基于深度学习的高分遥感影像典型地物检测方法应用实践  

Application practice of high resolution remote sensing imagestypical ground object detection methods based on deep learning

作　　者：孙广伟 李博 陈嘉浩 张大富[1] 范俊甫[1] SUN Guangwei;LI Bo;CHEN Jiahao;ZHANG Dafu;FAN Junfu(School of Architectural Engineering,Shandong University of Technology,Zibo 255049,China)

机构地区：[1]山东理工大学建筑工程学院,山东淄博255049

出　　处：《山东理工大学学报（自然科学版）》2021年第6期53-57,共5页Journal of Shandong University of Technology:Natural Science Edition

基　　金：山东省自然科学基金项目(ZR2020MD015);山东理工大学专业学位研究生教学案例库建设项目(4053-219059);山东理工大学实验室建设项目(2019014)。

摘　　要：传统的基于高分辨率遥感影像的典型地物检测方法难以兼顾检测精度、处理速度和自动化程度,而深度学习方法在图像处理领域中的应用为解决上述问题提供了可能。选用RSOD-Dataset数据集,基于TensorFlow深度学习框架,采用Faster R-CNN、YOLOv3和SSD三种经典深度学习目标检测算法,对高分辨率遥感影像数据进行预处理并完成模型训练,实现了对高分辨率遥感影像中典型地物目标的自动化检测实践,并对三种深度学习算法进行了对比分析。实验结果表明:SSD算法进行典型地物检测的均值平均精度指标(mAP)达到86.62%,每秒帧数(FPS)达到60.26,均优于Faster R-CNN算法和YOLOv3算法;在对飞机、储油罐、立交桥、体育场等典型地物的检测过程中有效地提升了检测精度、处理速度和计算的自动化水平,具有较为明显的应用价值。It is difficult to take into account detection accuracy,processing speed and automation degree in the traditional ground object detection method based on high-resolution remote sensing images,but the extensive application of deep learning in image processing makes it possible to solve the above problems.In this paper,based on the Faster R-CNN,YOLOv3(you only look once),SSD(single shot detector)deep learning object detection algorithm,the RSOD-Dataset is selected,and based on the TensorFlow deep learning framework,the high-resolution remote sensing images data were pre-processing and the model training was completed,the automatic detection of typical object targets in high-resolution remote sensing image data was realized and three deep learning algorithms were compared and analyzed.The experimental results show that the mean Average Precision(mAP)of SSD algorithm for typical ground object detection reaches 86.62%,and the number of frames per second(FPS)reaches 60.26,both better than the Faster R-CNN algorithm and YOLOv3 algorithm.The SSD object detection algorithm effectively improves the detection accuracy,processing speed,and calculation automation level in the detection of typical features such as aircraft,oil tanks,overpasses,and playgrounds,and has obvious practical application value.

关 键 词：高分影像 深度学习 目标检测 地物检测 

分 类 号：P209[天文地球—测绘科学与技术]