一种结合多尺度策略的光谱-空间注意力网络用于高光谱图像分类  

A spectral-spatial attention network combining multi-scale strategies for hyperspectral image classification

作　　者：田亮 陈昊兵 郑波尽[1] TIAN Liang;CHEN Haobing;ZHENG Bojin(College of Computer Science,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2024年第4期532-539,共8页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(62373382)。

摘　　要：针对高光谱图像(HSI)分类任务中光谱和空间局部细节特征提取不足的问题,提出了一种创新的光谱-空间注意力网络MSSAN.该网络结构包含光谱和空间特征提取模块,每个模块都集成了多尺度扩张卷积块、残差提取块、密集提取块和注意力机制.残差和密集提取块整合浅层和深层特征,多尺度扩张卷积块辅助提取局部细节特征.随后的注意力机制凸显关键特征,充分利用光谱和空间信息.对比实验显示:MSSAN在IP、UP和SV三个数据集上表现出色,优于目前的先进算法.消融实验验证了MSSAN各模块组合的有效性.To address the issue of inadequate extraction of spectral and spatial local details in hyperspectral image(HSI)classification,an innovative spectral-spatial attention network named MSSAN is proposed.It includes spectral and spatial feature extraction modules with multi-scale dilated convolution blocks,residual and dense extraction blocks,and attention mechanism.The network effectively integrates shallow and deep features,with the multi-scale dilated convolution block enhancing the extraction of local details.The subsequent attention mechanism highlights key features and makes full use of spectral and spatial information.Comparative experiments demonstrate MSSAN's superior performance on IP,UP,and SV datasets compared to existing advanced algorithms.Ablation experiments confirm the efficacy of MSSAN's modular combination.

关 键 词：高光谱图像分类 卷积神经网络 多尺度卷积 注意力机制 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]