基于自高斯与通道注意力的重塑卷积高光谱图像分类算法  

Hyperspectral image classification method based on reshaped convolutional channel attention with light self-Gaussian attention

作　　者：谭云飞 李明[1] 罗勇航 文贵豪 石超山 TAN Yunfei;LI Ming*;LUO Yonghang;WEN Guihao;SHI Chaoshan(College of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《辽宁工程技术大学学报(自然科学版)》2024年第6期102-111,共10页Journal of Liaoning Technical University (Natural Science)

基　　金：国家自然科学基金项目(61877051,61170192)。

摘　　要：针对传统卷积受限固有的网络结构,缺乏建立远程依赖关系的能力和分类精度较差等问题,提出一种基于自高斯与通道注意力的重塑卷积高光谱图像分类算法(RC-LSGA)模型。RC-LSGA模型首先采用卷积层提取浅层空间信息的特征,再使用通道注意力机制增强光谱特征,然后通过LSGA Transformer模块和重塑卷积分支对全局-局部特征信息进行提取,最后将获得的特征输入分类器实现分类。RC-LSGA模型能够有效区分不同波段信息,对PU、SA和LK数据集中类别识别的平均准确率分别达到98.20%、99.33%和99.46%。实验结果表明,在训练样本数量有限的情况下,RC-LSGA模型性能优异,在分类任务中实用价值较高。Aiming at the problems of traditional convolution limited inherent network structure,lack of ability to establish remote dependencies and poor classification accuracy,a reshaped convolution hyperspectral image classification algorithm model based on self-Gaussian and channel attention(RC-LSGA)is proposed.The RC-LSGA model first uses the convolutional layer to extract the features of shallow spatial information,and then uses the channel attention mechanism to enhance the spectral features.Then,the global-local feature information is extracted by the LSGA Transformer module and the reshaping convolution branch.Finally,the obtained features are input into the classifier to achieve classification.The RC-LSGA model can effectively distinguish different band information,and the average accuracy of category recognition in PU,SA and LK datasets is 98.20%,99.33%and 99.46%respectively.The experimental results show that the RC-LSGA model has excellent performance and high practical value in classification tasks when the number of training samples is limited.

关 键 词：高光谱图像分类 通道注意力 LSGA Transformer模块 重塑卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]