Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DeCompress: Denoising via Neural Compression

Ali Zafari, Xi Chen, Shirin Jalali

TL;DR

DeCompress addresses denoising when ground-truth clean images are unavailable and large paired datasets are impractical. It formulates denoising as learning a neural compression code by minimizing a rate-distortion objective $\mathbb{E}\left[ \frac{1}{n}\|\hat{Y}^n - Y^n\|_2^2 \right] + \lambda R$, where $R = \mathbb{E}[-\log_2 P(\lfloor C^m \rceil)]$ and $\hat{Y}^n = g_s(\lfloor C^m \rceil)$ with bottleneck $C^m = g_a(Y^n)$. Training uses uniform-noise augmentation to approximate quantization and an entropy model to estimate $R$, enabling learning from noisy data only. The results show DeCompress achieving competitive denoising performance against zero-shot baselines while requiring only a single noisy image or limited data, highlighting a practical compression-based paradigm for denoising in data-scarce settings. This approach potentially extends to other inverse problems by coupling neural compression with adaptive rate–distortion control.

Abstract

Learning-based denoising algorithms achieve state-of-the-art performance across various denoising tasks. However, training such models relies on access to large training datasets consisting of clean and noisy image pairs. On the other hand, in many imaging applications, such as microscopy, collecting ground truth images is often infeasible. To address this challenge, researchers have recently developed algorithms that can be trained without requiring access to ground truth data. However, training such models remains computationally challenging and still requires access to large noisy training samples. In this work, inspired by compression-based denoising and recent advances in neural compression, we propose a new compression-based denoising algorithm, which we name DeCompress, that i) does not require access to ground truth images, ii) does not require access to large training dataset - only a single noisy image is sufficient, iii) is robust to overfitting, and iv) achieves superior performance compared with zero-shot or unsupervised learning-based denoisers.

DeCompress: Denoising via Neural Compression

TL;DR

DeCompress addresses denoising when ground-truth clean images are unavailable and large paired datasets are impractical. It formulates denoising as learning a neural compression code by minimizing a rate-distortion objective , where and with bottleneck . Training uses uniform-noise augmentation to approximate quantization and an entropy model to estimate , enabling learning from noisy data only. The results show DeCompress achieving competitive denoising performance against zero-shot baselines while requiring only a single noisy image or limited data, highlighting a practical compression-based paradigm for denoising in data-scarce settings. This approach potentially extends to other inverse problems by coupling neural compression with adaptive rate–distortion control.

Abstract

Learning-based denoising algorithms achieve state-of-the-art performance across various denoising tasks. However, training such models relies on access to large training datasets consisting of clean and noisy image pairs. On the other hand, in many imaging applications, such as microscopy, collecting ground truth images is often infeasible. To address this challenge, researchers have recently developed algorithms that can be trained without requiring access to ground truth data. However, training such models remains computationally challenging and still requires access to large noisy training samples. In this work, inspired by compression-based denoising and recent advances in neural compression, we propose a new compression-based denoising algorithm, which we name DeCompress, that i) does not require access to ground truth images, ii) does not require access to large training dataset - only a single noisy image is sufficient, iii) is robust to overfitting, and iv) achieves superior performance compared with zero-shot or unsupervised learning-based denoisers.

Paper Structure

This paper contains 13 sections, 3 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Lossy Neural Codes for Denoising
  4. Compression-based denoising
  5. Neural Compression as Denoiser
  6. Experiments
  7. Training on one vs. many noisy images
  8. Relationship between $\lambda$ and $\sigma$?
  9. Experimental settings
  10. Neural compression denoisers
  11. Data preparation
  12. Baseline settings
  13. Conclusion

Figures (4)

  • Figure 1: Structure of the neural compression network used for denoising. Conv represents convolutional layer, and GDN is the activation function. More details on the structure of network can be found in Section \ref{['sec:exp:settings']}.
  • Figure 2: Eleven test images, with resolutions of $256 \times 256$ for images 1-7 and $512 \times 512$ for image 8-11.
  • Figure 3: The image randomly selected from ImageNet-1k used for training our neural compression (ILSVRC2012_val_00000059.jpeg, $500\times375$ pixels).
  • Figure 4: Visual comparison of denoising methods in Table \ref{['tab:main_results']} for Monarch image at noise level $\sigma=25$. The PSNR (dB) is above each denoised image.