Multi-Stage Adaptive Sampling Coupled With Residual Dense-Transformer for Image Compressive-Sensing Reconstruction

As an efficient signal acquisition and reconstruction technology, compressed sensing (CS) has key application value in single-pixel imaging, magnetic resonance imaging (MRI) and other fields. The core challenge is to achieve high image reconstruction through a small amount of measurement data while solving the problem of low fixed sampling efficiency. It is difficult for a single network to balance local details and global dependence. In this paper, Multi-Stage Adaptive Sampling Coupled with Residual Dense-Transformer (MASRDT-Net) is proposed, which combines a multi-stage adaptive sampling module and a hybrid reconstruction architecture. Its core design includes: (1) A multi-stage adaptive sampling module, which captures the image infrastructure at the initial sampling rate and distributes the remaining sampling rate to the iteration stage in proportion. Combined with a trainable floating-point perception matrix, it realizes "priority sampling in important areas" and improves measurement data utilization at low sampling rates. (2) The residual dense-Transformer hybrid reconstruction module strengthens local feature extraction through residual dense blocks (RDB), Swin Transformer controls complexity and captures global dependencies, and corrects errors with the "PreReconBlock-Transformer-PostReconBlock" iterative process. Experiments show that the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of MASRDT-Net are better than ReconNet, ISTA-Net, and other methods at different sampling rates, realizing the balance of "high-efficiency sampling-high-precision reconstruction", providing practical solutions for practical CS applications.