SA-CCSNet: Saliency-Aware Cascade Network for Image Compressed Sensing

Sampling matrix design and reconstruction scheme development are two critical issues in image compressed sensing (CS). For the first issue, uniform sampling matrices are commonly used, which ignore the characteristics of the image and lead to distortion in important image parts. For the second problem, traditional iterative optimization algorithms often suffer from high computational complexity, while deep learning-based non-iterative methods need more insight from the CS domain. To solve the two problems, we propose a saliency-aware cascade framework that includes three serial subnetworks, i.e., sampling subnetwork, initial recovery subnetwork, and deep recovery subnetwork. There are two branches, full image sampling and salient image sampling, in the sampling subnetwork to allocate more sampling resources for salient regions. Initial recovery is used to obtain a preliminary recovered image, and deep recovery subnetwork maps the traditional Iterative Shrinkage-Thresholding Algorithm (ISTA) to CNN convolutional modules for further optimization. The entire network is learned from data end-to-end using a hybrid loss. Experimental results show that our approach results in superior quality reconstruction images and higher metrics compared to alternative algorithms, particularly at extremely low sampling rates ranging from 0.01 to 0.25.