基于神经网络注意力架构搜索的光学遥感图像场景分类  

Neural Network Attention Architecture Search for Optical Remote Sensing Image Scene Classification

作　　者：曹斌 郑恩让[1] 沈钧戈 Cao Bin;Zheng Enrang;Shen Junge(School of Electrical and Control Engineering,Shaanxi University of Science and Technology,Xi’an 710021,China;Unmanned System Research Institute,Northwestern Polytechnical University,Xi’an 710072,China)

机构地区：[1]陕西科技大学电气与控制工程学院,陕西西安710021 [2]西北工业大学无人系统技术研究院,陕西西安710072

出　　处：《遥感技术与应用》2023年第4期913-923,共11页Remote Sensing Technology and Application

基　　金：国家自然科学基金项目(61603233);河南省水下重点实验室开放基金项目(D5204200587)。

摘　　要：针对光学遥感图像场景分类存在类别变化、样本数量变化,场景图像中背景与重要物体变换大、尺度变化多的问题,提出基于神经网络注意力架构搜索的光学遥感图像场景分类方法,由算法自适应在神经网络中搜索卷积、池化、注意力等操作,构建能完成光学遥感图像场景分类任务的神经网络。为保证搜索神经网络过程稳定性,提出两段式贪婪策略网络搜索方法,分阶段丢弃无用操作,减少搜索算法负担、提高搜索速度。最后为了关注各物体与场景关联信息,提出自上而下的网络连接策略,充分复用各阶段多尺度特征图的语义。实验结果证明:该方法相较于手工设计的经典深度学习方法具有更好的性能。在AID、NWPU、PATTERNET 3个遥感图像标准数据集上总体精度均超过经典方法。在AID数据集上准确率达到94.04%;在PATTERNET数据集上准确率达到99.62%;在NWPU数据集上达到95.49%。With majority problems in image scene of optical remote,changing category in classification,varia⁃tional size in sample,diverse changing of scale between backgrounds and essential objectives,for instance,new Classification Algorithm for scene classification of optical remote sensing image base on attention architecture search of neural network is proposed in this paper.This algorithm can search convolution,pooling,attention and other operations in the neural network,adaptively;and complete the construction task of scene classifica⁃tion for optical remote sensing images in neural network.Two-stage greedy algorithms network search is men⁃tioned in order to ensure the stability of neural search network.This method abandons useless operations in stage which can reduce algorithm burden and improve speed of search.Furthermore,a top-bottom connection strategy of network,which can fully reuse the semantics of multi-scale feature maps in each stage,is proposed to merge information between each object and scene.The experimental results proved that the method proposed in this paper has better performance than the classical deep learning method designed by hand.Overall,the accu⁃racy of this method in all three remote sensing image-standard data sets(AID,NWPU and PatterNet)is ex⁃ceeding the classic method.The accuracy rate of AID data set,PatterNet data set,and NWPU data set are 94.04%,99.62%,and 95.49%,respectively.

关 键 词：遥感 场景分类 神经网络架构搜索 贪婪算法 网络连接策略 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP18[自动化与计算机技术—控制科学与工程]