联合连续学习的残差网络遥感影像机场目标检测方法  

Airport Detection Method Combined with Continuous Learning of Residual-Based Network on Remote Sensing Image

作　　者：李竺强 朱瑞飞 马经宇[1] 孟祥玉 王栋 刘思言 Li Zhuqiang;Zhu Ruifei;Ma Jingyu;Meng Xiangyu;Wang Dong;Liu Siyan(Key Laboratory of Satellite Remote Sensinag Application Technology of Jilin Province,Chang Guang Satellite Technology Co.,Ltd.,Changchun,Jilin 130000,China;Changchaun Institute of Optics,Fine Mechanics and Phgysics,Chinese Academg of Sciences,Changchun,Jilin 130033,China;Jilin Province Land Survey&Planning Institute,Changchun,Jilin 130061,China)

机构地区：[1]长光卫星技术有限公司,吉林省卫星遥感应用技术重点实验室,吉林长春130000 [2]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [3]吉林省国土资源调查规划研究院,吉林长春130061

出　　处：《光学学报》2020年第16期173-185,共13页Acta Optica Sinica

基　　金：国家重点研发计划重点专项(2018YFB1004605);吉林省重点科技研发项目(20180201109GX)。

摘　　要：在现有的高分辨率、大尺度目标遥感图像的检测中,传统方法由于提取特征手段单一、速度慢而无法快速并准确地从光学遥感影像中实现机场目标的识别。受人类视觉系统层次认知的启发,提出了一种适用于中高分辨率光学遥感影像的机场目标检测网络(CLRNet)。首先构建深度残差块,并将其作为特征提取网络;然后基于生成的样本核心集,采用连续学习方式从海量遥感数据中逐次迭代,精调机场检测模型;经过连续学习得到了鲁棒性强、遗忘度低的检测模型,该模型可以准确快速地从海量复杂背景下的光学遥感影像中识别出机场目标,而且对薄云遮挡以及卫星拍摄不全的机场有较好的识别效果。选取国产吉林一号卫星影像数据集进行测试,结果表明:所提方法的检测精度mAP(IoU不小于0.5)可达0.9613,每景的检测时间为0.23s。In the existing high-resolution and large-scale target remote sensing image object detection,the traditional method cannot achieve airport target recognition from optical remote sensing images quickly and accurately due to the single feature extraction and slow speed.Inspired by the hierarchical cognition of the human visual system,the continuous learning of residual-based convolution neural network(CLRNet)suitable for medium and high resolution optical remote sensing images is proposed.Firstly,the depth residual block is constructed as the feature extraction network.Secondly,the continuous learning method is used to fine tune the airport detection model from the massive remote sensing data.After continuous learning process,the model with strong robustness and low forgetting degree is obtained.The model can accurately and quickly identify airport from optical remote sensing images under massive and complex backgrounds.Our model has a better recognition effect for airports covered by thin clouds or incompletely captured by satellites.The domestic Jilin-1 satellite image dataset is selected for testing.Experiments show that the accuracy of the detection method mAP(IoU is not less than 0.5)can reach 0.9613,and the detection speed can reach 0.23s per scene.

关 键 词：遥感 连续学习 核心集 机场检测 残差卷积神经网络 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]