文献类型：期刊文献

中文题名：融合改进密集连接和分布排序损失的遥感图像检测

英文题名：Improved YOLOv3 Remote Sensing Target Detection Based on Improved Dense Connection and Distributional Ranking Loss

作者：袁磊 刘紫燕 朱明成 马珊珊 陈霖周廷

第一作者：袁磊

机构：[1]贵州大学大数据与信息工程学院,贵阳550025;[2]贵州理工学院航空航天工程学院,贵阳550003

第一机构：贵州大学大数据与信息工程学院,贵阳550025

年份：2021

卷号：48

期号：9

起止页码：168-173

中文期刊名：计算机科学

外文期刊名：Computer Science

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：贵州省科学技术基金资助项目(黔科合基础[2016]1054);贵州省联合资金资助项目(黔科合LH字[2017]7226);贵州大学2017年度学术新苗培养及创新探索专项(黔科合平台人才[2017]5788);贵州省科技计划项目(黔科合基础[2017]1069);贵州省教育厅创新群体重大研究项目(黔教合KY字[2018]026);贵州省普通高等学校工程研究中心(黔教合KY字[2018]007);贵州省科技计划重点项目([2019]1416)。

语种：中文

中文关键词：遥感图像;目标检测;YOLOv3;基础网络;样本不平衡

外文关键词：Remote sensing image;Object detection;YOLOv3;Baseline;Sample imbalance

摘要：针对遥感图像中小目标尺寸较小、样本分布不均匀、特征不明显等问题,提出一种改进的YOLOv3目标检测算法。在使用Stitcher数据增强解决小目标样本分布不均匀的问题后,提出VOVDarkNet-53基础网络,将DarkNet-53基础网络中第4次下采样后的8个残差模块减少为4个残差模块。然后采用VOVNet的密集连接方式,使网络利用更多的浅层小目标特征信息,增加网络感受野。最后,采用分布排序损失改进YOLOv3中的分类损失,解决单阶段目标检测器正负样本不平衡的问题。实验使用YOLOv3目标检测算法和改进后的YOLOv3算法在HRRSD遥感数据集上进行对比。结果表明,改进后的YOLOv3算法对小目标和中目标的检测精确度分别提升了7.2%和2.1%,尽管对大目标的检测精度下降了1%,但在平均单张图片处理时间几乎不变的情况下,平均检测精度均值(mAP)提升了4.1%,召回率和准确率也有所提升。
Aiming at solving the problems of small object size,uneven sample distribution,and unclear features in remote sensing images,an improved YOLOv3 object detection algorithm is proposed.The Stitcher data enhancement method is used to solve the problem of uneven distribution of small object samples.The VOVDarkNet-53 is proposed.The residual modules of the fourth downsampling in DarkNet-53 are reduced from eight to four.And then the dense connection mode of VOVNet is adopted to extract lower features of small objects to increase the network receptive field.The distributional ranking loss is used to improve the classification loss in YOLOv3 to solve the problem of imbalance between positive and negative samples in single-stage object detector.Comparative experiments are carried out on HRRSD remote sensing datasets by using YOLOv3 object detection algorithm and improved YOLOv3 algorithm.The results demonstrate that the proposed algorithm can achieve better performance of higher detection accuracy of the improved YOLOv3 algorithm for small objects and medium objects are improved by 7.2%and 2.1%,respectively.Although the detection accuracy for large objects is reduced by 1%,the average detection accuracy(mAP)is improved by 4.1%,and the recall and accuracy are also improved.