Band-kernel Stochastic Learning for Unsupervised Blind Hyperspectral Image Super-Resolution.

Area of Science:

• Computer Vision
• Remote Sensing
• Signal Processing

Background:

• Hyperspectral image super-resolution (HSI-SR) is challenging due to high spectral dimensionality.
• Supervised HSI-SR methods require costly labeled data, limiting generalization.
• Existing unsupervised methods decouple band selection and kernel estimation, creating performance trade-offs.

Purpose of the Study:

• To develop a unified statistical framework for blind HSI-SR.
• To introduce the first unsupervised blind HSI-SR method, BKSR.
• To address limitations of existing supervised and unsupervised HSI-SR approaches.

Main Methods:

• Proposed BKX-HMM, a hidden Markov model (HMM) framework for unified band selection, kernel estimation, and HSI restoration.
• Developed BKSR, comprising Gibbs sampling-based band selection (GBS), test-time-training kernel estimation (TKE), and robust HSI restoration (RHR).
• Integrated a spectral hyper-Laplacian prior into a diffusion model for noise-robust HSI restoration.

Main Results:

• BKSR demonstrates superior performance over baseline methods on synthetic and real HSI datasets.
• The method effectively handles diverse scenarios, including unknown kernels and non-i.i.d. noise.
• Achieved comparable computational costs to traditional band selection methods.

Conclusions:

• BKSR offers an effective unsupervised solution for blind HSI-SR.
• The unified framework redefines performance trade-offs as a distributional fitting problem.
• BKSR provides a robust and generalizable approach to HSI-SR.