面向堆叠薄型钢板计数的改进YOLOv5网络  

An improved YOLOv5 network for stacked thin steel plate counting

作　　者：侯维岩[1] 刘忠英 郭怀远 翟鹏飞 HOU Weiyan;LIU Zhongying;GUO Huaiyuan;ZHAI Pengfei(School of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450001,China)

机构地区：[1]郑州大学电气与信息工程学院,河南郑州450001

出　　处：《安徽大学学报（自然科学版）》2024年第6期55-62,共8页Journal of Anhui University(Natural Science Edition)

基　　金：国家自然科学基金重大研究计划项目(9206710030)。

摘　　要：针对堆叠薄型钢板的人工计数烦琐、错误率高问题,提出改进的YOLOv5(you only look once的第5个版本)网络.使用SPD-Conv(symmetric positive definite convolution)代替CSP-Darknet(cross stage partial darknet),将轻量高效通道注意力(efficient channel attention,简称ECA)机制融入特征提取,将BiFPN(bi-directional feature pyramid network)代替YOLOv5的PANet(path aggregation network),提出HIOU_Loc(height intersection over union location)预测框去冗余方法.实验结果表明:相对于其他3种网络,该文网络的计数准确度(accuracy,简称ACC)和平均准确率均值(mean average precision,简称mAP)均最高;该文网络能适应不同检测环境,且均有优良的检测性能.To address the issues of labor-intensive manual counting and high error rates in stacked thin steel plate detection,an improved YOLOv5(you only look once,version 5)network was proposed.SPD-Conv(symmetric positive definite convolution)was employed to replace CSP-Darknet(cross stage partial darknet),and the efficient channel attention(ECA)mechanism was integrated into feature extraction.Additionally,BiFPN(bi-directional feature pyramid network)was used to replace thePANet(path aggregation network)in YOLOv5,and an HIOU_Loc(height intersection over union location)method for redundant bounding box prediction was introduced.The experimental results demonstrated that,compared to three other networks,the proposed network achieved the highest accuracy(ACC)and mean average precision(mAP).Furthermore,the proposed network was adaptable to various detection environments and exhibited excellent detection performance.

关 键 词：堆叠薄型钢板计数 YOLOv5 SPD-Conv 注意力机制 BiFPN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]