Variational Bayesian Inference-Inspired Network for Compressive Spectral Imaging

In Coded Aperture Snapshot Spectral Imaging (CASSI) systems, model-based approaches highly rely on the handcrafted priors, while data-driven methods overlook the physical degradation process of measurement. To bridge the gap between model-based and data-driven methods, in this paper, we present a Variational Bayesian Inference-Inspired Network (VBII-Net) for solving the compressive spectral image reconstruction, that is, HSI reconstruction and sampling processing are integrated into a Bayesian framework. Specifically, the reconstruction problem is converted to the likelihood term in the evidence lower bound (ELBO), which supervises the learning of the inference network. Consequently, the inference network for the original HSIs and the coded mask can be jointly learned under the guidance of deep priors, enabling them to interact with each other based on the likelihood. Additionally, a nested learned mask-feature reaggregation module (NLMR) is designed to guide the coded feature regression at each stage of deep feature extraction by reusing learned masks, further enhancing the framework’s efficiency. By reweighting only the high-fidelity regions, NLMR encourages the model to focus computation and learning capacity on informative areas, reducing unnecessary updates during training and inference. Therefore, it significantly reduces FLOPs and parameter count.Finally, the proposed method significantly outperformed state-of-the-art approaches on both simulated and real datasets, achieving an average of 0.06dB improvement over the best counterpart method, while decreasing FLOPs by up to 26%. Code will be released.