融合感受野扩增与特征增强的遥感小目标检测  

Remote sensing small target detection integrating receptive field amplification and feature enhancement

作　　者：帖军[1,2] 秦锦添 郑禄 郑明雪[3] 陈婷 TIE Jun;QIN Jintian;ZHENG Lu;ZHENG Mingxue;CHEN Ting(College of Computer Science South-Central Minzu University,Wuhan 430074,China;Hubei Provincial Engineering Research Center of Agricultural Blockchain and Intelligent Management,Wuhan 430074,China;Institute of Agricultural Economics and Technology,Hubei Academy of Agricultural Sciences,Wuhan 430064,China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]农业区块链与智能管理湖北省工程研究中心,武汉430074 [3]湖北省农业科学院农业经济技术研究所,武汉430064

出　　处：《激光杂志》2024年第12期81-91,共11页Laser Journal

基　　金：国家民委中青年英才培养计划(No.MZR20007);湖北省技术创新计划重点研发专项(No.2023BAB087);新疆维吾尔自治区区域协同创新专项(科技援疆计划)(No.2022E02035);武汉市知识创新专项曙光计划项目(No.2023010201020465)。

摘　　要：针对遥感图像小目标因背景复杂、尺寸较小以及排列密集等因素导致检测精度不高的问题,提出一种融合感受野扩增与特征增强的遥感小目标检测方法。该方法以YOLOv8s作为基线网络,首先对主干网络特征提取部分构造感受野扩增模块,通过双层路由注意力(BRA)高效捕获全局特征信息;其次在特征金字塔部分构建浅层特征融合结构,并在浅层特征图横向连接部分添加改进的坐标空间注意力(CSA),以增强小目标的特征信息;最后通过改进非极大值抑制(NMS)算法对检测结果进行后处理,以适应不同密度物体的检测。在DIOR遥感图像数据集上进行实验,预测框与真实框之间的交并比阈值(IoU)为0.5时的平均精度均值(mAP)达到90.3%,比原模型高出3%;IoU为0.5∶0.95时的mAP达到71.3%,比原模型高出6.1%,实验结果表明,改进模型对遥感图像小目标检测任务具有较好的应用价值。Aiming at the problem of low detection accuracy of small targets in remote sensing images due to complex background,small size and dense arrangement,a remote sensing small target detection method integrating receptive field amplification and feature enhancement is proposed.Using YOLOv8s as the baseline network,the method firstly constructs a receptive field amplification module for the feature extraction part of the backbone network,and efficiently captures the global feature information through the Bi-Level Routing Attention(BRA);secondly,it constructs a shallow feature fusion structure in the feature pyramid part,and adds the improved coordinate spatial attention(CSA)in the transverse connectivity part of the shallow feature map,in order to enhance the feature information of the small targets;Finally,the detection results are post-processed by an improved non-maximum suppression(NMS)algorithm to adapt to the detection of objects with different densities Experiments are carried out on the DIOR remote sensing image dataset,the mean average precision(mAP)reaches 90.3%when the intersection and concurrency ratio threshold(IoU)between the predicted frame and the real frame is 0.5,which is 3%higher than that of the original model;and the mAP reaches http∶//www.laserjournal.cn 71.3%when the IoU is 0.5∶0.95,which is 6.1%higher than that of the original model,and the experimental results show that the improved model has a good application value for the small target detection in remote sensing images.

关 键 词：感受野扩增 特征增强 遥感小目标 YOLOv8s 

分 类 号：TN911[电子电信—通信与信息系统]