Supervise-Assisted Self-Supervised Deep-Learning Method for Hyperspectral Image Restoration.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Hyperspectral image (HSI) restoration is complex due to inverse problems and distribution gaps in deep learning models.
• Existing methods struggle with noise inherent in HSI degradation, further complicating restoration.
• Data-driven approaches often fail on target HSIs because of dissimilarities between training and target data.

Purpose of the Study:

• To develop a robust deep learning method for restoring noisy hyperspectral images.
• To overcome limitations of existing methods in handling distribution gaps and noise.
• To improve the accuracy and visual quality of HSI restoration across various tasks.

Main Methods:

• A supervise-assisted self-supervised deep learning framework for HSI restoration.
• Supervised learning to establish a generalized prior from extensive datasets.
• Self-supervised learning utilizing target HSI specific priors and a novel noise-adaptive loss function.
• The noise-adaptive loss incorporates Stein's unbiased risk estimator (SURE) and total variation (TV) regularizer.

Main Results:

• The proposed method achieves superior performance in HSI denoising, compressive sensing, super-resolution, and inpainting.
• Outperforms state-of-the-art methods on benchmark datasets in both quantitative metrics and visual quality.
• Demonstrates effective restoration of noisy degraded HSIs by leveraging inner data statistics.

Conclusions:

• The supervise-assisted self-supervised method offers a significant advancement in HSI restoration.
• The noise-adaptive loss function is crucial for fine-tuning restoration networks in the presence of noise.
• This approach provides a generalized and effective solution for various HSI restoration challenges.