Enhanced deep unrolling networks for snapshot compressive hyperspectral imaging

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Snapshot compressive hyperspectral imaging (SCHI) presents a significant inverse problem for reconstructing full hyperspectral images from limited measurements.
• Existing methods often struggle with reconstruction accuracy and convergence speed.

Purpose of the Study:

• To propose an enhanced deep unrolling neural network (EDUNet) for accurate hyperspectral image reconstruction in SCHI.
• To improve the convergence and performance of hyperspectral image reconstruction algorithms.

Main Methods:

• EDUNet is developed by deep unrolling a proximal gradient descent algorithm, incorporating novel gradient-driven update and proximal mapping modules.
• The gradient-driven update module uses a memory-assisted descent for enhanced convergence.
• The proximal mapping module features cross-stage spectral self-attention and a spectral geometry consistency loss for improved spectral information capture.

Main Results:

• Experiments on benchmark datasets (KAIST, ICVL, Harvard) and real data show EDUNet outperforms 15 competing models.
• EDUNet achieved superior performance across PSNR, SSIM, SAM, and ERGAS metrics.
• The proposed modules effectively exploit spectral self-similarities and geometric layouts.

Conclusions:

• EDUNet offers a robust and effective solution for hyperspectral image reconstruction in SCHI.
• The novel architectural components and loss function contribute to significant performance gains.
• This work advances the state-of-the-art in compressive hyperspectral imaging reconstruction.