Degradation-learning spatial-sparsity transformation unfolding network for reflective spectral compressive imaging

Snapshot compressive spectral imaging has emerged as an efficient imaging paradigm that captures hyperspectral images (HSIs) through a single snapshot while maintaining high reconstruction quality. As two fundamental components of CASSI systems, the optical architecture and reconstruction algorithm jointly determine imaging performance. Reflective CASSI (R-CASSI) presents a promising advancement by integrating the structural compactness of single-disperser (SD) CASSI with the superior spatial resolution of dual-disperser (DD) CASSI, demonstrating substantial development potential. While deep learning-based reconstruction approaches, particularly deep unfolding algorithms, have become predominant in this field, current research on R-CASSI-specific algorithms remains preliminary, with most existing solutions primarily designed for SD-CASSI architectures. This paper proposes a Degradation-Learning Spatial-Sparsity Transformation Unfolding Network (DLSSTUN) for hyperspectral imaging reconstruction in R-CASSI systems. Spatial Sparsity Transformation (SST) module based on Swin Transformers is proposed to recover spatial-spectral correlations. The Dynamic Gradient Descent (DGD) mechanism is introduced to reconcile discrepancies between the sensing matrix and degradation matrix. Comprehensive experiments demonstrate that our DLSSTUN framework achieves state-of-the-art performance.