├── .DS_Store
├── Fig
    ├── Fig2.png
    └── Tab2.png
└── README.md


/.DS_Store:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TIDESlab/ITS/HEAD/.DS_Store


--------------------------------------------------------------------------------
/Fig/Fig2.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TIDESlab/ITS/HEAD/Fig/Fig2.png


--------------------------------------------------------------------------------
/Fig/Tab2.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/TIDESlab/ITS/HEAD/Fig/Tab2.png


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
  1 | # Boosting of Implicit Neural Representation-based Image Denoiser
  2 | 
  3 | [Zipei Yan](https://yanzipei.github.io/), [Zhengji Liu](https://epicwatermelon.github.io/), [Jizhou Li](http://jizhou.li/)
  4 | 
  5 | This paper is available on [arXiv](https://arxiv.org/abs/2401.01548) and [IEEEXplore](https://ieeexplore.ieee.org/document/10447327).
  6 | 
  7 | ## Abstract
  8 | Implicit Neural Representation (INR) has emerged as an effective
  9 | method for unsupervised image denoising. However, INR models
 10 | are typically overparameterized; consequently, these models are
 11 | prone to overfitting during learning, resulting in suboptimal results,
 12 | even noisy ones. To tackle this problem, we propose a general recipe
 13 | for regularizing INR models in image denoising. In detail, we propose
 14 | to iteratively substitute the supervision signal with the mean
 15 | value derived from both the prediction and supervision signal during
 16 | the learning process. We theoretically prove that such a simple iterative
 17 | substitute can gradually enhance the signal-to-noise ratio of the
 18 | supervision signal, thereby benefiting INR models during the learning
 19 | process. Our experimental results demonstrate that INR models
 20 | can be effectively regularized by the proposed approach, relieving
 21 | overfitting and boosting image denoising performance.
 22 | 
 23 | ## How to use ITS for boosting INR-based image denoiser?
 24 | 
 25 | Since ITS is INR model agnostic/independent, you can plug ITS into your own INR-based denoiser.
 26 | 
 27 | ### Iterative Substitution (ITS) for Renewing the Supervision Signal
 28 | 
 29 | For a given set of iterations $`\{N, 2N, ..., kN\}`$ where $N$ is a hyper-parameter, ITS renews the supervision signals with the mean value derived from both the prediction and supervision signal, which is formulated as follows:
 30 | 
 31 | $$
 32 | \boldsymbol{\hat{y}}^{kN+1} = \frac{\boldsymbol{y} + \boldsymbol{\hat{x}}^{kN}}{2},
 33 | $$
 34 | 
 35 | where $\boldsymbol{\hat{y}}^{kN+1}$ is the renewed supervision signal, $\boldsymbol{y}$ is the original noisy observation and $\boldsymbol{\hat{x}}^{kN}$ is the denoised results from INR model.
 36 | 
 37 | ### Implementation
 38 | ```python
 39 | import torch
 40 | 
 41 | 
 42 | """
 43 | num_iters: number of training iterations
 44 | inr_model: INR model
 45 | z: coordinates
 46 | y: noisy observation
 47 | N: every N-th iteration for renewing the supervision signal
 48 | """
 49 | 
 50 | y_hat = torch.clone(y)
 51 | 
 52 | for i in range(1, num_iters + 1):
 53 |     x_hat = inr_model(z)
 54 |     
 55 |     loss = ((x_hat - y_hat) ** 2).mean()
 56 |     
 57 |     optimizer.zero_grad()
 58 |     loss.backward()
 59 |     optimizer.step()
 60 |     
 61 |     # renew supervision signal at every N-th iteration
 62 |     if i % N == 0: 
 63 |         y_hat = (y + x_hat.detach()) / 2.
 64 | ```
 65 | 
 66 | ## Results
 67 | 
 68 | ![image](./Fig/Fig2.png)
 69 | 
 70 | ![image](./Fig/Tab2.png)
 71 | 
 72 | More results could be found in our paper.
 73 | 
 74 | ## Repos for INR-based image denoiser
 75 | 
 76 | DIP: https://github.com/DmitryUlyanov/deep-image-prior
 77 | 
 78 | SIREN: https://github.com/LeoZDong/siren_denoise
 79 | 
 80 | LINR: https://github.com/WenTXuL/LINR/tree/main
 81 | 
 82 | WIRE: https://github.com/vishwa91/wire/tree/main
 83 | 
 84 | ADMM-DIPTV: https://github.com/sedaboni/ADMM-DIPTV
 85 | 
 86 | DeepRED: https://github.com/GaryMataev/DeepRED
 87 | 
 88 | 
 89 | ## Citation
 90 | If this work is useful for your research, please kindly cite it:
 91 | ```
 92 | @inproceedings{yan2024its,
 93 | title={Boosting of Implicit Neural Representation-based Image Denoiser},
 94 | author={Yan, Zipei and Liu, Zhengji and Li, Jizhou},
 95 | booktitle={ICASSP},
 96 | year={2024},
 97 | }
 98 | ```
 99 | 
100 | ## Contact
101 | 
102 | Please contact: lijz AT ieee DOT org
103 | 


--------------------------------------------------------------------------------