A Refinement Boosted and Attention Guided Deep FISTA Reconstruction Framework for Compressive Spectral Imaging

Hyperspectral images (HSIs) contain rich spatial and spectral information. A double dispersers coded aperture snapshot spectral imaging (DD-CASSI) system takes advantage of compressive sensing (CS) theory to map 3-D HSI data into a single 2-D measurement. One of the key components of DD-CASSI is to reconstruct high-quality HSI from measurement. Traditional model-based methods use mathematical optimization to reconstruct HSIs according to prior knowledge. Current deep learning-based methods achieve pleasant results. However, fully learned deep learning methods lack interpretability, and model-based deep learning methods cannot achieve pleasant performance. In this article, we propose a novel HSI reconstruction framework na med refinement boosted and attention guided tensor fast iterative shrinkage-thresholding algorithm-Net (ReAttFISTA-Net), which combines model-based deep learning and fully learned deep learning reconstruction strategies. In this framework, we introduce an attention guided fusion mechanism, which enhances spatial–spectral information, refinement subnetwork, and auxiliary loss terms to improve the reconstruction performance. Extensive experimental results show that the proposed reconstruction algorithm outperforms the state-of-the-art algorithms on both simulation and real-world datasets.