基于卷积神经网络与注意力机制的高光谱图像分类  

Hyperspectral image classification based on convolutional neural network and attention mechanism

作　　者：高玉鹏 闫伟红[2] 潘新[1] GAO Yupeng;YAN Weihong;PAN Xin(College of Computer and Information Engineering,Inner Mongolia Agricultural University,Hohhot,Inner Mongolia 010018,China;Institute of Grassland Research,Chinese Academy of Agricultural Sciences,Hohhot,Inner Mongolia 010010,China)

机构地区：[1]内蒙古农业大学计算机与信息工程学院,内蒙古呼和浩特010018 [2]中国农业科学院草原研究所,内蒙古呼和浩特010010

出　　处：《光电子．激光》2024年第5期483-489,共7页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61962048,61562067);中央级基本科研业务费(1610332020020)资助项目。

摘　　要：由于浅层卷积神经网络(convolutional neural network,CNN)模型感受野的限制,无法捕获远距离特征,在高光谱图像(hyperspectral image,HSI)分类问题中无法充分利用图像空间-光谱信息,很难获得较高精度的分类结果。针对上述问题,本文提出了一种基于卷积神经网络与注意力机制的模型(model based on convolutional neural network and attention mechanism,CNNAM),该模型利用CA(coordinate attention)对图像通道数据进行位置编码,并利用以自注意力机制为核心架构的Transformer模块对其进行远距离特征提取以解决CNN感受野的限制问题。CNNAM在Indian Pines和Salinas两个数据集上得到的总体分类精度分别为97.63%和99.34%,对比于其他模型,本文提出的模型表现出更好的分类性能。另外,本文以是否结合CA为参考进行了消融实验,并证明了CA在CNNAM中发挥重要作用。实验证明将传统CNN与注意力机制相结合可以在HSI分类问题中获得更高的分类精度。Due to the limitation of the receptive field of the shallow convolutional neural network(CNN)model,long distance features cannot be captured,and the spatial-spectral information of the image cannot be fully utilized in hyperspectral image classification,so it is difficult to obtain high precision classification results.To solve these problems,this paper proposes a model based on convolutional neural network and attention mechanism(CNNAM).The model uses coordinate attention(CA)to encode the position of the image channel data,and uses the Transformer module with self-attention mechanism as the core architecture to extract the long distance features to solve the CNN receptive field limitation problem.The overall classification accuracy of the proposed model on Indian Pines and Salinas datasets is 97.63%and 99.34%.Compared with other models,the proposed model shows better classification performance.In addition,the ablation experiments are carried out based on whether CA is combined with CNNAM,and it is proved that CA playes an important role in CNNAM.Experiments show that the combination of traditional CNN and attention mechanism can achieve higher classification accuracy in HSI classification.

关 键 词：高光谱图像分类(HSI) 卷积神经网络(CNN) coordinate attention Transformer 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]