基于深度学习神经网络的SAR星上目标识别系统研究  

Research on SAR Satellite Target Recognition System Based on Deep Learning Neural Network

作　　者：袁秋壮 魏松杰 罗娜[1] 

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《上海航天》2017年第5期46-53,共8页Aerospace Shanghai

基　　金：国家自然科学基金资助(61472189);上海航天科技创新基金资助(F2016020013)

摘　　要：实现对地面目标的智能识别,对一种基于深度学习的卷积神经网络(CNN)的星载合成孔径雷达(SAR)星上目标识别系统进行了研究。系统由星上和地面两部分组成。其中:地面部分进行网络结构设计、SAR图像数据预处理、CNN模型训练、模型压缩及上传;星上部分接收上传模型并解压缩、目标识别、识别后粗筛图像下传地面;地面进行人工筛查,筛查后的正确图像作为训练数据对CNN模型进行再训练,逐步获得精度更高的模型。提出的CNN架构为卷积层2个、下采样层2个、Dropout层3个、Flatten层1个、全连接层2个,最终输出标签11类。为使训练后的CNN模型能部署到卫星上使用,采用数据精度压缩和剪枝两种数据深度压缩方法以减小数据存储量和减低网络复杂度。在Keras深度学习开源库环境中实现设计的CNN模型,对运动和静止目标获取与识别(MSTAR)数据库中的11类军事目标识别的实验结果表明:识别和分类的效果良好,整体识别成功率达96.29%;模型能压缩至原来的1/13,精度损失小于2%。To realize the intelligent recognition of ground target,a synthetic aperture radar（SAR）satellite target recognition system based on deep learning（DL）convolutional neural network（CNN）was studied in this paper.The system composed of two parts of onborne segment and ground segment.The network structure design,SAR image pretreatment,CNN model training,and the model compression and uploading were carried on in the ground segment.The compressed model receiving and decompressing,target recognition,recognized coarse screening image downloading were carried on in the onborne segment.Then the images downloaded were screened by manual labour in the ground segment.The screened images were served as the training data for CNN model again.It would obtain higher precision images step by step.There were 2 convolutional layers,2 lower sampling layers,3 Dropout layers,1 Flatten layer,2 connection layers and 11 output label in the CNN structure.To use the trained CNN model onborne,two deep compression method were adopted to decrease the data memory and network complex,which were data precision compression and pruning.The CNN model proposed was realized in Keras deep learning open source library.The experimental results in recognizing 11 classes of military targets in the database of Moving and Stationary Target Acquisition and Recognition（MSTAR）showed that the effect of recognition and classification was well.The overall recognition success rate was 96.29%.The model could be compressed to 1/13 of the original one while the accuracy loss rate was less than 2%.

关 键 词：SAR图像 智能目标识别 深度学习 卷积神经网络 深度压缩 数据精度压缩 剪枝 模型效率 

分 类 号：TN957.52[电子电信—信号与信息处理]