Coded Hyperspectral Image Reconstruction Using Deep External and Internal Learning.

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Conventional hyperspectral cameras face limitations in spatial and temporal resolution.
• Coded hyperspectral imaging offers a promising alternative to overcome these limitations.
• Reconstructing hyperspectral images (HSI) from coded data is an ill-posed inverse problem requiring accurate prior information.

Purpose of the Study:

• To develop an effective convolutional neural network (CNN) for coded hyperspectral image reconstruction.
• To address the low spatial and temporal resolution issues inherent in conventional hyperspectral imaging.
• To learn deep priors for accurate HSI recovery using both external and internal data characteristics.

Main Methods:

• A CNN-based channel attention reconstruction network was developed to leverage spatial-spectral correlations in HSI.
• The network was trained using an external hyperspectral dataset with adversarial loss to capture general spatial-spectral correlations.
• Internal learning with spatial-spectral constraints and total variation regularization was applied for customized reconstruction and to prevent overfitting.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques in coded hyperspectral image reconstruction.
• Quantitative metrics and perceptive quality assessments confirmed the effectiveness of the CNN-based approach.
• The method successfully reconstructed HSIs from both synthetic and real-world coded data.

Conclusions:

• The developed CNN-based method effectively reconstructs hyperspectral images from coded data, addressing resolution limitations.
• Learning deep priors from external datasets and internal image information with spatial-spectral constraints is crucial for accurate HSI recovery.
• This approach offers a significant advancement in coded hyperspectral imaging systems, providing high-quality reconstructions.