基于改进YOLOv5的钢筋节点检测方法  

Rebar Node Detection Method Based on Improved YOLOv5

作　　者：于铭铭 段浩 郭帅[2] 蒋海里 艾腾峰 

机构地区：[1]上海大学机电工程与自动化学院,上海200444 [2]上海大学工程训练技术中心,上海200444 [3]上海公路桥梁(集团)有限公司,上海200433

出　　处：《工业控制计算机》2025年第2期81-82,85,共3页Industrial Control Computer

基　　金：国家自然科学基金(U1913603)资助。

摘　　要：原YOLOv5算法参数量、计算量过大,不能部署在算力有限的嵌入式设备上。基于YOLOv5算法提出轻量化的钢筋节点检测算法YOLOv5-FAS,将YOLOv5的Backbone部分替换为FasterNet网络,减少参数量和计算量;将YOLOv5的head部分由PAN-FPN替换为AFPN,渐近融合低层纹理信息和高层语义信息,减少特征信息的丢失和退化;将YOLOv5的损失函数替换为ECIOU,提升模型定位精度。在钢筋节点私有数据集上的测试结果表明,与YOLOv5s相比,改进算法参数量和计算量分别减少39.4%和38%,检测精度提升0.3%,检测速度提升160%,该算法适用于钢筋捆扎机器人对钢筋节点的实时检测要求。The original YOLOv5 algorithm parameters are too computationally intensive and cannot be deployed on embedded devices with limited computing power.Based on the YOLOv5 algorithm,a lightweight steel node detection algorithm YOLOv5-FAS is proposed,which replaces the Backbone part of YOLOv5 with the FasterNet network to reduce the amount of parameters and calculations.The head part of YOLOv5 is replaced by PAN-FPN with AFPN,and asymptotically fuses low-level textures information and high-level semantic information to reduce the loss and degradation of feature information,replace the loss function of YOLOv5 with ECIOU to improve the model positioning accuracy.Test results on the steel node private data set show that compared with YOLOv5s,the improved algorithm has reduced parameters and calculations by 39.4%and 38%respectively,increased detection accuracy by 0.3%,and increased detection speed by 160%.This algorithm is suitable for steel bar bundling.Requirements for real-time detection of steel nodes by robots.

关 键 词：钢筋节点检测 轻量化 特征提取 特征融合 损失函数 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]