├── .gitignore ├── README.md ├── artifact reduction-deblocking-denoising ├── DnCNN-TIP2017.md ├── README.md └── image │ ├── QQ截图20190131131042.jpg │ ├── QQ截图20190131152446.jpg │ └── image ├── engineering practices └── README.md ├── image-color enhancement └── README.md ├── image.png ├── super resolution ├── DBPN-CVPR2018.md ├── DRCN-CVPR2016.md ├── EDSR-CVPR2017.md ├── ESPCN-CVPR2016.md ├── FSRCNN-ECCV2016.md ├── LapSRN-CVPR2017.md ├── RCAN-ECCV2018.md ├── RDN-CVPR2018.md ├── README.md ├── SRCNN-ECCV2014.md ├── SRDenseNet-ICCV2017.md ├── SRGAN-CVPR2017.md ├── VDSR-CVPR2016.md └── image │ ├── 2018 │ └── tmp │ ├── QQ截图20181217172119.jpg │ ├── QQ截图20181217174219.jpg │ ├── QQ截图20181217195042.jpg │ ├── QQ截图20181217210829.jpg │ ├── QQ截图20181217214229.jpg │ ├── QQ截图20181218103827.jpg │ ├── QQ截图20181218110410.jpg │ ├── QQ截图20181218110516.jpg │ ├── QQ截图20181218113624.jpg │ ├── QQ截图20181218151406.jpg │ ├── QQ截图20181218152442.jpg │ ├── QQ截图20181218153657.jpg │ ├── QQ截图20181219144544.jpg │ ├── QQ截图20181219151026.jpg │ ├── QQ截图20181219151924.jpg │ ├── QQ截图20181219201012.jpg │ ├── QQ截图20181219202943.jpg │ ├── QQ截图20181220111456.jpg │ ├── QQ截图20181220165055.jpg │ ├── QQ截图20181220165322.jpg │ ├── QQ截图20181220165343.jpg │ ├── QQ截图20181220170523.jpg │ ├── QQ截图20181220170857.jpg │ ├── QQ截图20181220175338.jpg │ ├── QQ截图20181220190202.jpg │ ├── QQ截图20181220192743.jpg │ ├── QQ截图20181221104125.jpg │ ├── QQ截图20181221104633.jpg │ ├── QQ截图20181221105931.jpg │ ├── QQ截图20181221111337.jpg │ ├── QQ截图20181221112557.jpg │ ├── QQ截图20181221192238.jpg │ ├── QQ截图20181221192701.jpg │ ├── QQ截图20181221192709.jpg │ ├── QQ截图20181221193543.jpg │ ├── QQ截图20181221193917.jpg │ ├── QQ截图20181224110415.jpg │ ├── QQ截图20181224114316.jpg │ ├── QQ截图20181224115027.jpg │ ├── QQ截图20181224144650.jpg │ └── image.png └── testimg ├── face_img ├── shg_109922235_1047_e288cfc099f448688c2e849874d6vide.f0.png ├── shg_1103787868_1047_272ffcb61d234f26b8e1d95dad75vide.f0.png ├── shg_1141429069_1047_a6657948b36e4e729001cd408ed0vide.f0.png ├── shg_1155406328_1047_5827fa8f4bc24f6b814e2e16d9c9vide.f0.png ├── shg_1198794399_1047_208aeb28bc4d433baf6a925abbf7vide.f0.png ├── shg_1205440453_1047_cc345d90174545f6a862e89a8dd1vide.f0.png ├── shg_1338231904_1047_87a97b248985431499b165ad3ee5vide.f0.png ├── shg_1385532731_1047_8499b3cb0f714a6ead7911d78395vide.f0.png ├── shg_1417505285_1047_7293d7ec0207475d874b349e2d48vide.f0.png ├── shg_141884399_1047_20aa6a362e594ba1adcb4b2d847cvide.f0.png ├── shg_1421560562_1047_25d19e035b144f42b7aaabd0b7a0vide.f0.png ├── shg_1422167910_1047_1b347480431c41a8915f5fad2ee9vide.f0.png ├── shg_1519789614_1047_6157c89c278844a5921cbf452446vide.f0.png ├── shg_1547737132_1047_8cb248ce2b8043728182067bbfa3vide.f0.png ├── shg_1675283760_1047_c14a319eb6804c77bcb4cb65d203vide.f0.png ├── shg_1710966806_1047_89474bec69224104ab01077ca879vide.f0.png ├── shg_1759379157_1047_8868e3d17dce43de8a09a911420dvide.f0.png ├── shg_17923854_1047_2c77a6e431fe4764acbf39f6e611vide.f0.png ├── shg_1794846591_1047_c5535b3b8c824e07b58ced58382avide.f0.png ├── shg_180587087_1047_5c45ccca5c5340fcbba7c59b2b32vide.f0.png ├── shg_1870121582_1047_ce1b321ad7894546995609f07077vide.f0.png ├── shg_1945479187_1047_769151977b9b434a9fd669601657vide.f0.png ├── shg_1949720811_1047_55fcdab1af9d483288330573fcb5vide.f0.png ├── shg_1963722372_1047_53e430b7f99a4674a28bfd9fbc66vide.f0.png ├── shg_2019470413_1047_9880ae8ff04a48c8b2b1f5b6e798vide.f0.png ├── shg_2055501841_1047_640c99d5ee5744ef90e56fd32bdcvide.f0.png ├── shg_2074638923_1047_a9b376c564234fef9850e5ead5e4vide.f0.png ├── shg_207542132_1047_3c359231920a4fa7bf113c9b8a7dvide.f0.png ├── shg_272642512_1047_a36ce564aedf4440835c4d85555dvide.f0.png ├── shg_288036325_1047_52d7e34fce2f4994a7c6f0ec94a3vide.f0.png ├── shg_296658743_1047_60d77406ed98421b95543be07411vide.f0.png ├── shg_30425128_1047_08bf53dffc814d6b91fb717f4360vide.f0.png ├── shg_319326083_1047_29d9b3b71e50486291842b80dc96vide.f0.png ├── shg_41540261_1047_19fde467661e4acf9d6fdfc52db0vide.f0.png ├── shg_427881023_1047_cc14422f3cd44a6b86c6026e2e72vide.f0.png ├── shg_47075428_1047_ede999db12144950aa0d99b2d873vide.f0.png ├── shg_486841007_1047_856e831c39cb42b89bce9f70b2edvide.f0.png ├── shg_511368361_1047_f91aef381f8c493192533dbf3a40vide.f0.png ├── shg_517307463_1047_d0962e3941dc4a6a9b504a7bad52vide.f0.png ├── shg_528635464_1047_3dbc904b8ad240eab7fabb2a0fc6vide.f0.png ├── shg_56082514_1047_3e1c1828e6f6484597c4d9f510c5vide.f0.png ├── shg_607450128_1047_e3daaf55c0604bc18bf64b9daafcvide.f0.png ├── shg_618483583_1047_ee8dd7eabd524e2da2bb4f944d94vide.f0.png ├── shg_689821267_1047_eb4a46c128ac4b5fbe3285260a58vide.f0.png ├── shg_728975743_1047_14773a7f5bda4a798d1fff8d7b57vide.f0.png ├── shg_730324952_1047_a0bdc8fb91a9442f86108eae9b7evide.f0.png ├── shg_732197352_1047_552e497a8c404a0cb362b06f300fvide.f0.png ├── shg_739699816_1047_17c7549339c84862a924b14d6899vide.f0.png ├── shg_746830622_1047_69d4dbc55efe4f5f88da24dee27evide.f0.png ├── shg_75109208_1047_8025a4abf7554f59b496744a022fvide.f0.png ├── shg_786224940_1047_8b090b3fe6764ac6bceb7c49d022vide.f0.png ├── shg_844600188_1047_bf1b247e40004553ad51b800e8a7vide.f0.png ├── shg_850139819_1047_219886a658b045509d37958e0ee6vide.f0.png ├── shg_901052700_1047_51242db02058439e8dfefc265022vide.f0.png ├── shg_921792563_1047_10dafa825fa34453a5eae5bbbeeevide.f0.png ├── shg_925784094_1047_f4f0c8a64bf94a06b690adefe572vide.f0.png ├── shg_936294450_1047_986afd82d2e14887a4571b26dc7bvide.f0.png ├── shg_943430993_1047_36282002403349c19209684d90d8vide.f0.png ├── shg_961532195_1047_975a8a2e5fa14544a0ae4c86565cvide.f0.png ├── shg_963030462_1047_4a11175e76d8408b9b075f298587vide.f0.png ├── tjg_111430007_1047_18bdceb379eb44b8b088b77ab3dcvide.f0.png ├── tjg_1346781597_1047_b7d5f768c76d47f3824fb0db844fvide.f0.png ├── tjg_1370035883_1047_b821a698143f413cbdf8ed3f29aavide.f0.png ├── tjg_1538781621_1047_d13dc2c686014960a5dc6a481e00vide.f0.png ├── tjg_1541618258_1047_71edf7553b2144eb967c633fa6f1vide.f0.png ├── tjg_1550436651_1047_7aafee7e7efc4021a39eaaaccc28vide.f0.png ├── tjg_156613776_1047_fbe34f9af7614eebb97383c3c738vide.f0.png ├── tjg_1585686676_1047_9c7cba6a7776464ebda5622a8028vide.f0.png ├── tjg_1649110556_1047_a9cd725e4c814a1dbd24d27c0386vide.f0.png ├── tjg_1795246937_1047_997aa204dbcb4206bcd8f250cbf2vide.f0.png ├── tjg_1917340715_1047_a7ef639c50d94a558e379b2adb12vide.f0.png ├── tjg_1966027031_1047_095f5ceaa0ba4b17857b85e2932evide.f0.png ├── tjg_2031177283_1047_a8563f4f0f5b46f2a18beed263c2vide.f0.png ├── tjg_2096875702_1047_8435d9fdf7144129bc0dd8acdde4vide.f0.png ├── tjg_2109838680_1047_11c7bfc133a34cf894c1d6a487a5vide.f0.png ├── tjg_250968726_1047_b4db44fb41454f8eb146fedbbf98vide.f0.png ├── tjg_29808111_1047_e0106dcffcd947da91409f934868vide.f0.png ├── tjg_417044029_1047_bc2ea64ac35d470393fd7d4ef309vide.f0.png ├── tjg_502311402_1047_c57b480d79354a31a03c8cb38b4fvide.f0.png ├── tjg_515353011_1047_f8b08f6220d84b0aacd8178bd532vide.f0.png ├── tjg_516558712_1047_eec38c5f00c64fb7a51681204ed4vide.f0.png ├── tjg_555070312_1047_82fd6f192f8a41a7a444e273fe56vide.f0.png ├── tjg_580342898_1047_712bb4acf6fd47a8b6a2f33f010cvide.f0.png ├── tjg_603434954_1047_c191f982a7974775abc05d71c4e1vide.f0.png ├── tjg_611863320_1047_445efb3125e64a9d815d0c3b9dd2vide.f0.png ├── tjg_718668875_1047_15c4a78bebf4444ab061613bae49vide.f0.png ├── tjg_772722521_1047_82e647e527d04484b42236a41905vide.f0.png ├── tjg_792711838_1047_a3704a779101425fb7a719f4ff31vide.f0.png ├── tjg_896054223_1047_4f27d6012de049bb9a74baf688d4vide.f0.png └── tjg_927249347_1047_cdd5726846054ec0aec90f0527d3vide.f0.png └── readme.md /.gitignore: -------------------------------------------------------------------------------- 1 | # Prerequisites 2 | *.d 3 | 4 | # Compiled Object files 5 | *.slo 6 | *.lo 7 | *.o 8 | *.obj 9 | 10 | # Precompiled Headers 11 | *.gch 12 | *.pch 13 | 14 | # Compiled Dynamic libraries 15 | *.so 16 | *.dylib 17 | *.dll 18 | 19 | # Fortran module files 20 | *.mod 21 | *.smod 22 | 23 | # Compiled Static libraries 24 | *.lai 25 | *.la 26 | *.a 27 | *.lib 28 | 29 | # Executables 30 | *.exe 31 | *.out 32 | *.app 33 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # AI-video-enhance 2 | This repository collects the state-of-the-art algorithms for video/image enhancement using deep learning (AI) in recent years, including reviews and engineering practices. 3 | 4 | 视频/图像增强手段包括去伪影/去块/去噪(artifacts reduction, deblocking, denoising)、超分辨率(super resolution)、图像/色彩增强(image/color enhancement)，这里主要针对这几种领域收集一些优秀的AI算法，并在工程上进行一些代码试验。 5 | 6 | # super resolution 7 | 8 | - [SRCNN-ECCV2014](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRCNN-ECCV2014.md) 9 | - [ESPCN-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/ESPCN-CVPR2016.md) 10 | - [FSRCNN-ECCV2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/FSRCNN-ECCV2016.md) 11 | - [VDSR-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/VDSR-CVPR2016.md) 12 | - [DRCN-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/DRCN-CVPR2016.md) 13 | - [LapSRN-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/LapSRN-CVPR2017.md) 14 | - [SRDenseNet-ICCV2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRDenseNet-ICCV2017.md) 15 | - [SRGAN-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRGAN-CVPR2017.md) 16 | - [EDSR-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/EDSR-CVPR2017.md) 17 | - [DBPN-CVPR2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/DBPN-CVPR2018.md) 18 | - [RDN-CVPR2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/RDN-CVPR2018.md) 19 | - [RCAN-ECCV2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/RCAN-ECCV2018.md) 20 | 21 | # 相关资源 22 | - **视频增强** https://github.com/yulunzhang/video-enhancement 23 | - **压缩编码** https://github.com/flyywh/Image-compression-and-video-coding 24 | - **video SR** 25 | [https://github.com/LoSealL/VideoSuperResolution](https://github.com/LoSealL/VideoSuperResolution) 26 | - **image SR** 27 | [https://github.com/icpm/super-resolution](https://github.com/icpm/super-resolution) 28 | [https://github.com/YapengTian/Single-Image-Super-Resolution](https://github.com/YapengTian/Single-Image-Super-Resolution) 29 | - **Awesome Deep Vision** https://github.com/kjw0612/awesome-deep-vision 30 | - **Awesome Computer Vision** https://github.com/jbhuang0604/awesome-computer-vision 31 | - **denoising** https://github.com/flyywh/Image-Denoising-State-of-the-art 32 | 33 | -------------------------------------------------------------------------------- /artifact reduction-deblocking-denoising/DnCNN-TIP2017.md: -------------------------------------------------------------------------------- 1 | # DnCNN-TIP2017 # 2 | 3 | Beyond a Gaussian Denoiser,号称不仅仅适用于高斯白噪声，也可用于未知level的高斯白噪声。所以文中用同一个DnCnn同时训练几种任务：Blind Gaussian denoising（多level）, single image super-resolution（SISR ,降采样再上采样产生的噪声,多尺度） and JPEG image deblocking（多种质量系数）。文中也提及了， 很多深度学习去噪算法如MLP、TRND都是针对某level的噪声训练一个模型，不太通用。 4 | 5 | 6 | DnCnn利用残差学习和批处理归一化加速训练过程，提高去噪性能。 7 | 如图网络结果比较简单，第一层使用conv+relu，然后接着d层的conv+batch normalization+relu,最后再一个conv层。conv使用3x3的核，最后一层卷积生成1通道的噪声残差，然后与输入相减得到去噪的图。其中每个卷积的特征图个数为64。 8 | 9 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/artifact%20reduction-deblocking-denoising/image/QQ%E6%88%AA%E5%9B%BE20190131131042.jpg) 10 | 11 | **实验部分** 12 | 13 | DnCNN-S：高斯噪声level分为σ = 15,25 and 50三个模型，使用400张180x180的图片生成不同level的训练数据（实验中说大量的训练数据仅仅带来微弱的效果提升），其中图片分成40x40的块，batchs大小为128。 14 | DnCNN-B：单个模型用于blind Gaussian denoising，噪声级别范围为σ ∈ [0, 55]，图片分成50x50的块，batchs大小为128。 15 | DnCNN-3：单模型处理三种任务，包括高斯盲去噪、SISR、JPEG deblocking,盲去噪产生[0, 55]级别的数据集，SISR使用factors2，3，4的上下采样，deblocking使用quality factor为5 to 99 using the MATLAB 16 | JPEG encoder。块大小为50x50，batchs大小为128。 17 | 18 | 看效果。 19 | 20 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/artifact%20reduction-deblocking-denoising/image/QQ%E6%88%AA%E5%9B%BE20190131152446.jpg) 21 | 22 | ## 开源实现 ## 23 | 24 | **论文链接** [https://arxiv.org/pdf/1608.03981v1.pdf](https://arxiv.org/pdf/1608.03981v1.pdf) 25 | **tensorflow** [https://github.com/wbhu/DnCNN-tensorflow](https://github.com/wbhu/DnCNN-tensorflow) 26 | 27 | **PyTorch** [https://github.com/SaoYan/DnCNN-PyTorch](https://github.com/SaoYan/DnCNN-PyTorch) 28 | 29 | **keras** [https://github.com/husqin/DnCNN-keras](https://github.com/husqin/DnCNN-keras) 30 | 31 | **matlab** [https://github.com/cszn/DnCNN](https://github.com/cszn/DnCNN) 32 | 33 | -------------------------------------------------------------------------------- /artifact reduction-deblocking-denoising/README.md: -------------------------------------------------------------------------------- 1 | 去噪声相关 2 | -------------------------------------------------------------------------------- /artifact reduction-deblocking-denoising/image/QQ截图20190131131042.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/artifact reduction-deblocking-denoising/image/QQ截图20190131131042.jpg -------------------------------------------------------------------------------- /artifact reduction-deblocking-denoising/image/QQ截图20190131152446.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/artifact reduction-deblocking-denoising/image/QQ截图20190131152446.jpg -------------------------------------------------------------------------------- /artifact reduction-deblocking-denoising/image/image: -------------------------------------------------------------------------------- 1 | 2 | -------------------------------------------------------------------------------- /engineering practices/README.md: -------------------------------------------------------------------------------- 1 | # engineering practices 2 | 3 | 一些工程实践 4 | -------------------------------------------------------------------------------- /image-color enhancement/README.md: -------------------------------------------------------------------------------- 1 | 图像增强，色彩增强相关。 2 | -------------------------------------------------------------------------------- /image.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/image.png -------------------------------------------------------------------------------- /super resolution/DBPN-CVPR2018.md: -------------------------------------------------------------------------------- 1 | # DBPN # 2 | 3 | 文中说，现有的深度前馈 feed-forward SR网络可以学习低分辨率输入的表示，以及从低分辨率输入到高分辨率输出的非线性映射关系，然而却不能完全解决LR与HR之间相互的依赖关系（没有反馈feedback connections连接--）。 4 | 5 | 所以使用Deep Back-Projection Networks 网络来探索up-和down-采样层，为每个阶段的投影误差projection errors提供误差反馈机制。上采样和下采样阶段是相互连接的，每个阶段代表不同类型的图像退化和高分辨率组件。好高深，继续虐一下。 6 | 7 | DBPN使用了iterative errorcorrecting feedback mechanism，迭代误差校正机制，同时计算上采样、下采样的投影误差projection errors来指导重建高质量的HR，这里下面再解释。 8 | 9 | 上下采用之间相互连接（不像前馈SR结构都是只有input到output的连接映射，这种方式在大倍上采样中首先与LR空间的有限特征，所以不能很好求得LR到HR的映射关系），所以DBPN不仅用上采样层获得HR特征，同时也用下采样层获得LR的反馈特征。 10 | 11 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221104125.jpg) 12 | 13 | 为什么要这样呢？作者总结了目前SR结构的分类，如上图： 14 | 15 | - （a）为预先通过传统插值如bicubic来上采样得到预超分的图，再通过卷积学习到最后的HR预测结果，如SRCNN、VDSR等。但是传统插值上采样也许会引起噪音。 16 | - （b）为LR图像经过不断特征学习与传播，最后通过反卷积或sub-pixel卷积来重建HR结果，如FSRCNN、ESPCN、EDSR等。这种方式受限于网络深度和容量，所以学习不到非常复杂的特征映射关系（比如EDSR也是这种方式，对于高倍上采样就需要非常多层卷积和每次非常多的特征图）。 17 | - （c）为金字塔模式，如LapSRNA，通过单向推理得的逐级学习特征然后上采样x2的方式，多级后最后可以得到多倍的HR图像 18 | - （d）本文提出的方式，首先输入LR经过卷积提取特征，然后使用Back-projection的迭代来最小化重建误差，分为up-和downprojection，不断迭代（每个阶段由一个up-和一个downprojection构成），这种结构通过将构建误差分布到每个阶段，得到不同深度的SR特征，然后不同深度的SR特征通过concat连接起来，最后经过卷积获得最好逇HR预测。 19 | 20 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221104633.jpg) 21 | 22 | 如上图，up-projection单元，由阶段t-1到t的过程是，首先低分辨率图像Lt-1经过反卷积上采样到高分辨率H0t，H0t经过卷积降采样到低分辨率L0t，低分辨率L0t与输入Lt-1计算残差（error），然后残差经过反卷积得到高分辨率残差H1t，然后与高分辨率的H0t相加得到输出t阶段的上采样高分辨率图像Ht 23 | 24 | 25 | 同理down-projection单元，首先高分辨率图像H经过卷积下采样到低分辨率L0t，然后反卷积到高分辨率H0t并与Ht计算高分辨率残差，残差经过卷积下采样得到低分辨率残差L1t，然后与低分辨率L0t相加得到t阶段的下采样的低分辨率Lt图像。 26 | 27 | 这两个单元使用比较大的卷积核，如8* 8或12* 12。大卷积核因为难以收敛和可能产生局部最优解的缺点很少在先前文献中被使用，但是这里的迭代单元可以抑制这种缺点。 28 | 29 | 30 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221111337.jpg) 31 | 32 | 介绍完两个组件，再看文中还提出了Dense DBPN来改善提升DBPN效果（为了减缓梯度下降了，受启发于DenseNets）。其实也就是在原来网络加入dense，所有up-projection输出都连接到后面的down-projection作为输入(特征reuse)，同理down-projection的输出都连接到后面的up-projection作为输入。如上图所示为Dense projection units，前面阶段的输出concat作为此阶段的输入，其中不同原始DenseNets，D-DBPN避免使用dropout和batch norm，因为它们移除了特征范围，不适合SR任务，所以这里concat后使用1* 1的卷积来feature pooling and dimensional reduction。据说此法非常有效，最后看下图为D-DBPN的网络结构。 33 | 34 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221105931.jpg) 35 | 36 | 直接看效果，X8超分，效果顶呱呱的 37 | 38 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221112557.jpg) 39 | 40 | ## 开源实现 ## 41 | 42 | **论文链接** [https://www.toyota-ti.ac.jp/Lab/Denshi/iim/members/muhammad.haris/projects/DBPN.html](https://www.toyota-ti.ac.jp/Lab/Denshi/iim/members/muhammad.haris/projects/DBPN.html) 43 | 44 | **tensorflow** [https://github.com/tlokeshkumar/DBPN-tf](https://github.com/tlokeshkumar/DBPN-tf) 45 | 46 | 47 | **PyTorch** [https://github.com/alterzero/DBPN-Pytorch](https://github.com/alterzero/DBPN-Pytorch) 48 | 49 | **caffe** [https://github.com/alterzero/DBPN-caffe](https://github.com/alterzero/DBPN-caffe) 50 | 51 | **Keras** [https://github.com/rajatkb/DBPN-Keras](https://github.com/rajatkb/DBPN-Keras) 52 | -------------------------------------------------------------------------------- /super resolution/DRCN-CVPR2016.md: -------------------------------------------------------------------------------- 1 | # DRCN # 2 | 3 | 和VDSR同一个作者，网络深度差不多，就是结构上采用了Recursive Neural Network递归神经网络。 4 | 5 | 作者认为，多数计算机视觉任务通过增加网络深度来扩大感受野，通常一个卷积a conv layer增加了参数量，然后接着一个池化a pool layer来丢弃像素信息。 然而，super-resolutionand denoising超分、去噪等图像重建问题是非常看重图像细节的，因此一般DL都不会使用池化层，那么纯粹增加卷积层指挥增加更多参数量，一方面，容易导致过拟合，所有需要更多的训练数据，另一方面huge参数使模型难以存储和恢复。 6 | 7 | 所以为了解决这些问题，DRCN使用了a deeply-recursive convolutional network递归卷积网络。通过多次重复应用相同的卷积层（而不会增加参数量了），网络感受野达到41* 41（其中SRCNN的是13* 13）。 8 | 9 | 当然，另一个问题梯度爆炸还是会有的，训练收敛很难达到，文中使用了两种方法解决：首先，所有递归都设计成有监督的方式，每个递归都进行一次HR预测的重建，然后所有递归结果求均值得到最后的HR预测图像。另外，像VDSR一样，使用了resnet残差思想来重建HR图像（在超分这种场景input与output之间高度相关时，残差尤其有效）。 10 | 11 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218151406.jpg) 12 | 13 | 看图所示，与VDSR类似，先进行传统插值放大到目标HR的分辨率大小，DRCN结构：首先是embedding network，两层卷积用于提取输入图像的特征图，然后是inference network，就是递归卷积结构了，所有递归层使用同一个参数权重W，最后是reconstruction network，重建输出HR。 14 | 15 | 看下面展开递归部分，三个图实属同理，其中H1到HD使用同样的权重W，然后每个递归层d都加上input进行局部重建HR（也就是每个递归都经过ReconNet），然后所有重建进行求和去均值来获得最后HR的输出（称为重建层共享所有递归的局部预测）。最后的loss也设计为D+1个输出（D个递归，1个最终输出）与ground true的MSE函数。 16 | 17 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218152442.jpg) 18 | 19 | DRCN的主要特征：recursive-supervision and skipconnection，递归监督和残差跳连接 20 | 21 | 实验部分：文中使用了16个递归，所以大概输入到输出经过了20个卷积，所有卷积核大小为3* 3，训练图像以步长为21分割为41* 41大小的patchs，然后以64批量输入进行随机梯度训练。学习率初始化为0.01，然后每5次epochs进行一次除以10的衰减，最小为10^-6。Titan X GPU上训练用了6天，擦了。 22 | 23 | 直接上效果。 24 | 25 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218153657.jpg) 26 | 27 | 28 | ## 开源实现 ## 29 | 30 | **论文链接** [https://cv.snu.ac.kr/research/DRCN/](https://cv.snu.ac.kr/research/DRCN/) 31 | 32 | **tensorflow** [https://github.com/jiny2001/deeply-recursive-cnn-tf](https://github.com/jiny2001/deeply-recursive-cnn-tf) 33 | 34 | -------------------------------------------------------------------------------- /super resolution/EDSR-CVPR2017.md: -------------------------------------------------------------------------------- 1 | # EDSR # 2 | 3 | 又一个resnet的超分结构，NTIRE2017的冠军作品。 4 | 5 | based on the SRResNet architecture, 6 | we first optimize it by analyzing and removing 7 | unnecessary modules to simplify the network architecture. 8 | 9 | 是基于SRResNet（SRGAN的生成网络）的优化，移除了一些不必要的结构，简化Residual blocks结构。 10 | 11 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220175338.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220175338.jpg) 12 | 13 | 如图，原始res block是经过卷积、批量归一化（BN）、激活ReLU、卷积、BN，然后加上输入并最后relu激活得到输出；SRResNet的block去掉了最后一个激活，EDSR则去掉了所有BN，所以网络结构更简洁（特别BN耗内存也耗时），这样就可以增加更多的层来训练更复杂的网络。 14 | 15 | EDSR整个网络结构如下图所示，整个结构基于SRResNet，相比之下，ResBlock去除了BN层，并且block的特征数相同，如64或256，一共指定了32个blocks。 16 | 17 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220190202.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220190202.jpg) 18 | 19 | 看结构，是LR输入一个conv提取特征，然后B个resblock提取全局的高层特征，然后一个conv后加上低层特征作为特征的学习结果，然后上采样后，进行一次conv得到HR图像，其中上采样支持X2、X3和X4等倍数（另外，文中也提供了MDSR多倍数超分的结构）。 20 | 21 | 22 | **实验部分** 23 | 24 | 使用了最近发布的2K图DIV2K数据集，然后将图像裁剪为48* 48的RGB通道的LR图像，同时进行翻转和旋转扩展数据集（注意，是使用了RGB三通道，不是Y通道而已了）。loss函数使用了L1函数。 25 | 26 | 结果又多屌直接看图 27 | 28 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220192743.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220192743.jpg) 29 | 30 | ## 开源实现 ## 31 | 32 | **论文链接** [https://arxiv.org/abs/1707.02921](https://arxiv.org/abs/1707.02921) 33 | 34 | **tensorflow** [https://github.com/jmiller656/EDSR-Tensorflow](https://github.com/jmiller656/EDSR-Tensorflow) 35 | 36 | 37 | **PyTorch** [https://github.com/thstkdgus35/EDSR-PyTorch](https://github.com/thstkdgus35/EDSR-PyTorch) 38 | 39 | **Torch ** [https://github.com/LimBee/NTIRE2017](https://github.com/LimBee/NTIRE2017) 40 | 41 | -------------------------------------------------------------------------------- /super resolution/ESPCN-CVPR2016.md: -------------------------------------------------------------------------------- 1 | # ESPCN # 2 | 3 | we present the first convolutional neural network 4 | (CNN) capable of real-time SR of 1080p videos on a single 5 | K2 GPU. 6 | 7 | 作者声称K2 GPU上可以做到1080P视频的实时超分。 8 | 9 | ESPCN也是最后一层才做上采样，意味着输入和特征映射等都是LR分辨率的大小，这样可以提升速度效率，并可以用更小的核覆盖更大的区域信息，从而更有可能达到实时。 10 | 11 | 文中说了，很多传统算法在第一层先将LR图像经过bicubic为HR图像作为CNN网络的输入，一方面加大了计算耗时和内存消耗，另一方面，经过bicubic后的LR图像并没有带来更好的用于超分学习的输入信息，甚至可能容易带来一些噪点。所以文中直接是端到端的LR到HR的CNN超分网络，通过网络学习LR到HR的直接映射关系，也就是所有细节、轮廓等信息都通过网络来学习。ESPCN通过L个层来学习输入图像LR的n个特征图，也更加容易学习到更复杂更好的LR到HR的映射关系，最后一层通过文中提出的亚像素卷积来重建超分图像HR。 12 | 13 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181217210829.jpg) 14 | 15 | 文中将HR图像通过r倍的scale降采样为LR图像，然后经过高斯模糊做了处理。LR表示为H × W × C的张量，HR表示为rH × rW × C的张量。然后对C个通道（为了实时，文中仅用了图像的Y通道）的LR图像做了L层（L文中取3）的卷积学习，第一层的卷积核为5×5，输出特征图个数64，第二层的卷积核为3×3，输出特征图个数32，最，最后一层卷积核3×3，最后的输出特征图个数为C×r×r（这里假设超分2倍，r=2，C=1，那么输出的特征图个数为4个）。 16 | 17 | 对学习到的特征图，最后使用 Efficient sub-pixel convolution layer：亚像素卷积来重建HR图像，直接看下图，其实就是每个通道有r×r个特征图，然后r×r个特征图相同位置的所有像素点（r×r个像素）排列成长为r宽为r的单通道像素图，这就是亚像素重建。 18 | 19 | 本算法只考虑y通道的超分。 20 | 21 | ![效果](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181217214229.jpg) 22 | 23 | ## 开源实现 ## 24 | 25 | **论文链接** [https://arxiv.org/pdf/1609.05158.pdf](https://arxiv.org/pdf/1609.05158.pdf) 26 | 27 | **tensorflow** [https://github.com/drakelevy/ESPCN-TensorFlow](https://github.com/drakelevy/ESPCN-TensorFlow) [https://github.com/kweisamx/TensorFlow-ESPCN](https://github.com/kweisamx/TensorFlow-ESPCN) 28 | 29 | **PyTorch** [https://github.com/leftthomas/ESPCN](https://github.com/leftthomas/ESPCN) 30 | 31 | **caffe** [https://github.com/wangxuewen99/Super-Resolution/tree/master/ESPCN](https://github.com/wangxuewen99/Super-Resolution/tree/master/ESPCN) 32 | 33 | -------------------------------------------------------------------------------- /super resolution/FSRCNN-ECCV2016.md: -------------------------------------------------------------------------------- 1 | # FSRCNN # 2 | 3 | FSRCNN是对SRCNN的一个改进版，速度更快，性能更好。FSRCNN与SRCNN都是香港中文大学Dong Chao等人提出的。 4 | 5 | 1.不需要像SRCNN那样经过bicubic前处理成目标分辨率的大小，可以直接原图LR卷积生成特征图，然后在最后一层通过反卷积层直接将低分辨率特征图生成高分辨率的超分结果图HR。 6 | 2.将中间maping层前面加入shrinking和后面加入expanding，可以减少需要进行maping的特征维度，降低运算量。 7 | 3.增加了mapping的层数，减少卷积核大小，这里因为特征图大小是LR大小，更小的卷积核就可以达到SRCNN那种9*9卷积核覆盖到的信息，并且通过多层mapping学习更加复杂的特征图之间的映射关系。 8 | 9 | FSRCNN相对于SRCNN速度上提升了40倍左右，并且超分质量也得到保障，号称是CPU下可以达到实时速度的（24fps）。 10 | 11 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181217195042.jpg) 12 | 13 | 图中可以看出三个主要改变： 14 | 15 | - 原始小图直接输入，不用bicubic处理。然后是特征提取层，SRCNN采用了9*9的核，因为其输入是通过bicubic的处理，所以卷积核比较大（9*9大小），在FSRCNN就采用了5*5的核，图像输入通道为1（Y通道），输出特征图个数为d个。 16 | - 非线性mapping改为shringing层、多层mapping、expanding层组成。改进SRCNN中mapping特征通道个数过大的缺点（导致运算量大），FSRCNN首先使用1*1的卷积做特征线性融合处理，输出s个通道的特征图，s远小于d，这样相当于降低特征维度；然后经过多个非线性mapping的映射学习保证质量，这里用了m个3*3的卷积层来mapping，输出同样为s通道的特征图，最后通过1*1的卷积做expanding处理，生成了d通道特征图（其实是shringing的逆运算，做特征图个数的恢复），据文中提到可以提升0.3db的psnr指标。 17 | - 最后一层采用反卷积层重建高分辨率图像，可以很方便指定多种倍数的超分（X2、X3、X4等），这里使用的卷积核为9*9。 18 | - 另外FSRCNN使用了PRelu激活函数，不同于SRCNN使用的Relu激活函数f(xi)=max(xi,0)，PRelu定义为f(xi)=max(xi,0)+ai*min(0,xi)，ai用于当xi为负数的情况下。PRelu主要避免了Relu中0梯度引起的“dead features”，可以充分利用所有网络参数。 19 | 20 | ## 训练与实验 ## 21 | 文中对训练集合做了一些扩张，首先提出了一个新的数据集，包含100张轮廓清晰、少平滑区域（如天空、大海）的适合超分训练的图像。然后还通过scaling（0.6-0.9）、rotation（90-270度）手段来扩充数据集。 22 | 这里将原清晰大图下采样为指定倍数n的小图，然后再分割为很多块patchs，用于与对应的清晰ground truth组合pairs作为训练集。 23 | 24 | ![不同参数下的性能表现](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218103827.jpg) 25 | 26 | ## 开源实现 ## 27 | 28 | **算法主页** [http://mmlab.ie.cuhk.edu.hk/projects/FSRCNN.html](http://mmlab.ie.cuhk.edu.hk/projects/FSRCNN.html) 29 | 30 | **tensorflow** [https://github.com/yifanw90/FSRCNN-TensorFlow](https://github.com/yifanw90/FSRCNN-TensorFlow) 31 | 32 | **PyTorch** [https://github.com/yippp/FSRCNN](https://github.com/yippp/FSRCNN) 33 | 34 | -------------------------------------------------------------------------------- /super resolution/LapSRN-CVPR2017.md: -------------------------------------------------------------------------------- 1 | # LapSRN # 2 | 金字塔超分算法。特点： 3 | 4 | - 不要求input做传统的bicubic插值，因此可以降低计算复杂度 5 | - 使用Charbonnier损失函数来训练，可以获得很高质量的重建 6 | - 单个前向推理可以实现多种scale倍数的超分预测 7 | 8 | 作者认为现在超分的方法存在几个问题： 9 | 10 | - 很多算法预先使用bicubic将LR插值到目标HR的大小，然后再进行预测推理，这个前处理加大了不必要的计算损耗，输入大分辨率的图，也让推理更慢，传统插值也很可能带来视觉伪影。当然，有些算法使用deconvolution反卷积来重建HR，或使用sub-pixel convolution来重建HR，这些办法受限于网络容量所以并不能学习到input到output的很复杂的 映射关系。 11 | - 一般算法使用l2范数来作为loss，因此很容易使预测变得模糊，这也是因为l2很难捕获HR图像块的underlying multi-modal distributions（底层多模态分布），导致学习到的预测相对人类视觉感官上比较平滑。 12 | - 另外，现存方法对重建HR一般只做一次上采样操作，这让产生大倍数的超分重建非常困难（如8倍），此外，这些方法也不能产生多个分辨率的中间预测，所以要适应不同倍数超分的需求就要训练多个模型，使工程上落地更加困难。 13 | 14 | 15 | LapSRN基于CNN级联的方式实现一个类金字塔支持多scale级别的的超分模型，每一个级别中进行一次X2的上采样，3个级别就可以达到X8的超分预测了。 16 | 17 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181219144544.jpg) 18 | 19 | 1. 如图，第一级别中，输入图片经过几个卷积提取特征图，然后经过a transposed convolutional layer反卷积来上采样2倍得到X2的特征图，这里出现2个红色箭头。 20 | 1. 向下箭头为一个卷积预测LR到HR的残差residuals，并与LR经过反卷积上采样的图相加得到这一级别的HR预测，此时得到X2的超分图。 21 | 1. 另一个向右箭头继续重复1的过程，继续提取特征，这也是下一个级别预测了。多个级别就组成了X4、X8的超分预测了。 22 | 2. 最后的loss函数使用Charbonnier损失函数来改善性能。看下图和看原文。 23 | 3. 每层上采样都使用一个事先初始化的the bilinear kernel 来做卷积。 24 | 25 | ![loss函数](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181219151026.jpg) 26 | 27 | LapSRN对于X4超分没有直接上采样，而是采样级别方式，两次级联的X2上采样，这样得到的效果会更好。 28 | 29 | 30 | **实验部分** 31 | 文中每个卷积层使用3* 3的核，生成64个特征图。 反卷积核初始化为4*4 的 a bilinear filter双线性滤波器。每个卷积或反卷积后面紧跟着leaky rectified linear units (LReLUs) with a negative slope of 0.2。 32 | 33 | 训练集部分使用64个128* 128的小patchs作为一组随机梯度优化学习。然后训练集做了Scaling缩放、Rotation旋转、Flipping翻转等。 34 | 35 | 总体看，LapSRNA通过级联上采样，逐级预测残差的方式，在做高倍超分也能输出中间的超分预测。逐级上采样使得其速度更快，每一级别都有loss监督使得其效果更佳。 36 | 37 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181219151924.jpg) 38 | 39 | ## 开源实现 ## 40 | 41 | **论文链接** [http://vllab.ucmerced.edu/wlai24/LapSRN/](http://vllab.ucmerced.edu/wlai24/LapSRN/) 42 | 43 | **tensorflow** [https://github.com/zjuela/LapSRN-tensorflow](https://github.com/zjuela/LapSRN-tensorflow) 44 | 45 | **pytorch** [https://github.com/twtygqyy/pytorch-LapSRN](https://github.com/twtygqyy/pytorch-LapSRN) 46 | 47 | -------------------------------------------------------------------------------- /super resolution/RCAN-ECCV2018.md: -------------------------------------------------------------------------------- 1 | # RCAN # 2 | 3 | 非常深的residual channel attention networks（RCAN），文中两个亮点： 4 | 5 | - residual in residual (RIR) structure基于残差的残差结构 6 | - channel attention mechanism通道注意力机制 7 | 8 | 作者认为目前大部分算法 treat channel-wise features equally，就是对每个通道特征都是同样的重要程度，这样缺乏灵活处理不同类型的信息，如高频、低频信息。往往高频细节信息更难学习，图像HR可以看做是恢复尽可能多高频信息的一个过程，因为LR包括的低频信息可以直接传给HR。所以RCAN提出了channel attention mechanism作用于不同通道的特征图，提高不同通道的特征图之间的差异性。而RIR的结构是解决非常深的网络提出来的一种结构。 9 | 10 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181224110415.jpg) 11 | 12 | 如图，RCAN包括浅层特征提取（用一层卷积）、RIR深层特征提取、上采样和重建三个部分。 13 | 14 | RIR结构为核心部分，包括了G个 residual groups (RG) 和一个 long skip connection (LSC，看做全局残差)，每个RG又包括B个 residual channel attention blocks (RCAB）和一个 short skip connection (SSC，看做局部残差)。RIR结构这种残差堆起来的方式允许网络深度达到400层以上，这就可以获得超好的性能了。 15 | 16 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181224114316.jpg) 17 | 18 | RCAB如上图所示，首先经过一个卷积，一个relu激活，然后一个卷积，最后接着Channel Attention（CA）结构，最后与输入相加。CA负责提取所有特征通道之间的统计信息来增强每个特征通道，提高通道之间的区分能力（Channel attention）。就是没有类似现有算法那样，将每个通道都同等对待。 19 | 20 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181224115027.jpg) 21 | 22 | 如图，channel attention mechanism通道注意力的机制（本文主要的亮点之处），首先使用全局池化HGP统计z，也就是每通道H* W的特征图变成了1* 1的一个统计值。然后接着一个卷积WD将特征通道数降维channel-downscaling为C/r，然后使用relu激活范数，接着一个卷积WU将通道扩维 channel-upscaling为C，接着使用sigmoid函数作用后得到最后的sc，最后作用于输入（与输入矩阵进行点乘）。整个CA过程就是通过某些机制来给每个通道的feature计算一个channel attention值sc，最后作用于对应的通道特征图，起到通道之间的特征程度改变（重要度改变了）。 23 | 24 | **实现细节** 25 | 26 | 27 | 文中使用G10个RG结构，每个RG结构使用20个RCAB。除了channel-downscaling与 channel-upscaling（卷积核1* 1，特征数C/r=4），所有卷积核使用3* 3，特征数为C=64。最后输出三通道彩色图。训练集使用2K图DIV2K。 28 | 29 | 如图，可以看到细节学习能力非常强。 30 | 31 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181224144650.jpg) 32 | 33 | ## 开源实现 ## 34 | 35 | **论文链接** [https://arxiv.org/pdf/1807.02758.pdf](https://arxiv.org/pdf/1807.02758.pdf) 36 | 37 | **tensorflow** [https://github.com/kozistr/rcan-tensorflow](https://github.com/kozistr/rcan-tensorflow) 38 | 39 | **PyTorch** [https://github.com/yulunzhang/RCAN](https://github.com/yulunzhang/RCAN) 40 | 41 | -------------------------------------------------------------------------------- /super resolution/RDN-CVPR2018.md: -------------------------------------------------------------------------------- 1 | # RDN # 2 | 3 | RDN-residual dense network,残差和稠密网络的结合。 4 | 5 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221192238.jpg) 6 | 7 | 如上图，比较了EDSR和SRDenseNet的block结构，文中提出了Residual dense block（RDB），是由dense和residual组成的，dense可以充分利用前面层的特征，residual可以使得学习更快更容易收敛。RDB是组成RDN网络的核心组件。如下图，RDN为C个卷积组成，每个卷积的输出都作为后面卷积的输入（dense），最后经过1* 1卷积做特征融合，并与输入相加得到此RDB层的特征输出。RDB做了局部残差特征的学习和融合。 8 | 9 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221193543.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221193543.jpg) 10 | 11 | RDN，如下图所示，由前到后共4个部分，首先是LR经过卷积提取特征F-1，然后再次经过卷积得到F0，称为特征提取层，接着是不同层次的高层特征提取，通过D个RDB组成的中间网络，这可以学习到不同层次的LR特征，然后下一部分是特征融合，使用concat连接所有RDB的特征，经过1* 1卷积做特征融合和降维，然后经过一个卷积后再加上低层特征F-1得到FDF，最后一部分就是上采样到HR目标大小的特征，然后使用最后一个卷积层来重建HR图像。 12 | 13 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221192701.jpg) 14 | 15 | RDN卷积层均使用3* 3的核（除了全局和局部特征融合之后的卷积核为1* 1），然后特征提取、局部、全局特征融合后的卷积最后是输出64个特征图。RDN里面的卷积生成的特征图个数为G。最后RDN输出的是3通道的彩色图像。 16 | 17 | 总体看，结构不是很难理解，不过很大很复杂，首先是D个RDB，每个RDB里面有C个卷积，每个卷积生成G个特征图，并且局部和全局通过dense和residual来连接。D、C、G的组合可以构造非常大和复杂的网络结果，文中的结果使用了D=16，C=8和G=64。 18 | 19 | show time 20 | 21 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181221193917.jpg) 22 | 23 | ## 开源实现 ## 24 | 25 | **论文链接** [https://arxiv.org/pdf/1802.08797.pdf](https://arxiv.org/pdf/1802.08797.pdf) 26 | 27 | **tensorflow** [https://github.com/hengchuan/RDN-TensorFlow](https://github.com/hengchuan/RDN-TensorFlow) [https://github.com/seathiefwang/RDN-Tensorflow](https://github.com/seathiefwang/RDN-Tensorflow) 28 | 29 | 30 | **PyTorch** [https://github.com/lingtengqiu/RDN-pytorch](https://github.com/lingtengqiu/RDN-pytorch) 31 | 32 | **Keras** [https://github.com/rajatkb/RDNSR-Residual-Dense-Network-for-Super-Resolution-Keras](https://github.com/rajatkb/RDNSR-Residual-Dense-Network-for-Super-Resolution-Keras) 33 | 34 | **Torch** [https://github.com/yulunzhang/RDN](https://github.com/yulunzhang/RDN) 35 | -------------------------------------------------------------------------------- /super resolution/README.md: -------------------------------------------------------------------------------- 1 | # AI for image/video SR # 2 | 3 | - [SRCNN-ECCV2014](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRCNN-ECCV2014.md) 4 | - [ESPCN-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/ESPCN-CVPR2016.md) 5 | - [FSRCNN-ECCV2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/FSRCNN-ECCV2016.md) 6 | - [VDSR-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/VDSR-CVPR2016.md) 7 | - [DRCN-CVPR2016](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/DRCN-CVPR2016.md) 8 | - [LapSRN-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/LapSRN-CVPR2017.md) 9 | - [SRDenseNet-ICCV2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRDenseNet-ICCV2017.md) 10 | - [SRGAN-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/SRGAN-CVPR2017.md) 11 | - [EDSR-CVPR2017](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/EDSR-CVPR2017.md) 12 | - [DBPN-CVPR2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/DBPN-CVPR2018.md) 13 | - [RDN-CVPR2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/RDN-CVPR2018.md) 14 | - [RCAN-ECCV2018](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/RCAN-ECCV2018.md) 15 | -------------------------------------------------------------------------------- /super resolution/SRCNN-ECCV2014.md: -------------------------------------------------------------------------------- 1 | # SRCNN # 2 | SRCNN为首个使用CNN进行端到端的单图像超分的算法。算法对输入的低分辨率图片首先做一个前处理操作：使用传统算法bicubic将低分辨率图片超分到目标高分辨率的大小，称为Y，然后通过CNN学习到的一种映射F(Y)得到最后的超分结果。映射F的目的是尽可能地使Y接近ground truth。 3 | 4 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181217172119.jpg) 5 | 6 | F的设计主要包括了三个操作： 7 | 8 | - Patch extraction and representation：特征学习，首先将Y分成很多固定大小的小块(patches)，然后通过卷积学习到特征向量，最后生成了n1个特征图，文中n1为64，卷积核大小为9*9。 9 | - Non-linear mapping:特征图重新非线性映射，将上一层的特征图映射到另一个高维特征图(mapped vector)，最后得到n2个特征图，这里采用了1*1的卷积核，n2为32。 10 | - Reconstruction:重建，将上层高维特征图重建生成一个高分辨率图像，也是通过通过卷积方式生成了c通道的图像，卷积核大小为5*5。 11 | 12 | 效果 13 | 14 | ![效果展示](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181217174219.jpg) 15 | 16 | 17 | ## 开源实现 ## 18 | 19 | **算法主页** [http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html](http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html) 20 | 21 | **tensorflow** [https://github.com/tegg89/SRCNN-Tensorflow](https://github.com/tegg89/SRCNN-Tensorflow) 22 | 23 | **caffe** [https://github.com/SolessChong/srcnn-caffe](https://github.com/SolessChong/srcnn-caffe) 24 | -------------------------------------------------------------------------------- /super resolution/SRDenseNet-ICCV2017.md: -------------------------------------------------------------------------------- 1 | # SRDenseNet # 2 | 3 | dense skip connections,本算法结合densenet和skip connections一起设计了基于CNN的非常深的超分结构。densenet是用于深层结构中缓解梯度消失问题的手段。densenet在每个block中间每一层的特征图都作为后面所有层的输入（如下图红色框里面的dense block），使得前面所有层学习到的特征都concat起来（不是resnet那种加起来），使得特征图个数变大了，这也是前面学习特征图的一种利用，加强了特征传播，减少了学习的参数量，当然，还缓解了梯度消失问题（不过确定是比较耗时和耗内存）。 4 | 5 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181219201012.jpg) 6 | 7 | 文中提出三种结构，如上图， 8 | 9 | - （a）输入LR图像经过卷积提取低层特征（low-level features），然后经过8个dense block提取高层特征（high level features），最后只对高层特征通过两个反卷积层（3* 3的核，256特征图）进行上采样和一个重建层得到HR的输出。 10 | - （b）在最后将低层特征和高层特征concat起来，同时利用来做反卷积，最后通过重建层3* 3卷积操作获得单通道的HR图像。 11 | - （c）提取高层特征的block，后面都利用上低层特征，那么所有block处理后，最后特征图数量一定很大，所以文中使用一个1* 1的卷积层bottleneck 来降低特征通道数。最后降低后的特征图256个再送入反卷积进行上采样（反卷积就是学习到一个上采样的核，可以看做卷积的逆）。 12 | 13 | 8个dense block一共有64个卷积层，最后block输出128个特征图。所以学习能力还是非常劲爆的，看图： 14 | 15 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181219202943.jpg) 16 | 17 | ## 开源实现 ## 18 | 19 | **论文链接** [http://openaccess.thecvf.com/content_ICCV_2017/papers/Tong_Image_Super-Resolution_Using_ICCV_2017_paper.pdf](http://openaccess.thecvf.com/content_ICCV_2017/papers/Tong_Image_Super-Resolution_Using_ICCV_2017_paper.pdf) 20 | 21 | **tensorflow** [https://github.com/ppooiiuuyh/SR_SRDenseNet_tensorflow](https://github.com/ppooiiuuyh/SR_SRDenseNet_tensorflow) 22 | 23 | **pytorch** [https://github.com/wxywhu/SRDenseNet-pytorch](https://github.com/wxywhu/SRDenseNet-pytorch) 24 | 25 | **caffe** [https://github.com/xueshengke/SRDenseNet-Caffe](https://github.com/xueshengke/SRDenseNet-Caffe) 26 | 27 | -------------------------------------------------------------------------------- /super resolution/SRGAN-CVPR2017.md: -------------------------------------------------------------------------------- 1 | # SRGAN # 2 | 3 | 文中提出未解决问题：: how do we recover the finer texture details 4 | when we super-resolve at large upscaling factors？ 精确度和速度在超分上都已经有很大突破，但是传统办法使用MSE作为loss函数（MSE正相关psnr，得到的psnr虽然很高），最后预测的结果容易丢失高频细节，导致比较平滑，不符合视觉感受。 5 | 6 | The ill-posed nature of the underdetermined SR problem 7 | is particularly pronounced for high upscaling factors, for 8 | which texture detail in the reconstructed SR images is 9 | typically absent. 10 | 11 | SR的主要问题，尤其在高倍上采样时，出现高频纹理细节缺失。 12 | 13 | 所以文中引入如GAN网络到图像超分上，并使用了perceptual 14 | loss感知损失（由对抗损失adversarial loss加上内容损失content loss），SRGAN也是首个X4超分可以达到逼真自然状态的框架，MOS值高哈。 15 | 16 | GAN，简单就是通过生成网络loss与对抗网络loss的设计上同时追求损失最小，那生成网络极力生成一个内容接近ground truth的图片，对抗网络极力判别出来生成的图片是假图。这样相互对抗，最后达到一个平衡，不分伯仲的状态。 17 | 18 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220111456.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220111456.jpg) 19 | 20 | 如图，SRGAN的生成网络是一个resnet（称为SRResNet），content loss使用基于VGG框架的特征图计算的loss（对像素空间的变化更稳定）。图所示，生成网络G使用了B个残差blocks，每个block包括3*3 卷积核，生成64个特征图，后接着batch-normalization layers and ParametricReLU，然后再重复一次刚才操作，接着加上input，经过B个blockS后，做上几个卷积就，最后上采样使用已经训练好的sub-pixel convolution层，然后卷积生成HR图像。 21 | 22 | 23 | 对抗网络D使用8个卷积层，每层卷积核3* 3，第一层生成特征图个数64，后面层依次*2，最后一层生成215个特征图。然后接着两个dense层和一个sigmoid激活得到最后的分类概率。 24 | 25 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165055.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165055.jpg) 26 | 27 | 如图，损失函数的设计为content loss加上adversarial loss，content loss可选，为直接计算的MSE，或是经VGG-19框架提取特征后的某一层的MSE； 28 | 29 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165322.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165322.jpg) 30 | 31 | adversarial loss为G预测后的结果输入到D对抗后的概率输出的-log（概率输出追求最大，所以加上-作为loss）。 32 | 33 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165343.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220165343.jpg) 34 | 35 | **实验部分** 36 | 37 | SRGAN使用ImageNet的随机350k张图片作为训练集。LR是通过HR做4倍的下采样获得的，每一批输入为16个96* 96的HR子patchs。 38 | 39 | 不用对抗网络的生成网络为SRResNet，也做了实验，也分别对MSE和VGG loss做了实验：SRResNet-MSE, SRResNet-VGG22, SRGAN-MSE 40 | , SRGAN-VGG22 41 | , SRGAN-VGG54，其中SRGAN-VGG54结合的mos值是最好的，也就是视觉效果最好了（虽然psnr，ssim比其他会有降低）。 42 | 43 | 对于感知loss文中做了比较，直接mse来获得loss，使用VGG22获得低层特征求loss，使用vgg54获得高层特征获得loss，趋势是高频细节可以恢复得越来越好（表现为mos值提升）。如图所示： 44 | 45 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220170523.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220170523.jpg) 46 | 47 | 48 | 直接上SRGAN效果图，细节很丰满（虽然有些细节和ground truth不一致）。x4超分能在高频细节达到这个效果真是令人震撼。 49 | 50 | ![https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220170857.jpg](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181220170857.jpg) 51 | 52 | ## 开源实现 ## 53 | 54 | **论文链接** [https://arxiv.org/pdf/1609.04802v2.pdf](https://arxiv.org/pdf/1609.04802v2.pdf) 55 | 56 | **tensorflow** [https://github.com/brade31919/SRGAN-tensorflow](https://github.com/brade31919/SRGAN-tensorflow) [https://github.com/tensorlayer/srgan](https://github.com/tensorlayer/srgan) 57 | 58 | 59 | **caffe** [https://github.com/ShenghaiRong/caffe_srgan](https://github.com/ShenghaiRong/caffe_srgan) 60 | 61 | 62 | **PyTorch** [https://github.com/aitorzip/PyTorch-SRGAN](https://github.com/aitorzip/PyTorch-SRGAN) 63 | 64 | **Keras** [https://github.com/titu1994/Super-Resolution-using-Generative-Adversarial-Networks](https://github.com/titu1994/Super-Resolution-using-Generative-Adversarial-Networks) 65 | 66 | -------------------------------------------------------------------------------- /super resolution/VDSR-CVPR2016.md: -------------------------------------------------------------------------------- 1 | # VDSR # 2 | 3 | 作者称为高精度非常深的网络，受启发于VGG-net，VDSR一共用了20层网络，通过很多层的小卷积核作用是可以搜索到前后上下文关系、更大感受野一种方法。而深层网络的缺点是梯度弥散、梯度消失、不容易收敛，文中为了解决收敛问题，使用了残差结构resnet，仅仅学习图像的residuals残差，残差是目标图与原图的差值，更容易被学习到，同时并使用很大的学习率，VDSR还提供用户自定义scale指定不同超分倍数。 4 | 5 | 6 | 利用残差好处：很多网络如SRCNN都是通过学习到的特征图来重建所有HR的细节，这样如在深层网络下the vanishing/exploding gradients problem渐变消失/爆炸问题就很严重了，文中旨在用残差解决这个问题。往往HR与LR之间很多共性，两者相减即为残差图，残差图包含更少的信息，甚至很多像素位置为所，所以更容易被学习到。 7 | 8 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218110410.jpg) 9 | 10 | 如图，VDSR很简单粗暴，首先LR通过常规插值超分到目标HR的大小，称为ILR，然后直接做d层卷积操作，其中第一层做 3x3 卷积操作，产生64个特征图，接着中间做d-2层的3x3的卷积，然后最后一层做3*3的卷积产生3通道残差图，最后与ILR相加输出HR 预测图。 11 | 12 | 这里深层网络训练过程也会更慢，如果学习率很小，如10^-5，那VDSR至少要在GPU上 训练一周，所以大学习率High Learning Rates是有助于训练上的收敛的，但是也容易引起vanishing/exploding gradients问题，所以文中建议使用大学习率时结合Adjustable Gradient Clipping（常用于递归神经网络）来遏制梯度爆炸。如下图可以看到大学习率的收敛性能。 13 | 14 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218110516.jpg) 15 | 16 | 总结下，VDSR主要特征有：网络更深性能更好、残差学习、单模型多scale多种倍数的超分。所以性能是棒棒的，如图所示： 17 | 18 | ![](https://github.com/jlygit/AI-video-enhance/blob/master/super%20resolution/image/QQ%E6%88%AA%E5%9B%BE20181218113624.jpg) 19 | 20 | ## 开源实现 ## 21 | 22 | **论文链接** [https://cv.snu.ac.kr/research/VDSR/](https://cv.snu.ac.kr/research/VDSR/) 23 | 24 | **tensorflow** [https://github.com/Jongchan/tensorflow-vdsr](https://github.com/Jongchan/tensorflow-vdsr) [https://github.com/kweisamx/TensorFlow-VDSR](https://github.com/kweisamx/TensorFlow-VDSR) 25 | 26 | **PyTorch** [https://github.com/twtygqyy/pytorch-vdsr](https://github.com/twtygqyy/pytorch-vdsr) 27 | 28 | **caffe** [https://github.com/huangzehao/caffe-vdsr](https://github.com/huangzehao/caffe-vdsr) 29 | -------------------------------------------------------------------------------- /super resolution/image/2018/tmp: -------------------------------------------------------------------------------- 1 | 2 | -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181217172119.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181217172119.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181217174219.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181217174219.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181217195042.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181217195042.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181217210829.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181217210829.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181217214229.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181217214229.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218103827.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218103827.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218110410.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218110410.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218110516.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218110516.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218113624.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218113624.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218151406.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218151406.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218152442.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218152442.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181218153657.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181218153657.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181219144544.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181219144544.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181219151026.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181219151026.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181219151924.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181219151924.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181219201012.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181219201012.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181219202943.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181219202943.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220111456.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220111456.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220165055.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220165055.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220165322.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220165322.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220165343.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220165343.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220170523.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220170523.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220170857.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220170857.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220175338.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220175338.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220190202.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220190202.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181220192743.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181220192743.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221104125.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221104125.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221104633.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221104633.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221105931.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221105931.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221111337.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221111337.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221112557.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221112557.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221192238.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221192238.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221192701.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221192701.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221192709.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221192709.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221193543.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221193543.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181221193917.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181221193917.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181224110415.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181224110415.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181224114316.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181224114316.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181224115027.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181224115027.jpg -------------------------------------------------------------------------------- /super resolution/image/QQ截图20181224144650.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/QQ截图20181224144650.jpg -------------------------------------------------------------------------------- /super resolution/image/image.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/super resolution/image/image.png -------------------------------------------------------------------------------- /testimg/face_img/shg_109922235_1047_e288cfc099f448688c2e849874d6vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_109922235_1047_e288cfc099f448688c2e849874d6vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1103787868_1047_272ffcb61d234f26b8e1d95dad75vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1103787868_1047_272ffcb61d234f26b8e1d95dad75vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1141429069_1047_a6657948b36e4e729001cd408ed0vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1141429069_1047_a6657948b36e4e729001cd408ed0vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1155406328_1047_5827fa8f4bc24f6b814e2e16d9c9vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1155406328_1047_5827fa8f4bc24f6b814e2e16d9c9vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1198794399_1047_208aeb28bc4d433baf6a925abbf7vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1198794399_1047_208aeb28bc4d433baf6a925abbf7vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1205440453_1047_cc345d90174545f6a862e89a8dd1vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1205440453_1047_cc345d90174545f6a862e89a8dd1vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1338231904_1047_87a97b248985431499b165ad3ee5vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1338231904_1047_87a97b248985431499b165ad3ee5vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1385532731_1047_8499b3cb0f714a6ead7911d78395vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1385532731_1047_8499b3cb0f714a6ead7911d78395vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1417505285_1047_7293d7ec0207475d874b349e2d48vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1417505285_1047_7293d7ec0207475d874b349e2d48vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_141884399_1047_20aa6a362e594ba1adcb4b2d847cvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_141884399_1047_20aa6a362e594ba1adcb4b2d847cvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1421560562_1047_25d19e035b144f42b7aaabd0b7a0vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1421560562_1047_25d19e035b144f42b7aaabd0b7a0vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1422167910_1047_1b347480431c41a8915f5fad2ee9vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1422167910_1047_1b347480431c41a8915f5fad2ee9vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1519789614_1047_6157c89c278844a5921cbf452446vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1519789614_1047_6157c89c278844a5921cbf452446vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1547737132_1047_8cb248ce2b8043728182067bbfa3vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1547737132_1047_8cb248ce2b8043728182067bbfa3vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1675283760_1047_c14a319eb6804c77bcb4cb65d203vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1675283760_1047_c14a319eb6804c77bcb4cb65d203vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1710966806_1047_89474bec69224104ab01077ca879vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1710966806_1047_89474bec69224104ab01077ca879vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1759379157_1047_8868e3d17dce43de8a09a911420dvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1759379157_1047_8868e3d17dce43de8a09a911420dvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_17923854_1047_2c77a6e431fe4764acbf39f6e611vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_17923854_1047_2c77a6e431fe4764acbf39f6e611vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1794846591_1047_c5535b3b8c824e07b58ced58382avide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1794846591_1047_c5535b3b8c824e07b58ced58382avide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_180587087_1047_5c45ccca5c5340fcbba7c59b2b32vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_180587087_1047_5c45ccca5c5340fcbba7c59b2b32vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1870121582_1047_ce1b321ad7894546995609f07077vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1870121582_1047_ce1b321ad7894546995609f07077vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1945479187_1047_769151977b9b434a9fd669601657vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1945479187_1047_769151977b9b434a9fd669601657vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1949720811_1047_55fcdab1af9d483288330573fcb5vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1949720811_1047_55fcdab1af9d483288330573fcb5vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_1963722372_1047_53e430b7f99a4674a28bfd9fbc66vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_1963722372_1047_53e430b7f99a4674a28bfd9fbc66vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_2019470413_1047_9880ae8ff04a48c8b2b1f5b6e798vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_2019470413_1047_9880ae8ff04a48c8b2b1f5b6e798vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_2055501841_1047_640c99d5ee5744ef90e56fd32bdcvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_2055501841_1047_640c99d5ee5744ef90e56fd32bdcvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_2074638923_1047_a9b376c564234fef9850e5ead5e4vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_2074638923_1047_a9b376c564234fef9850e5ead5e4vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_207542132_1047_3c359231920a4fa7bf113c9b8a7dvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_207542132_1047_3c359231920a4fa7bf113c9b8a7dvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_272642512_1047_a36ce564aedf4440835c4d85555dvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_272642512_1047_a36ce564aedf4440835c4d85555dvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_288036325_1047_52d7e34fce2f4994a7c6f0ec94a3vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_288036325_1047_52d7e34fce2f4994a7c6f0ec94a3vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_296658743_1047_60d77406ed98421b95543be07411vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_296658743_1047_60d77406ed98421b95543be07411vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_30425128_1047_08bf53dffc814d6b91fb717f4360vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_30425128_1047_08bf53dffc814d6b91fb717f4360vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_319326083_1047_29d9b3b71e50486291842b80dc96vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_319326083_1047_29d9b3b71e50486291842b80dc96vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_41540261_1047_19fde467661e4acf9d6fdfc52db0vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_41540261_1047_19fde467661e4acf9d6fdfc52db0vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_427881023_1047_cc14422f3cd44a6b86c6026e2e72vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_427881023_1047_cc14422f3cd44a6b86c6026e2e72vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_47075428_1047_ede999db12144950aa0d99b2d873vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_47075428_1047_ede999db12144950aa0d99b2d873vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_486841007_1047_856e831c39cb42b89bce9f70b2edvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_486841007_1047_856e831c39cb42b89bce9f70b2edvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_511368361_1047_f91aef381f8c493192533dbf3a40vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_511368361_1047_f91aef381f8c493192533dbf3a40vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_517307463_1047_d0962e3941dc4a6a9b504a7bad52vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_517307463_1047_d0962e3941dc4a6a9b504a7bad52vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_528635464_1047_3dbc904b8ad240eab7fabb2a0fc6vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_528635464_1047_3dbc904b8ad240eab7fabb2a0fc6vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_56082514_1047_3e1c1828e6f6484597c4d9f510c5vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_56082514_1047_3e1c1828e6f6484597c4d9f510c5vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_607450128_1047_e3daaf55c0604bc18bf64b9daafcvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_607450128_1047_e3daaf55c0604bc18bf64b9daafcvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_618483583_1047_ee8dd7eabd524e2da2bb4f944d94vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_618483583_1047_ee8dd7eabd524e2da2bb4f944d94vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_689821267_1047_eb4a46c128ac4b5fbe3285260a58vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_689821267_1047_eb4a46c128ac4b5fbe3285260a58vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_728975743_1047_14773a7f5bda4a798d1fff8d7b57vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_728975743_1047_14773a7f5bda4a798d1fff8d7b57vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_730324952_1047_a0bdc8fb91a9442f86108eae9b7evide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_730324952_1047_a0bdc8fb91a9442f86108eae9b7evide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_732197352_1047_552e497a8c404a0cb362b06f300fvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_732197352_1047_552e497a8c404a0cb362b06f300fvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_739699816_1047_17c7549339c84862a924b14d6899vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_739699816_1047_17c7549339c84862a924b14d6899vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_746830622_1047_69d4dbc55efe4f5f88da24dee27evide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_746830622_1047_69d4dbc55efe4f5f88da24dee27evide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_75109208_1047_8025a4abf7554f59b496744a022fvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_75109208_1047_8025a4abf7554f59b496744a022fvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_786224940_1047_8b090b3fe6764ac6bceb7c49d022vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_786224940_1047_8b090b3fe6764ac6bceb7c49d022vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_844600188_1047_bf1b247e40004553ad51b800e8a7vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_844600188_1047_bf1b247e40004553ad51b800e8a7vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_850139819_1047_219886a658b045509d37958e0ee6vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_850139819_1047_219886a658b045509d37958e0ee6vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_901052700_1047_51242db02058439e8dfefc265022vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_901052700_1047_51242db02058439e8dfefc265022vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_921792563_1047_10dafa825fa34453a5eae5bbbeeevide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_921792563_1047_10dafa825fa34453a5eae5bbbeeevide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_925784094_1047_f4f0c8a64bf94a06b690adefe572vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_925784094_1047_f4f0c8a64bf94a06b690adefe572vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_936294450_1047_986afd82d2e14887a4571b26dc7bvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_936294450_1047_986afd82d2e14887a4571b26dc7bvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_943430993_1047_36282002403349c19209684d90d8vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_943430993_1047_36282002403349c19209684d90d8vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_961532195_1047_975a8a2e5fa14544a0ae4c86565cvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_961532195_1047_975a8a2e5fa14544a0ae4c86565cvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/shg_963030462_1047_4a11175e76d8408b9b075f298587vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/shg_963030462_1047_4a11175e76d8408b9b075f298587vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_111430007_1047_18bdceb379eb44b8b088b77ab3dcvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_111430007_1047_18bdceb379eb44b8b088b77ab3dcvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1346781597_1047_b7d5f768c76d47f3824fb0db844fvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1346781597_1047_b7d5f768c76d47f3824fb0db844fvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1370035883_1047_b821a698143f413cbdf8ed3f29aavide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1370035883_1047_b821a698143f413cbdf8ed3f29aavide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1538781621_1047_d13dc2c686014960a5dc6a481e00vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1538781621_1047_d13dc2c686014960a5dc6a481e00vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1541618258_1047_71edf7553b2144eb967c633fa6f1vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1541618258_1047_71edf7553b2144eb967c633fa6f1vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1550436651_1047_7aafee7e7efc4021a39eaaaccc28vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1550436651_1047_7aafee7e7efc4021a39eaaaccc28vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_156613776_1047_fbe34f9af7614eebb97383c3c738vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_156613776_1047_fbe34f9af7614eebb97383c3c738vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1585686676_1047_9c7cba6a7776464ebda5622a8028vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1585686676_1047_9c7cba6a7776464ebda5622a8028vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1649110556_1047_a9cd725e4c814a1dbd24d27c0386vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1649110556_1047_a9cd725e4c814a1dbd24d27c0386vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1795246937_1047_997aa204dbcb4206bcd8f250cbf2vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1795246937_1047_997aa204dbcb4206bcd8f250cbf2vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1917340715_1047_a7ef639c50d94a558e379b2adb12vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1917340715_1047_a7ef639c50d94a558e379b2adb12vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_1966027031_1047_095f5ceaa0ba4b17857b85e2932evide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_1966027031_1047_095f5ceaa0ba4b17857b85e2932evide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_2031177283_1047_a8563f4f0f5b46f2a18beed263c2vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_2031177283_1047_a8563f4f0f5b46f2a18beed263c2vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_2096875702_1047_8435d9fdf7144129bc0dd8acdde4vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_2096875702_1047_8435d9fdf7144129bc0dd8acdde4vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_2109838680_1047_11c7bfc133a34cf894c1d6a487a5vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_2109838680_1047_11c7bfc133a34cf894c1d6a487a5vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_250968726_1047_b4db44fb41454f8eb146fedbbf98vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_250968726_1047_b4db44fb41454f8eb146fedbbf98vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_29808111_1047_e0106dcffcd947da91409f934868vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_29808111_1047_e0106dcffcd947da91409f934868vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_417044029_1047_bc2ea64ac35d470393fd7d4ef309vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_417044029_1047_bc2ea64ac35d470393fd7d4ef309vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_502311402_1047_c57b480d79354a31a03c8cb38b4fvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_502311402_1047_c57b480d79354a31a03c8cb38b4fvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_515353011_1047_f8b08f6220d84b0aacd8178bd532vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_515353011_1047_f8b08f6220d84b0aacd8178bd532vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_516558712_1047_eec38c5f00c64fb7a51681204ed4vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_516558712_1047_eec38c5f00c64fb7a51681204ed4vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_555070312_1047_82fd6f192f8a41a7a444e273fe56vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_555070312_1047_82fd6f192f8a41a7a444e273fe56vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_580342898_1047_712bb4acf6fd47a8b6a2f33f010cvide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_580342898_1047_712bb4acf6fd47a8b6a2f33f010cvide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_603434954_1047_c191f982a7974775abc05d71c4e1vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_603434954_1047_c191f982a7974775abc05d71c4e1vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_611863320_1047_445efb3125e64a9d815d0c3b9dd2vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_611863320_1047_445efb3125e64a9d815d0c3b9dd2vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_718668875_1047_15c4a78bebf4444ab061613bae49vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_718668875_1047_15c4a78bebf4444ab061613bae49vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_772722521_1047_82e647e527d04484b42236a41905vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_772722521_1047_82e647e527d04484b42236a41905vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_792711838_1047_a3704a779101425fb7a719f4ff31vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_792711838_1047_a3704a779101425fb7a719f4ff31vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_896054223_1047_4f27d6012de049bb9a74baf688d4vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_896054223_1047_4f27d6012de049bb9a74baf688d4vide.f0.png -------------------------------------------------------------------------------- /testimg/face_img/tjg_927249347_1047_cdd5726846054ec0aec90f0527d3vide.f0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/jlygit/AI-video-enhance/b0c8d3d1a9cebfcbec179d9ba66a15512c76af91/testimg/face_img/tjg_927249347_1047_cdd5726846054ec0aec90f0527d3vide.f0.png -------------------------------------------------------------------------------- /testimg/readme.md: -------------------------------------------------------------------------------- 1 | test for img 2 | --------------------------------------------------------------------------------