一种在轨卫星光学遥感舰船检测算法  

Algorithm of Ship Detection for Optical Remote Sensing Satellite on Oribit

作　　者：瞿涛 韩传钊[2] QU Tao;HAN Chuanzhao(School of Computer Science,Wuhan University,Wuhan 430072,China;Institute of Beijing Remote Sensing Information,Beijing 100192,China)

机构地区：[1]武汉大学计算机学院,武汉430072 [2]北京市遥感信息研究所,北京100192

出　　处：《航天器工程》2021年第2期31-39,共9页Spacecraft Engineering

基　　金：教育部高校国防基础科研项目(JCKY2018110C163)。

摘　　要：为了解决资源有限的在轨处理环境对光学遥感舰船检测网络速度和精度的影响,提出了应用倒置残差结构的舰船检测算法。采用端到端网络作为基础模型,裁剪网络的层数和3/4通道数,构建2层卷积倒置残差模块,使用可分离卷积和1×1卷积对普通卷积层进行替换,只保留1个检测分支,使用金字塔检测方式实现多尺度目标检测网络。基于高分一号卫星数据集的试验结果表明:相较于当前的主流网络,算法模型在平均精度(AP)相差不大的情况下检测速度有很大提升。和YOLOv3相比,检测速度提升了72%,参数量减少了99.5%;和YOLOv2相比,检测速度和AP分别提升了58%和0.9。文章提出的算法可为计算资源极其有限的卫星在轨舰船检测实现提供有效的理论技术支撑。A ship detection algorithm using inverted residual structure is proposed in order to reduce the influence of resource limitation on the speed and accuracy of optical remote sensing on-orbit detection network.An end-to-end network is used as the basic model.After the number of layers is cut and 3/4 channels of the network is removed,the two-layer convolutional inversion residual module is constructed.The convolutional layer is replaced by depth-wise separable convolution and 1×1 convolution,retaining only one detection branch.The multi-scale target detection network is realized by the pyramid detection method.The test results based on Gaofen-1 satellite dataset show that compared with current popular networks,the detection speed of the proposed model is greatly improved while the difference in AP(average precision)is relatively small.Compared with YOLOv3,the detection speed is increased by 72%and the number of parameters decreases by 99.5%.Compared with YOLOv2,the detection speed and AP are improved by 58%and 0.9,respectively.The proposed algorithm provides an effective theoretical and technical support for on-orbit satellite ship detection with very limited computing resources.

关 键 词：在轨卫星 光学遥感 舰船检测 倒置残差结构 网络压缩 金字塔检测 

分 类 号：V19[航空宇航科学与技术—人机与环境工程]