.png){kind=icon}
.png){kind=icon}
Return
A spatial-spectral sparse dynamic graph learning network for hyperspectral image denoising
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132198.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hyperspectral image (HSI) denoising is a crucial task that significantly affects subsequent hyperspectral analysis. However, previous deep learning-based HSI denoising methods often struggle to capture sparse long-range spatial correlations, restricting their ability to effectively exploit global critical spatial-spectral information and resulting in issues such as detail loss and blurry edges. This paper proposes a novel HSI denoising framework using graph neural networks (GNNs), called a spatial-spectral sparse dynamic graph learning network (SSDGLN), which dynamically integrates global spatial-spectral information and effectively leverages global spectral correlations. Its core is a learnable threshold-based spatial-spectral dynamic graph (SSDG) module. This module first constructs a learnable threshold-based sparse graph attention (LTGA) block, effectively modeling the sparse long-range correlations in spatial positions and collecting globally similar spectral features. Then LTGA is combined with dynamic convolution to capture both global contextual features and subtle local features. Afterward, a global multi-scale feature learning module uses different scales of SSDG to enhance edge details. Also, a dense spatial-spectral feature extraction module is devised by stacking SSDG modules with dense connections to promote multi-level feature reuse. Both simulated and real-data experiments validate that SSDGLN outperforms mainstream HSI denoising methods in quantitative metrics and visual fidelity, providing a reliable strategy for applying GNN to HSI denoising.
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W