COAST: COntrollable Arbitrary-Sampling NeTwork for Compressive Sensing.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Deep network-based compressive sensing (CS) methods often require separate models for each sampling matrix, leading to inefficiency and poor generalization.
• Existing CS models struggle with arbitrary or unseen sampling matrices, limiting their practical applicability.

Purpose of the Study:

• To develop a single deep learning model capable of solving CS problems for arbitrary sampling matrices, including previously unseen ones.
• To enhance the efficiency and generalization ability of deep network-based CS methods.

Main Methods:

• Propose a novel COntrollable Arbitrary-Sampling neTwork (COAST) based on an optimization-inspired deep unfolding framework.
• Introduce a random projection augmentation (RPA) strategy for diverse training across the sampling space.
• Develop a controllable proximal mapping module (CPMM) and a plug-and-play deblocking (PnP-D) strategy for dynamic feature modulation and artifact elimination.

Main Results:

• The COAST model demonstrates the ability to handle arbitrary sampling matrices using a single network.
• Achieves state-of-the-art performance in CS reconstruction with significantly reduced computational speed.
• Experimental results validate the model's effectiveness on widely used benchmark datasets.

Conclusions:

• The proposed COAST network offers an efficient and generalized solution for compressive sensing with arbitrary sampling matrices.
• COAST overcomes the limitations of traditional CS methods by enabling a single model to adapt to diverse sampling strategies.
• The developed strategies like RPA, CPMM, and PnP-D contribute to the model's robust performance and interpretability.