Snapshot hyperspectral imaging method based on a transformer and auxiliary learning tasks

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Hyperspectral imaging (HSI) captures detailed spectral information but faces reconstruction challenges.
• Coded aperture snapshot spectral imaging (CASSI) offers a single-shot acquisition method.
• Noise in measurements significantly degrades the quality of reconstructed hyperspectral images.

Purpose of the Study:

• To develop a noise-resistant algorithm for CASSI reconstruction.
• To improve the spatial and spectral fidelity of reconstructed hyperspectral images.
• To enhance the regularization capability of CASSI reconstruction algorithms.

Main Methods:

• A spectral awareness network (SANet) was proposed for CASSI reconstruction.
• An auxiliary learning task using panchromatic (PAN) image reconstruction was incorporated.
• Spatial details from the reconstructed PAN image were used to improve the CASSI reconstruction network.

Main Results:

• The SANet algorithm demonstrated superior performance compared to state-of-the-art methods.
• Reconstructed hyperspectral images showed higher structural similarity index (SSIM) and spectral angle mapper (SAM) values.
• The method proved robust against added Gaussian and Poisson noise.

Conclusions:

• The proposed SANet algorithm effectively reconstructs hyperspectral images from CASSI measurements.
• The integration of PAN image reconstruction enhances spatial detail and regularization.
• SANet offers a robust solution for noisy CASSI data, improving HSI reconstruction quality.