Pixel adaptive deep-unfolding neural network with state space model for image deraining.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Rain streaks degrade image quality and hinder computer vision tasks.
• Deep unfolding neural networks (DUNs) show promise for image deraining but have limitations.
• Existing DUNs struggle with global structure perception and adaptive step size control.

Purpose of the Study:

• To develop an advanced image deraining method addressing limitations of current deep unfolding networks.
• To enhance the perception of both local and global image structures.
• To improve the adaptability of deraining methods to diverse input images.

Main Methods:

• Proposes a pixel adaptive deep unfolding network incorporating state space models (SSMs).
• Introduces an adaptive pixel-wise gradient descent (APGD) module for flexible step size adjustment.
• Employs a stage fusion proximal mapping (SFPM) module with a dual-branch architecture (CNNs and SSMs).
• Utilizes Fourier transform for stage feature fusion to minimize information loss.

Main Results:

• The proposed method demonstrates superior performance in quantitative metrics and visual quality on public datasets.
• Achieves effective deraining by enhancing global structure perception and adaptive gradient descent.
• State space models provide efficient long-range dependency modeling with linear complexity.

Conclusions:

• The developed pixel adaptive deep unfolding network offers a significant advancement in image deraining.
• The integration of APGD and SFPM modules effectively overcomes previous limitations in DUNs.
• The method achieves state-of-the-art results, offering improved visual quality for rainy images.