| 摘 要: 现有基于深度学习的高光谱图像融合方法在处理复杂地物交界区域时极易引发特征串扰与边缘平滑现象。为解决这一问题，本文提出了一种基于超像素自适应注意力网络(Superpixel Adaptive Attention Network，SAAN)的高光谱图像融合算法。首先，利用多光谱图像生成超像素标签图，用于构建后续特征聚合与注意力掩膜所需的结构先验。随后，设计了一种结构引导的交叉注意力机制，自适应评估局部同质性并生成精准的注意力权重。最后，在加权聚合阶段，利用注意力权重动态指导特征重构。实验结果表明，SAAN模型在所选的两个实验数据集上均取得了卓越表现，显著优于其他先进对比算法。 |
| 关键词: 深度学习,高光谱图像融合,超像素分割,交叉注意力机制,双流网络 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Hyperspectral image fusion based on superpixel adaptive attention network |
|
Li Ketong, Liu Cong
|
University of Shanghai for Science and Technology
|
| Abstract: Existing deep learning-based hyperspectral image fusion methods are highly prone to induce feature crosstalk and edge over-smoothing when processing complex boundary regions of ground objects. To solve this problem, this paper proposes a hyperspectral image fusion algorithm based on the Superpixel Adaptive Attention Network (SAAN). First, a superpixel label map derived from the multispectral image is used to construct the structural prior for subsequent feature aggregation and attention masking. Subsequently, a structure-guided cross-attention mechanism is designed to adaptively evaluate local homogeneity and generate accurate attention weights. Finally, in the weighted aggregation stage, the attention weights are used to dynamically guide feature reconstruction. Experimental results show that the SAAN model achieves excellent performance on the two selected experimental datasets, and is significantly superior to other state-of-the-art comparison algorithms. |
| Keywords: deep learning hyperspectral image fusion superpixel segm-entation cross-attention mechanism dual-stream network |
