Gates-Controlled Deep Unfolding Network for Image Compressed Sensing
Tiancheng Li, Qiurong Yan, Quan Zou, Qianling Dai
Abstract
Deep Unfolding Networks (DUNs) have demonstrated remarkable success in compressed sensing by integrating optimization solvers with deep neural networks. The issue of information loss during the unfolding process has received significant attention. To address this issue, many advanced deep unfolding networks utilize memory mechanisms to augment the information transmission during iterations. However, most of these networks only use the memory module to enhance the proximal mapping process instead of adjusting the entire iteration. In this paper, we propose an LSTM-inspired proximal gradient descent module called the Gates-Controlled Iterative Module(GCIM), leading to a Gates-Controlled Deep Unfolding Network(GCDUN) for compressed sensing. We utilize the gate units to modulate the information flow through the iteration by forgetting the redundant information before the gradient descent, providing necessary features for the proximal mapping stage, and selecting the key information for the next stage. To reduce parameters, we propose a parameter-friendly version called Recurrent Gates-Controlled Deep Unfolding Networks (RGCDUN), which also achieves great performance but with much fewer parameters. Extensive experiments manifest that our networks achieve excellent performance. The source codes are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/coder0856/GCDUN/</uri> .