A Dual-Strategy Learning Framework for Hyperspectral Image Super-Resolution

Hyperspectral image super-resolution (HSI SR) has achieved remarkable success with deep neural networks. Currently, most methods in HSI SR assume a predetermined degradation model during training to synthesize low-resolution images. These methods falter when confronted with HSI exhibiting degradatio...

Full description

Saved in:
Bibliographic Details
Main Authors: Shuying Li, Ruichao Sun, San Zhang, Qiang Li
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10891577/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Hyperspectral image super-resolution (HSI SR) has achieved remarkable success with deep neural networks. Currently, most methods in HSI SR assume a predetermined degradation model during training to synthesize low-resolution images. These methods falter when confronted with HSI exhibiting degradation patterns and their limited flexibility restricts practical application. In addition, these methods focus on the complex network designs for superior performance, which entail high resource consumption and limit their broad application. To address these issues, in this article, we propose a dual-strategy learning framework exploring meta-transfer learning for HSI blind SR. This framework can be applied to any SR network and facilitate performance enhancement. First, we pretrain a three-channel SR model on natural image data to address the issue of insufficient HSI data. Furthermore, we innovatively propose a transfer scheme, which directly applies our pretrained three-channel SR model to HSI, thereby significantly enhancing the spectral fidelity. To enhance the model's performance under specific degradation conditions, we incorporate meta-learning, enabling it to adapt to input images after a few iterations. Besides, we introduce attention-based knowledge distillation to equip our final network with the implicit representation capability of a meta network under a lightweight premise. Extensive experiments on three benchmark datasets demonstrate that the proposed approach outperforms existing methods in various degradations.
ISSN:1939-1404
2151-1535