USB-Net: Unfolding Split Bregman Method With Multi-Phase Feature Integration for Compressive Imaging.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Existing unfolding-based compressive imaging methods face challenges with inefficient feature extraction and information loss, particularly at low sampling ratios, leading to degraded image quality.
• These limitations manifest as significant detail degradation and distortion in reconstructed images, hindering practical applications.

Purpose of the Study:

• To introduce USB-Net, a novel deep unfolding network designed to overcome the limitations of existing compressive imaging reconstruction techniques.
• To enhance feature extraction, optimize data fidelity, and achieve high-quality image reconstruction, especially under challenging low sampling conditions.

Main Methods:

• Developed USB-Net, a deep unfolding method inspired by the Split Bregman algorithm and incorporating a multi-phase feature integration strategy.
• Utilized a customized Depthwise Attention Block for efficient feature extraction and to address the Split Bregman method's splitting operator.
• Introduced Auxiliary Iteration Modules (X(k), D(k), B(k)) and Iterative Fusion Modules to improve the decomposition strategy and integrate multi-phase iterative insights.

Main Results:

• USB-Net demonstrated significant improvements in image reconstruction quality compared to state-of-the-art methods.
• The method achieved superior performance on various compressive imaging tasks, including general compressive sensing, CS-magnetic resonance imaging, and snapshot compressive imaging.
• Experiments confirmed the network's effectiveness in mitigating detail degradation and distortion, even at low sampling ratios.

Conclusions:

• USB-Net effectively addresses the challenges of feature extraction and information loss in compressive imaging reconstruction.
• The proposed method leverages multi-phase iterative insights to enhance data fidelity and produce high-quality reconstructed images.
• USB-Net exhibits strong generalizability and outperforms existing approaches across diverse imaging applications.