改进残差神经网络在遥感图像分类中的应用  

Application of Improved Residual Network in Sensing Image Classification

作　　者：刘春容 宁芊[1,2] 雷印杰 陈炳才[3] LIU Chun-rong;NING Qian;LEI Yin-jie;CHEN Bing-cai(College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China;College of Physics and Electronic Engineering, Xinjiang Normal Universit, Urumqi 830054, China;College of Computer Science and Technology, Dalian University of Technology, Dalian 116024, China)

机构地区：[1]四川大学电子信息学院,成都610065 [2]新疆师范大学物理与电子工程学院,乌鲁木齐830054 [3]大连理工大学计算机科学与技术学院,大连116024

出　　处：《科学技术与工程》2021年第31期13421-13429,共9页Science Technology and Engineering

基　　金：国家自然科学基金(U1903215);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(2019E0214)。

摘　　要：针对传统卷积神经网络随着深度加深而导致网络退化以及计算量大等问题,提出一种改进残差神经网络的遥感图像场景分类方法。该方法以残差网络ResNet50作为主框架,在残差结构中引入深度可分离卷积和分组卷积,减少了网络的参数量和计算量,加快模型收敛的同时也提升了分类精度。此外在网络中嵌入多尺度squeeze and excitation block模块对通道特征进行重校准,提取出更加重要的特征信息,进一步提升了网络的分类性能。在航空图像数据集(aerial image dataset,AID)和UCMerced_Land Use两个公开数据集上的分类精度分别为91.92%和93.52%,相比常规残差网络分类精度分别提高了3.38%和10.24%,证明所提方法在遥感图像场景分类任务中的可行性和有效性。In terms of that the deepening of traditional convolution neural network led to various problems,including the degradation of the network and the large amount of calculation,a remote sensing image scene classification method was proposed to improve the residual neural network.In this method,the residual network ResNet50 was taken as the main framework,and the deep separable convolution and grouping convolution were introduced into the residual structure,and thus the amount of network parameters and calculation were reduced,the convergence of the model was sped up and the classification accuracy was improved.Furthermore,the multi-scale squeeze and excitation block module was embedded into the network to recalibrate the channel features and extract more important feature information,and the classification performance of the network was further improved.The classification accuracy of AID(aerial image dataset)and UCMerced_Land Use reached 91.92%and 93.52%,respectively,which were 3.38%and 10.24%higher than that of the conventional residual network.It was proved that the proposed method in remote sensing image scene classification task is feasible and effective.

关 键 词：遥感图像 场景分类 残差神经网络 分组卷积 深度可分离卷积 多尺度缩聚与激发模块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]