Hyperspectral image reconstruction via patch attention driven network.

Area of Science:

• Computer Vision
• Image Processing
• Spectroscopy

Background:

• Coded aperture snapshot spectral imaging (CASSI) enables 3D hyperspectral image (HSI) acquisition using 2D compressive measurements.
• Reconstructing HSIs from these measurements is an ill-posed inverse problem, posing significant challenges in the field.

Purpose of the Study:

• To propose a novel network architecture for accurate HSI recovery from CASSI measurements.
• To enhance the performance of HSI reconstruction by addressing the ill-posed nature of the inverse problem.

Main Methods:

• A multilevel residual network driven by a patch-wise attention mechanism was developed.
• A complementary input method was introduced for data pre-processing to integrate measurements and coded aperture information.
• The patch attention module adaptively generates clues by analyzing feature distribution and global correlations.

Main Results:

• The proposed network architecture demonstrated superior performance compared to existing state-of-the-art methods.
• Extensive simulation experiments validated the effectiveness of the novel approach.
• The method successfully improved the accuracy of hyperspectral image reconstruction.

Conclusions:

• The developed network architecture offers a significant advancement in HSI recovery from CASSI data.
• The combination of patch attention and complementary input pre-processing effectively tackles the challenges of ill-posed inverse problems.
• This work provides a robust solution for enhancing hyperspectral imaging applications.