Snapshot Compressive Imaging via Degradation Cue and Spectral Latent Diffusion

Area of Science:

• Computational imaging
• Signal processing
• Hyperspectral imaging

Background:

• Snapshot spectral compressive imaging (SSCI) aims to reconstruct 3D hyperspectral images from 2D measurements.
• Current methods struggle to fully utilize degradation information and image priors, leading to inaccurate prior modeling and performance gaps, especially in complex scenarios.
• High compression in measurements like those from a coded aperture snapshot spectral imager (CASSI) causes loss of spectral-spatial context, which existing priors fail to capture effectively.

Purpose of the Study:

• To develop a novel reconstruction method that addresses the limitations of existing approaches in leveraging degradation and image priors for hyperspectral imaging.
• To improve the accuracy of hyperspectral image reconstruction by effectively capturing spectral-spatial context and enhancing prior information.
• To introduce a method that overcomes the performance gap observed in complex scenarios due to inaccurate prior modeling and context loss.

Main Methods:

• The proposed method, Degradation Cue Learning and Spectral Latent Diffusion (DCL-SLD), integrates two key modules: Degradation Cue Learning (DCL) and Spectral Latent Diffusion (SLD).
• The DCL module utilizes a pre-trained image encoder and feature distribution transmission to extract and integrate degraded spatial information, enabling reconstruction via learned visual context.
• The SLD module employs a latent diffusion model based on spectral correlations to generate a low-rank vector representation, preserving contextual relationships within the high-dimensional spectral structure.

Main Results:

• The DCL-SLD method demonstrated superior performance compared to state-of-the-art methods in hyperspectral image reconstruction.
• Extensive experiments on both simulated and real datasets validated the effectiveness of the proposed approach.
• The method significantly improves the exploitation of contextual information in both spatial and spectral domains, leading to more accurate recovery.

Conclusions:

• The DCL-SLD method offers a significant advancement in hyperspectral image reconstruction by effectively learning degradation cues and utilizing spectral latent diffusion.
• The integrated approach enhances the utilization of prior information and spectral-spatial context, overcoming limitations of previous methods.
• The proposed technique shows strong potential for applications requiring accurate 3D hyperspectral image recovery from compressed 2D measurements.