改进的YOLOv5-ResNet相似目标检测方法  

Improved YOLOv5-ResNet Method for Detecting Similar Objects

作　　者：赵桂平 邓飞[1] 王昀[2] 唐云[1] ZHAO Gui-ping;DENG Fei;WANG Yun;TANG Yun(College of Computer and Network Security,Chengdu University of Technology,Chengdu 610059,China;Institute of Seismic Acquisition Technology,Institute of Petroleum Geophysical Exploration Technology,Sinopec,Nanjing 211103,China)

机构地区：[1]成都理工大学计算机与网络安全学院(牛津布鲁克斯学院),成都610059 [2]中国石油化工股份有限公司石油物探技术研究院地震采集技术研究所,南京211103

出　　处：《科学技术与工程》2022年第30期13406-13416,共11页Science Technology and Engineering

基　　金：国家自然科学基金重点项目(41930112);中石化地球物理实验室基金(33550006-22-FW0399-0022)。

摘　　要：针对在相似目标检测问题中,以YOLOv5为代表的一步法漏检错检率高、以Faster R-CNN为代表的两步法检测速度慢的问题,提出了一种改进的YOLOv5-ResNet相似目标检测网络模型。该模型以YOLOv5框架为基础,借鉴了两步法的优点。在边框生成方面,改进了特征融合结构,强化了模型的特征提取能力,降低了总体漏检、误检率。在类别预测方面,引入SE(squeeze and excitation)模块,在通道方向上施加注意力机制,降低网络检测时的计算量,并保持了较高的准确率。在斯坦福宠物狗数据集和自制音符卡片数据集上的实验结果表明,本文提出的相似目标快速检测模型不仅在识别精度方面略高于Faster R-CNN,而在速度方面仅次于YOLOv5,检测帧率约为YOLOv5的72%,能够满足相似目标检测的实时需要。Aiming at the problem of similar object detection,a high rate of missed detection and false detection was performed by the one-stage method represented by YOLOv5,a low speed of detection was performed by the two-stage method represented by Faster R-CNN,an improved YOLOv5-ResNet model for detecting similar objects was proposed.The model was based on YOLOv5 framework and drew on the advantages of two-stage method.In the aspect of generating bounding boxes,the structure of feature fusion was improved to strengthen the feature extraction ability of the model on reducing the overall rate of missed detection and false detection.In the aspect of classification,squeeze and excitation(SE)module was introduced to exert attention mechanism on channels,which was used to reduce the amount of calculation and maintain high accuracy during detecting.The results on Stanford dogs dataset and custom music note card dataset show that the detecting accuracy of the model is slightly higher than Faster R-CNN and the detecting FPS(frames per second)is second only to YOLOv5,which reaches about 72%of YOLOv5’s,meets the demand of real-time detection.

关 键 词：目标检测 相似目标 YOLOv5-ResNet 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]