摘要
针对高分辨率遥感图像在目标检测中存在准确率低、目标漏检的问题,提出一种基于改进Faster R⁃CNN的遥感目标检测算法。采用ResNet⁃50作为主体网络进行特征提取,降低模型参数量和硬件资源占用,将ResNet⁃50的多尺度特征进行融合,进一步丰富中小目标的细节信息和位置信息。根据遥感目标尺寸的实际分布特点,采用K⁃means算法生成聚类中心,针对遥感目标尺度差异过大的问题,对聚类中心进行均值化操作,生成自适应锚点框参数,增强了区域建议网络(RPN)对多尺度目标的搜索能力,节约了人工根据经验设置锚点框参数的时间和精力。实验结果表明,改进算法能够有效地在多种复杂背景下检测不同尺度的遥感目标,在TRGS⁃HRRSD公共数据集上获得了83.76%的平均精度,召回率达到78.6%。
In view of the low accuracy and missing inspection in the object detection of high⁃resolution remote sensing image,a remote sensing object detection algorithm based on improved Faster R⁃CNN is proposed.ResNet⁃50 is adopted as the main network to extract features,reduce the number of model parameters and the occupancy of hardware resources,and integrate the multi⁃scale features of ResNet⁃50 to further enrich the detail and positioning information of small and medium⁃sized objects.The K⁃means algorithm is used to generate the clustering centers according to the actual distribution characteristics of the scale of the remote sensing objects.Since the scale differences among the remote sensing objects are excessively large,the clustering centers are averaged to generate the adaptive anchor frame parameters,which enhances the searching ability of RPN(region proposal network)for multi⁃scale objects,and saves the time and energy of manually setting anchor frame parameters according to experience.The experimental results show that the improved algorithm can effectively detect remote sensing objects with different scales under a variety of complex backgrounds,and its average precision obtained with the TRGS⁃HRRSD public data set is 83.76%,and its recall rate reaches 78.6%.
作者
马宇
单玉刚
袁杰
MA Yu;SHAN Yugang;YUAN Jie(School of Electrical Engineering,Xinjiang University,Urumqi 830001,China;College of Education,Hubei University of Arts and Science,Xiangyang 441000,China)
出处
《现代电子技术》
2022年第3期58-63,共6页
Modern Electronics Technique
基金
国家自然科学基金项目(61863033)
新疆维吾尔自治区“天山青年计划”:优秀青年科技人才培养项目(2019Q018)
湖北省教育厅科学技术研究项目(B2016175)
湖北文理学院博士基金项目(2015B002)
