Learning Retinex Prior for Compressive Hyperspectral Image Reconstruction.

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Coded aperture snapshot spectral compressive imaging (CASSI) reconstructs hyperspectral images (HSIs) from 2D measurements.
• Deep unfolding networks struggle with global illumination discrepancies and artifacts in CASSI reconstructions.
• Existing methods often fail to capture intrinsic spectral characteristics due to spatial-spectral degradation.

Purpose of the Study:

• To develop a novel deep learning network for high-fidelity HSI reconstruction from CASSI measurements.
• To address global illumination discrepancies and artifacts inherent in CASSI degradation models.
• To introduce the Retinex prior and an Adaptive Token Selection Transformer (ATST) for improved reconstruction.

Main Methods:

• Proposed a Retinex Prior-Driven Unfolding Network (RPDUN) integrating Retinex prior into a multi-stage network.
• Incorporated Retinex prior as a regularization term for global illumination adjustment.
• Introduced an Adaptive Token Selection Transformer (ATST) to mitigate noise in the reflectance domain.

Main Results:

• RPDUN effectively compensates for spatial-spectral degradation and captures intrinsic spectral characteristics.
• ATST module adaptively filters tokens, reducing noise and artifacts in the reflectance map.
• Achieved new state-of-the-art performance on simulated and real-world datasets, enhancing reconstruction quality.

Conclusions:

• RPDUN offers a significant advancement in hyperspectral image reconstruction for CASSI systems.
• The integration of Retinex prior and ATST provides superior performance and computational efficiency.
• This work represents the first application of the Retinex prior in HSI reconstruction.