Mask-Guided Spatial-Spectral MLP Network for High-Resolution Hyperspectral Image Reconstruction

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Hyperspectral image (HSI) reconstruction is crucial for spectral compressive imaging (CASSI) systems, especially in dynamic environments.
• Deep unfolding frameworks have advanced HSI reconstruction but suffer from large models and high computational costs.
• Existing methods iteratively solve sub-problems, increasing model complexity and resource demands.

Purpose of the Study:

• To develop a computationally efficient and accurate HSI reconstruction method for CASSI systems.
• To address the limitations of deep unfolding frameworks in terms of model size and overhead.
• To disentangle degradation and latent target representations using a novel deep learning approach.

Main Methods:

• A lightweight MLP block captures non-local similarities and long-range dependencies in spatial and spectral domains.
• An attention-based mask modeling module generates spatial-spectral-adaptive degradation representations.
• Multi-level fusion and deeper supervision enhance information flow and feature extraction.
• A dual-domain loss combines projection and reconstruction losses for consistent optical detection.

Main Results:

• The proposed method achieves superior reconstruction accuracy compared to state-of-the-art approaches on benchmark HSI datasets.
• Significant reductions in computational and memory costs are demonstrated.
• The method effectively disentangles degradation information from latent target representations.
• Experimental results validate the efficiency and effectiveness of the proposed technique.

Conclusions:

• The presented method offers a simple yet effective solution for HSI reconstruction in CASSI.
• It overcomes the computational and memory limitations of traditional deep unfolding methods.
• The approach demonstrates strong performance in both accuracy and efficiency, paving the way for practical applications.