SALSA-Net: Explainable Deep Unrolling Networks for Compressed Sensing.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Compressed sensing (CS) faces challenges in efficiency and accuracy.
• Deep unrolling networks (DUNs) offer explainability, speed, and performance for CS.
• Classical deep networks have limitations in CS problem-solving.

Purpose of the Study:

• To propose SALSA-Net, a novel deep unrolling model for image compressed sensing.
• To improve the efficiency and accuracy of CS reconstruction.
• To leverage the strengths of both classical algorithms and deep learning.

Main Methods:

• Developed SALSA-Net by unrolling and truncating the Split Augmented Lagrangian Shrinkage Algorithm (SALSA).
• Implemented a network structure with gradient update, threshold denoising, and auxiliary update modules.
• Introduced learned sampling to replace traditional methods for better feature preservation and efficiency.
• Optimized all parameters via end-to-end learning with forward constraints for faster convergence.

Main Results:

• SALSA-Net demonstrated significant reconstruction performance improvements over state-of-the-art methods.
• The model inherited the explainable recovery and high-speed advantages of DUNs.
• Learned sampling enhanced feature information preservation and sampling efficiency.

Conclusions:

• SALSA-Net offers a promising approach to address the efficiency and accuracy challenges in image CS.
• The proposed model effectively combines the interpretability of SALSA with the learning capabilities of deep neural networks.
• SALSA-Net represents a significant advancement in deep unrolling networks for compressed sensing applications.