基于Ghost模块的改进YOLOv5目标检测算法  

Improved YOLOv5 target detection algorithm based on Ghost module

作　　者：李宇翔 王帅[1,2] 陈伟[1,3] 田子建[1] 侯麟朔[1] LI Yuxiang;WANG Shuai;CHEN Wei;TIAN Zijian;HOU Linshuo(College of Mechatronics&Information Engineering,China University of Mining&Technology(Beijing),Beijing 100083,China;Ordos Division,Department of Inner Mongolia Administration Coal Mine Safety,Ordos 017000,China;School of Computer Science&Technology,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083 [2]内蒙古煤矿安全监察局鄂尔多斯监察分局,内蒙古鄂尔多斯017000 [3]中国矿业大学计算机科学与技术学院,江苏徐州221116

出　　处：《现代电子技术》2023年第3期29-34,共6页Modern Electronics Technique

基　　金：国家自然科学基金项目(52074305)。

摘　　要：现有以YOLOv5为代表的目标检测技术,存在骨干网络对特征提取不充分以及颈部层未高效融合浅层位置信息和深层高级语义信息等问题,这会导致检测精度较低,小目标误检、漏检。针对此问题,从兼顾实时性与检测精度出发,对YOLOv5进行改进,提出一种改进网络YOLOv5-CBGhost。首先在骨干网络中引入Ghost模块对模型进行轻量化处理,引入CA模块来更好地获得全局感受野,提高模型获取目标位置的准确度;然后借鉴双向加权特征金字塔网络,对原PAN结构进行改进,有效减少了特征冗余以及参数量,并通过跨层加权连接融合更多特征,提高了模型的目标检测精度;最后,增加多检测头以获取图片更丰富的高层语义信息,有效增加了检测精度。通过在PASCAL VOC2007+2012数据集上实验,YOLOv5-CBGhost的目标精度达到81.8%,相较于YOLOv5s,提高了3.0%,计算量减少42.5%,模型大小减少3.5%。The existing object detection technology represented by YOLOv5 has problems such as insufficient feature extraction of the backbone network and the failure of efficient fusion of shallow position information and deep high-level semantic information in the neck layer,which will lead to low detection accuracy,false detection of small targets and missed detection.To address this problem,an improved network YOLOv5-CBGhost is proposed in this paper by improving YOLOv5 from the perspective of taking into account real-time performance and detection accuracy.Firstly,the Ghost module is introduced into the backbone network to execute the lightweight processing for the model,and Coordinate attention(CA)module is introduced to better obtain the global sensory field and improve the model accuracy of obtaining the target location.The original PAN structure is improved by means of the bi-directional weighted feature pyramid network,which effectively reduces the feature redundancy and the number of parameters,and improves the model′s target detection accuracy by fusing more features through cross-layer connection.Finally,multiple detection heads are added to obtain richer high-level semantic information of the pictures,which effectively increases the detection accuracy.The result of experiment on the PASCAL VOC2007+2012 dataset show that the target detection accuracy of YOLOv5-CBGhost reaches 81.8%,which is 3.0%higher than that of YOLOv5s,the amount of computation is reduced by 42.5%,and the model size is reduced by 3.5%.

关 键 词：目标检测 YOLOv5改进 Ghost模块 模型处理 PAN结构改进 特征融合 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]