Hybrid Multi-Dimensional Attention U-Net for Hyperspectral Snapshot Compressive Imaging Reconstruction

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Spectral snapshot compressive imaging (SCI) captures 3D spatial-spectral data in a single 2D measurement.
• Reconstructing hyperspectral images from compressed 2D measurements is an ill-posed problem.
• Existing methods using 2D convolutions and basic attention struggle to extract spectral features effectively.

Purpose of the Study:

• To develop an efficient and high-quality hyperspectral image reconstruction method.
• To address the limitations of existing SCI reconstruction techniques.
• To balance reconstruction performance with computational cost.

Main Methods:

• Proposed a hybrid multi-dimensional attention U-Net (HMDAU-Net) for end-to-end hyperspectral image reconstruction.
• Integrated 3D and 2D convolutions within an encoder-decoder architecture.
• Employed attention gates to enhance salient features and reduce noise in skip connections.

Main Results:

• HMDAU-Net effectively utilizes spectral information for reconstruction.
• The hybrid approach balances computational cost and reconstruction quality.
• Achieved superior performance compared to state-of-the-art reconstruction algorithms.

Conclusions:

• HMDAU-Net offers an effective solution for hyperspectral image reconstruction from compressive measurements.
• The proposed architecture demonstrates significant improvements in image quality and efficiency.
• This method advances the field of spectral snapshot compressive imaging.