Dilated projection correction network based on autoencoder for hyperspectral image super-resolution.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Deep learning methods have advanced hyperspectral image (HSI) super-resolution (SR).
• Existing methods often overlook unique HSI characteristics, limiting performance.
• There is a need for HSI SR methods that leverage spectral information more effectively.

Purpose of the Study:

• To improve the spatial resolution of hyperspectral images (HSIs) by incorporating prior information.
• To develop an HSI SR method that fully utilizes spectral information and reduces computational complexity.
• To enhance the accuracy and efficiency of HSI super-resolution.

Main Methods:

• A novel autoencoder-based dilated projection correction network (aeDPCN) is proposed.
• The HSI SR problem is transformed from the image domain to the abundance domain.
• A coarse-to-fine super-resolution strategy with back-projection is employed for abundance embeddings.

Main Results:

• The proposed aeDPCN method significantly improves spatial resolution in HSIs.
• The abundance domain transformation reduces computational complexity while preserving spectral information.
• Experiments show superior accuracy and efficiency compared to state-of-the-art methods.

Conclusions:

• The aeDPCN method offers a stable and effective solution for HSI super-resolution, even at large upscaling factors.
• Transforming HSI SR to the abundance domain is a promising approach for leveraging spectral information.
• The method demonstrates state-of-the-art performance in both accuracy and computational efficiency.