DPCS-Net: A Fixed-Point Reconstruction Network Based on Compressed Sensing

Deep Unfolding Networks (DUNs) are renowned for their interpretability and exceptional performance, which have revitalized the field of Compressed Sensing (CS). However, existing DUNs often encounter issues of insufficient feature extraction and feature loss between modules during the iterative process. This paper proposes a novel deep CS framework based on the traditional Fixed-Point Continuation (FPC) optimization algorithm–Deep Compressed Sensing Network (DPCS-Net). DPCS-Net deep unfolds the optimization process of traditional CS algorithms, mapping each iterative step to a layer in the deep neural network. The network includes a Frequency Domain Feature Extraction (FDFE) module to effectively separate multi-scale features, and an Iterative Deep Unfolding (GDU+MMN) module for multi-stage feature extraction and feature fusion, thereby enhancing image reconstruction accuracy. Experimental results demonstrate that DPCS-Net outperforms existing compressed sensing image reconstruction methods on several standard datasets, proving its excellent reconstruction capability and robustness.