WINNet: Wavelet-Inspired Invertible Network for Image Denoising.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Image denoising seeks to recover clean images from noisy observations.
• Model-based methods offer interpretability and generalization but may lack performance.
• Learning-based methods excel in performance but often struggle with generalization and interpretability.

Purpose of the Study:

• To develop a novel image denoising approach combining the strengths of wavelet-based and learning-based methods.
• To enhance generalization ability and interpretability in image denoising.

Main Methods:

• Proposed a wavelet-inspired invertible network (WINNet) integrating lifting-inspired invertible neural networks (LINNs), sparsity-driven denoising, and noise estimation.
• LINNs learn a non-linear redundant transform inspired by wavelet lifting schemes for noise removal.
• A noise estimation network adaptively adjusts thresholds within LINNs.

Main Results:

• WINNet demonstrates high interpretability and strong generalization to unseen noise levels.
• Achieved competitive performance in non-blind/blind image denoising and image deblurring tasks.
• The network effectively utilizes a redundant multi-scale representation for denoising.

Conclusions:

• WINNet successfully merges the interpretability of model-based approaches with the performance of learning-based methods.
• The proposed architecture offers a robust solution for diverse image denoising challenges.
• WINNet shows promise for various image restoration applications beyond denoising.