Noise Prior Knowledge Informed Bayesian Inference Network for Hyperspectral Super-Resolution.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep learning (DL) models are prevalent in hyperspectral image super-resolution (HS-SR).
• Existing DL models often overlook image prior information and lack interpretability.
• Current DL approaches for HS-SR are typically built from generic components, not tailored for the task.

Purpose of the Study:

• To develop an interpretable deep learning network for HS-SR that incorporates prior knowledge.
• To address the limitations of current black-box DL models in HS-SR.
• To improve the accuracy and understanding of HS-SR models.

Main Methods:

• Proposed BayeSR network, embedding Bayesian inference with Gaussian noise prior into a deep neural network.
• Constructed a Bayesian inference model solvable via proximal gradient algorithm.
• Unfolded the iterative algorithm into a network architecture, converting noise matrix operations into channel attention.

Main Results:

• The BayeSR network explicitly encodes prior knowledge from observed images.
• The model considers the intrinsic generation mechanism of HS-SR.
• Qualitative and quantitative experiments show BayeSR outperforms state-of-the-art methods.

Conclusions:

• BayeSR offers a superior and more interpretable approach to HS-SR compared to existing methods.
• The network effectively leverages Gaussian noise priors and image characteristics.
• The proposed method demonstrates significant advancements in hyperspectral image super-resolution.