Hyperspectral image reconstruction via local-nonlocal priors and spectral smoothness regularization

Compressive hyperspectral imaging utilizes a two-dimensional detector to capture the three-dimensional spectral data cube in a compressed manner, enabling snapshot spectral imaging. Recovering the original hyperspectral image (HSI) from the compressed measurement is an ill-posed problem, which is typically addressed by introducing regularization terms or by using deep learning approaches. Recently, methods that combine both strategies have attracted significant research attention. However, these approaches usually simply embed neural network as a feature extractor into traditional optimization frameworks, or directly integrate regularization terms with neural networks in a sequential processing pipeline. As a result, they fail to construct unified representations that faithfully reflect the low-rank property, spectral smoothness, and the inherent spatial-spectral relationship of HSI. In this work, we proposed a multi-prior fusion framework for HSI reconstruction. The proposed approach utilizes deep image prior (DIP), enabling the extraction of local structures in HSI directly from the compressed measurement by employing an untrained U-Net. Meanwhile, nonlocal self-similarity (NSS) prior is incorporated to further enforce global correlation, and total variation (TV) regularization ensures spectral smoothness, thereby achieving unified optimization across local-nonlocal and spatial-spectral domains. Compared with existing methods, the proposed approach achieves superior reconstruction results on synthetic datasets. In addition, we built an experimental optical setup and performed ablation experiments, which further validated the effectiveness of the proposed method.