Representation Sampling and Hybrid Transformer Network for Image Compressed Sensing

Abstract

Deep unrolling networks (DUNs) have attracted substantial attention in the field of image compressed sensing (CS) due to their superior performance and good interpretability by recasting optimization algorithms as deep networks. However, existing DUNs suffer from low sampling efficiency, and the improvement in reconstruction quality heavily relies on large model complexity. To address these issues, we propose a lightweight Representation Sampling and Hybrid Transformer Network (RHT-Net). Firstly, we propose a Representation-CS (RCS) model to extract high-level features to achieve efficient sampling. This sampling strategy leads to highly dense, semantically rich and extremely compact features without observing the original pixels, which also reduces the cross-domain loss during iteration. Secondly, we design a Tri-Scale Sparse Denoising (TSSD) module in the deep unrolling stages to extend sparse proximal projections, leveraging multi-scale auxiliary variables to enhance multi-feature flow and memory effects. Thirdly, we develop a hybrid Transformer module that includes a Global Cross Attention (GCA) block and a Window Local Attention (WLA) block, using the measurements to cross-estimate the reconstruction error, thereby generating finer spatial details and improving local recovery. Experiments demonstrate that RHT-Net enhanced version outperforms the current state-of-the-art methods by up to 1.17dB in PSNR. The lightweight RHT-Net achieves a 0.43dB gain while reducing model parameters by up to 22 times. The code will be released publicly at https://github.com/songhp/RHTNet.