基于光谱-空间注意力残差网络的高光谱图像分类  

Spectral-spatial Attention Residual Networks for Hyperspectral Image Classification

作　　者：汪菲菲 赵慧洁[1,2,3] 李娜 李思远[4] 蔡昱 WANG Feifei;ZHAO Huijie;LI Na;LI Siyuan;CAI Yu(Key Laboratory of Precision Opto-Mechatronics Technology,Ministry of Education,School of Instrumentation and Optoelectronic Engineering,Beihang University,Beijing 100191,China;Institute of Artificial Intelligence,Beihang University,Beijing 100191,China;Aerospace Optical-Microwave Integrated Precision Intelligent Sensing,Key Laboratory of Ministry of Industry and Information Technology,Beihang University,Beijing 100191,China;Key Laboratory of Spectral Imaging Technology,Xi′an Institute of Optics and Precision Mechanics,Chinese Academy of Sciences,Xi′an 710119,China;China Academy of Launch Vehicle Technology,Beijing 100076,China)

机构地区：[1]北京航空航天大学,仪器科学与光电工程学院精密光机电一体化技术教育部重点实验室,北京100191 [2]北京航空航天大学人工智能研究院,北京100191 [3]北京航空航天大学“空天光学-微波一体化精准智能感知”工信部重点实验室,北京100191 [4]中国科学院西安光学精密机械研究所,光谱成像技术重点实验室,西安710119 [5]中国运载火箭技术研究院,北京100076

出　　处：《光子学报》2023年第12期200-218,共19页Acta Photonica Sinica

基　　金：国家自然科学基金(No.61975004);预研项目(No.6230111002)。

摘　　要：在高光谱图像分类任务中,引入注意力改变提取到的光谱和空间特征权重,有效突出重要特征,提高分类准确率。将注意力机制、残差网络和特征提取模块集成到分类框架中,引入中心区域光谱注意力机制,在避免干扰像素对波段权重影响的同时,利用周围像素增强中心像素波段,增强光谱特征的鲁棒性进而提取有效的光谱特征。并在此基础上提出了光谱-空间注意力残差网络,该网络可以从高光谱图像中连续提取到丰富的光谱特征和空间特征,并通过残差网络连接特征提取模块,缓解了精度下降问题,保证网络良好的分类性能。在4个公开数据集上,所提出的分类算法和其他算法相比,各项指标均达到最优。Hyperspectral image classification is a research hotspot in the field of hyperspectral image processing and application.Classification models predict the class of each pixel by analyzing the spectral and spatial information of each pixel and compare it to the actual features.In the hyperspectral classification task,the spatial context information of the data can be used to improve the classification accuracy,so this paper uses the powerful learning ability of 3D-CNN to extract effective spectral and spatial features into hyperspectral images,and then fuses the extracted spectra and spatial features to enhance the flow between different levels of the network,thereby improving the classification efficiency.Although CNN operations can mine deeper feature information as the network deepens,CNN is ineffective in modeling long-distance dependencies,so consider combining CNN with attention mechanisms.This combination can focus on the local position of the given information,assign corresponding weights to it,emphasize the key features in the feature map,adjust the global information of the attention statistics image through weight re-annotation,retain the features that are more conducive to the classification task,and improve the representation ability of extracted features.But the common attention mechanism is to calculate the average globally,that is,the pixel values of the entire image block,inevitably introducing information from different categories of pixels around it,which is not needed in classification tasks.Another spectral attention mechanism based on the center pixel provides weight values that ignore the effects of surrounding pixels in the same category.Therefore,a simple spectral attention mechanism in the central region is proposed, in which the central region is selected with the central pixel as the reference and the surrounding3×3 range as the central region, on the one hand, the range contains certain spectral information of thesame category, and on the other hand, the interference of different categ

关 键 词：光谱-空间特征 残差网络 高光谱分类 光谱注意力机制 空间注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]