改进YOLOv5s的遥感图像检测研究  

Improved YOLOv5s remote sensing image detection research

作　　者：钱承山[1] 沈有为 孙宁 卢峥松 戴仁天 Qian Chengshan;Shen Youwei;Sun Ning;Lu Zhengsong;Dai Rentian(College of Automation,Nanjing University of Information Science&Technology,Nanjing 211800,China;Wuxi University,Wuxi 214000,China)

机构地区：[1]南京信息工程大学自动化学院,南京211800 [2]无锡学院,无锡214000

出　　处：《国外电子测量技术》2022年第11期57-66,共10页Foreign Electronic Measurement Technology

基　　金：无锡市现代产业发展资金项目(20201012);2022江苏省研究生科研实践创新计划(SJCX22_0349);江苏省科技副总项目(FZ20200099)资助。

摘　　要：针对遥感图像中目标排列紧密,背景复杂的问题,设计Transformer和卷积的双向交互模块(CTN)作为网络特征提取结构,使模型能够弱化背景噪声带来的干扰且能更好的捕获全局信息。其次,为了加强特征提取网络在复杂背景下的提取能力,构建了DenseBlock模块和ConvBlock模块,所设计的模块能增强模型在多目标下多尺度学习的能力,相比原网络能提取出更丰富的语义信息。最后对数据集中所有实例分布进行统计分析,其存在的许多小目标容易使原网络存在漏检误检的现象,针对这种情况,在检测头部分额外添加了一个检测头来缓解目标尺度变化带来的负面影响,同时去除对检测效果提升不明显的特征提取分支及检测分支,使用K-means++重新聚类得到最优锚框并分配至裁剪后的3个预测特征层。实验结果表明,改进的网络能有效改善遥感图像的漏检与误检的情况,在目标密集分布的情况下提升YOLOv5s的检测能力,改进的网络能更快收敛,均值平均精度(mean average precision, mAP)相比于原YOLOv5s算法提高了3.1%。Aiming at the problem that the targets in remote sensing images are closely arranged and the background is complex, a bidirectional interaction module(CTN) of transformer and convolution is designed as a network feature extraction structure, so that the model can weaken the interference caused by background noise and better capture global information. Secondly, in order to strengthen the extraction ability of the feature extraction network in complex backgrounds, the DenseBlock module and the ConvBlock module are constructed. The designed modules can enhance the model′s ability to learn at multiple scales under multiple targets. Compared with the original network, it can extract richer semantic information. Finally, the distribution of all instances in the data set is statistically analyzed, and many small targets exist in the original network, which can easily cause the phenomenon of missed detection and false detection in the original network. In response to this situation, an additional detection head is added to the detection head part to alleviate the negative impact of target scale changes, at the same time, the feature extraction branches and detection branches that are not obvious for detection improvement are also removed. K-means++ is used to re-cluster to get the optimal anchor box and assign the obtained anchor box to the three cropped prediction feature layers. The experimental results show that the improved network can effectively improve the missed detection and false detection of remote sensing images, and improve the detection ability of YOLOv5 s in the case of dense distribution of targets. Compared with the original YOLOv5 s algorithm, the improved network can converge faster and the mean average precision(mAP) is 3.1% than the original algorithm.

关 键 词：目标检测 YOLOv5s 遥感目标 自注意力机制 网络结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]