]): 337 | text = ctcBeamSearch( 338 | mat[i], 339 | self.vocab, 340 | self.ignored_index, 341 | lm_model, 342 | lm_factor, 343 | beam_width, 344 | ) 345 | texts.append(text) 346 | return texts 347 | 348 | 349 | def four_point_transform(image, rect): 350 | (tl, tr, br, bl) = rect 351 | 352 | widthA = np.sqrt(((br[0] - bl[0])**2) + ((br[1] - bl[1])**2)) 353 | widthB = np.sqrt(((tr[0] - tl[0])**2) + ((tr[1] - tl[1])**2)) 354 | maxWidth = max(int(widthA), int(widthB)) 355 | 356 | # compute the height of the new image, which will be the 357 | # maximum distance between the top-right and bottom-right 358 | # y-coordinates or the top-left and bottom-left y-coordinates 359 | heightA = np.sqrt(((tr[0] - br[0])**2) + ((tr[1] - br[1])**2)) 360 | heightB = np.sqrt(((tl[0] - bl[0])**2) + ((tl[1] - bl[1])**2)) 361 | maxHeight = max(int(heightA), int(heightB)) 362 | 363 | dst = np.array( 364 | [[0, 0], [maxWidth - 1, 0], [maxWidth - 1, maxHeight - 1], 365 | [0, maxHeight - 1]], 366 | dtype="float32", 367 | ) 368 | 369 | # compute the perspective transform matrix and then apply it 370 | M = cv2.getPerspectiveTransform(rect, dst) 371 | warped = cv2.warpPerspective(image, M, (maxWidth, maxHeight)) 372 | 373 | return warped 374 | 375 | 376 | def group_text_box( 377 | polys, 378 | slope_ths: float = 0.1, 379 | ycenter_ths: float = 0.5, 380 | height_ths: float = 0.5, 381 | width_ths: float = 1.0, 382 | add_margin: float = 0.05, 383 | ): 384 | # poly top-left, top-right, low-right, low-left 385 | horizontal_list, free_list, combined_list, merged_list = [], [], [], [] 386 | 387 | for poly in polys: 388 | slope_up = (poly[3] - poly[1]) / np.maximum(10, (poly[2] - poly[0])) 389 | slope_down = (poly[5] - poly[7]) / np.maximum(10, (poly[4] - poly[6])) 390 | if max(abs(slope_up), abs(slope_down)) < slope_ths: 391 | x_max = max([poly[0], poly[2], poly[4], poly[6]]) 392 | x_min = min([poly[0], poly[2], poly[4], poly[6]]) 393 | y_max = max([poly[1], poly[3], poly[5], poly[7]]) 394 | y_min = min([poly[1], poly[3], poly[5], poly[7]]) 395 | horizontal_list.append([ 396 | x_min, x_max, y_min, y_max, 0.5 * (y_min + y_max), y_max - y_min 397 | ]) 398 | else: 399 | height = np.linalg.norm([poly[6] - poly[0], poly[7] - poly[1]]) 400 | margin = int(1.44 * add_margin * height) 401 | 402 | theta13 = abs( 403 | np.arctan( 404 | (poly[1] - poly[5]) / np.maximum(10, (poly[0] - poly[4])))) 405 | theta24 = abs( 406 | np.arctan( 407 | (poly[3] - poly[7]) / np.maximum(10, (poly[2] - poly[6])))) 408 | # do I need to clip minimum, maximum value here? 409 | x1 = poly[0] - np.cos(theta13) * margin 410 | y1 = poly[1] - np.sin(theta13) * margin 411 | x2 = poly[2] + np.cos(theta24) * margin 412 | y2 = poly[3] - np.sin(theta24) * margin 413 | x3 = poly[4] + np.cos(theta13) * margin 414 | y3 = poly[5] + np.sin(theta13) * margin 415 | x4 = poly[6] - np.cos(theta24) * margin 416 | y4 = poly[7] + np.sin(theta24) * margin 417 | 418 | free_list.append([[x1, y1], [x2, y2], [x3, y3], [x4, y4]]) 419 | horizontal_list = sorted(horizontal_list, key=lambda item: item[4]) 420 | 421 | # combine box 422 | new_box = [] 423 | for poly in horizontal_list: 424 | 425 | if len(new_box) == 0: 426 | b_height = [poly[5]] 427 | b_ycenter = [poly[4]] 428 | new_box.append(poly) 429 | else: 430 | # comparable height and comparable y_center level up to ths*height 431 | if (abs(np.mean(b_height) - poly[5]) < height_ths * 432 | np.mean(b_height)) and (abs(np.mean(b_ycenter) - poly[4]) < 433 | ycenter_ths * np.mean(b_height)): 434 | b_height.append(poly[5]) 435 | b_ycenter.append(poly[4]) 436 | new_box.append(poly) 437 | else: 438 | b_height = [poly[5]] 439 | b_ycenter = [poly[4]] 440 | combined_list.append(new_box) 441 | new_box = [poly] 442 | combined_list.append(new_box) 443 | 444 | # merge list use sort again 445 | for boxes in combined_list: 446 | if len(boxes) == 1: # one box per line 447 | box = boxes[0] 448 | margin = int(add_margin * box[5]) 449 | merged_list.append([ 450 | box[0] - margin, box[1] + margin, box[2] - margin, 451 | box[3] + margin 452 | ]) 453 | else: # multiple boxes per line 454 | boxes = sorted(boxes, key=lambda item: item[0]) 455 | 456 | merged_box, new_box = [], [] 457 | for box in boxes: 458 | if len(new_box) == 0: 459 | x_max = box[1] 460 | new_box.append(box) 461 | else: 462 | if abs(box[0] - x_max) < width_ths * ( 463 | box[3] - box[2]): # merge boxes 464 | x_max = box[1] 465 | new_box.append(box) 466 | else: 467 | x_max = box[1] 468 | merged_box.append(new_box) 469 | new_box = [box] 470 | if len(new_box) > 0: 471 | merged_box.append(new_box) 472 | 473 | for mbox in merged_box: 474 | if len(mbox) != 1: # adjacent box in same line 475 | # do I need to add margin here? 476 | x_min = min(mbox, key=lambda x: x[0])[0] 477 | x_max = max(mbox, key=lambda x: x[1])[1] 478 | y_min = min(mbox, key=lambda x: x[2])[2] 479 | y_max = max(mbox, key=lambda x: x[3])[3] 480 | 481 | margin = int(add_margin * (y_max - y_min)) 482 | 483 | merged_list.append([ 484 | x_min - margin, x_max + margin, y_min - margin, 485 | y_max + margin 486 | ]) 487 | else: # non adjacent box in same line 488 | box = mbox[0] 489 | 490 | margin = int(add_margin * (box[3] - box[2])) 491 | merged_list.append([ 492 | box[0] - margin, 493 | box[1] + margin, 494 | box[2] - margin, 495 | box[3] + margin, 496 | ]) 497 | # may need to check if box is really in image 498 | return merged_list, free_list 499 | 500 | 501 | def get_image_list(horizontal_list: list, 502 | free_list: list, 503 | img: np.ndarray, 504 | model_height: int = 64): 505 | image_list = [] 506 | maximum_y, maximum_x = img.shape 507 | 508 | max_ratio_hori, max_ratio_free = 1, 1 509 | for box in free_list: 510 | rect = np.array(box, dtype="float32") 511 | transformed_img = four_point_transform(img, rect) 512 | ratio = transformed_img.shape[1] / transformed_img.shape[0] 513 | crop_img = cv2.resize( 514 | transformed_img, 515 | (int(model_height * ratio), model_height), 516 | interpolation=Image.ANTIALIAS, 517 | ) 518 | # box : [[x1,y1],[x2,y2],[x3,y3],[x4,y4]] 519 | image_list.append((box, crop_img)) 520 | max_ratio_free = max(ratio, max_ratio_free) 521 | 522 | max_ratio_free = math.ceil(max_ratio_free) 523 | 524 | for box in horizontal_list: 525 | x_min = max(0, box[0]) 526 | x_max = min(box[1], maximum_x) 527 | y_min = max(0, box[2]) 528 | y_max = min(box[3], maximum_y) 529 | crop_img = img[y_min:y_max, x_min:x_max] 530 | width = x_max - x_min 531 | height = y_max - y_min 532 | ratio = width / height 533 | crop_img = cv2.resize( 534 | crop_img, 535 | (int(model_height * ratio), model_height), 536 | interpolation=Image.ANTIALIAS, 537 | ) 538 | image_list.append(( 539 | [ 540 | [x_min, y_min], 541 | [x_max, y_min], 542 | [x_max, y_max], 543 | [x_min, y_max], 544 | ], 545 | crop_img, 546 | )) 547 | max_ratio_hori = max(ratio, max_ratio_hori) 548 | 549 | max_ratio_hori = math.ceil(max_ratio_hori) 550 | max_ratio = max(max_ratio_hori, max_ratio_free) 551 | max_width = math.ceil(max_ratio) * model_height 552 | 553 | image_list = sorted( 554 | image_list, key=lambda item: item[0][0][1]) # sort by vertical position 555 | return image_list, max_width 556 | 557 | 558 | def diff(input_list): 559 | return max(input_list) - min(input_list) 560 | 561 | 562 | def get_paragraph(raw_result, 563 | x_ths: int = 1, 564 | y_ths: float = 0.5, 565 | mode: str = "ltr"): 566 | # create basic attributes 567 | box_group = [] 568 | for box in raw_result: 569 | all_x = [int(coord[0]) for coord in box[0]] 570 | all_y = [int(coord[1]) for coord in box[0]] 571 | min_x = min(all_x) 572 | max_x = max(all_x) 573 | min_y = min(all_y) 574 | max_y = max(all_y) 575 | height = max_y - min_y 576 | box_group.append([ 577 | box[1], min_x, max_x, min_y, max_y, height, 0.5 * (min_y + max_y), 0 578 | ]) # last element indicates group 579 | # cluster boxes into paragraph 580 | current_group = 1 581 | while len([box for box in box_group if box[7] == 0]) > 0: 582 | # group0 = non-group 583 | box_group0 = [box for box in box_group if box[7] == 0] 584 | # new group 585 | if len([box for box in box_group if box[7] == current_group]) == 0: 586 | # assign first box to form new group 587 | box_group0[0][7] = current_group 588 | # try to add group 589 | else: 590 | current_box_group = [ 591 | box for box in box_group if box[7] == current_group 592 | ] 593 | mean_height = np.mean([box[5] for box in current_box_group]) 594 | # yapf: disable 595 | min_gx = min([box[1] for box in current_box_group]) - x_ths * mean_height 596 | max_gx = max([box[2] for box in current_box_group]) + x_ths * mean_height 597 | min_gy = min([box[3] for box in current_box_group]) - y_ths * mean_height 598 | max_gy = max([box[4] for box in current_box_group]) + y_ths * mean_height 599 | add_box = False 600 | for box in box_group0: 601 | same_horizontal_level = (min_gx <= box[1] <= max_gx) or (min_gx <= box[2] <= max_gx) 602 | same_vertical_level = (min_gy <= box[3] <= max_gy) or (min_gy <= box[4] <= max_gy) 603 | if same_horizontal_level and same_vertical_level: 604 | box[7] = current_group 605 | add_box = True 606 | break 607 | # cannot add more box, go to next group 608 | if not add_box: 609 | current_group += 1 610 | # yapf: enable 611 | # arrage order in paragraph 612 | result = [] 613 | for i in set(box[7] for box in box_group): 614 | current_box_group = [box for box in box_group if box[7] == i] 615 | mean_height = np.mean([box[5] for box in current_box_group]) 616 | min_gx = min([box[1] for box in current_box_group]) 617 | max_gx = max([box[2] for box in current_box_group]) 618 | min_gy = min([box[3] for box in current_box_group]) 619 | max_gy = max([box[4] for box in current_box_group]) 620 | 621 | text = "" 622 | while len(current_box_group) > 0: 623 | highest = min([box[6] for box in current_box_group]) 624 | candidates = [ 625 | box for box in current_box_group 626 | if box[6] < highest + 0.4 * mean_height 627 | ] 628 | # get the far left 629 | if mode == "ltr": 630 | most_left = min([box[1] for box in candidates]) 631 | for box in candidates: 632 | if box[1] == most_left: 633 | best_box = box 634 | elif mode == "rtl": 635 | most_right = max([box[2] for box in candidates]) 636 | for box in candidates: 637 | if box[2] == most_right: 638 | best_box = box 639 | text += " " + best_box[0] 640 | current_box_group.remove(best_box) 641 | 642 | result.append([ 643 | [ 644 | [min_gx, min_gy], 645 | [max_gx, min_gy], 646 | [max_gx, max_gy], 647 | [min_gx, max_gy], 648 | ], 649 | text[1:], 650 | ]) 651 | 652 | return result 653 | 654 | 655 | def printProgressBar( 656 | prefix="", 657 | suffix="", 658 | decimals: int = 1, 659 | length: int = 100, 660 | fill: str = "█", 661 | printEnd: str = "\r", 662 | ): 663 | """ 664 | Call in a loop to create terminal progress bar 665 | @params: 666 | prefix - Optional : prefix string (Str) 667 | suffix - Optional : suffix string (Str) 668 | decimals - Optional : positive number of decimals in percent complete (Int) 669 | length - Optional : character length of bar (Int) 670 | fill - Optional : bar fill character (Str) 671 | printEnd - Optional : end character (e.g. "\r", "\r\n") (Str) 672 | """ 673 | 674 | def progress_hook(count, blockSize, totalSize): 675 | progress = count * blockSize / totalSize 676 | percent = ("{0:." + str(decimals) + "f}").format(progress * 100) 677 | filledLength = int(length * progress) 678 | bar = fill * filledLength + "-" * (length - filledLength) 679 | print(f"\r{prefix} |{bar}| {percent}% {suffix}", end=printEnd) 680 | 681 | return progress_hook 682 | 683 | 684 | def reformat_input(image): 685 | """ 686 | :param image: image file path or bytes or array 687 | :return: 688 | img (array): (original_image_height, original_image_width, 3) 689 | img_cv_grey (array): (original_image_height, original_image_width, 3) 690 | """ 691 | if type(image) == str: 692 | if image.startswith("http://") or image.startswith("https://"): 693 | tmp, _ = urlretrieve( 694 | image, 695 | reporthook=printProgressBar( 696 | prefix="Progress:", 697 | suffix="Complete", 698 | length=50, 699 | ), 700 | ) 701 | img_cv_grey = cv2.imread(tmp, cv2.IMREAD_GRAYSCALE) 702 | os.remove(tmp) 703 | else: 704 | img_cv_grey = cv2.imread(image, cv2.IMREAD_GRAYSCALE) 705 | image = os.path.expanduser(image) 706 | img = load_image(image) # can accept URL 707 | elif type(image) == bytes: 708 | nparr = np.frombuffer(image, np.uint8) 709 | img = cv2.imdecode(nparr, cv2.IMREAD_COLOR) 710 | img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 711 | img_cv_grey = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) 712 | 713 | elif type(image) == np.ndarray: 714 | if len(image.shape) == 2: # grayscale 715 | img_cv_grey = image 716 | img = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR) 717 | elif len(image.shape) == 3 and image.shape[2] == 3: # BGRscale 718 | img = image 719 | img_cv_grey = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) 720 | elif len(image.shape) == 3 and image.shape[2] == 4: # RGBAscale 721 | img = image[:, :, :3] 722 | img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR) 723 | img_cv_grey = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) 724 | 725 | return img, img_cv_grey 726 | --------------------------------------------------------------------------------