Equilibrium Image Denoising With Implicit Differentiation.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Learning-based image denoising often uses unrolled architectures with fixed, stacked blocks.
• Training deeper networks and selecting the optimal number of blocks pose significant challenges, potentially degrading performance.

Purpose of the Study:

• To propose a novel implicit model for iterative image denoising, addressing limitations of explicit unrolled networks.
• To develop a parameter-efficient model that avoids training difficulties and manual tuning of iteration numbers.

Main Methods:

• The proposed model utilizes implicit differentiation for gradient calculation, simplifying the backward pass.
• A single implicit layer, formulated as a fixed-point equation, solves for noise features.
• Infinite iterations are simulated using accelerated black-box solvers to achieve an equilibrium for denoising.

Main Results:

• The implicit layer effectively captures non-local self-similarity priors crucial for image denoising.
• The model demonstrates enhanced training stability and improved denoising performance compared to explicit methods.
• Extensive experiments confirm superior qualitative and quantitative results over state-of-the-art explicit denoisers.

Conclusions:

• The implicit modeling approach offers a more stable and efficient alternative for iterative image denoising.
• This method achieves state-of-the-art performance by leveraging implicit layers and avoiding common training pitfalls.
• The proposed technique represents a significant advancement in learning-based image restoration.