基于改进YOLOv7的钢卷端面缺陷检测  

The end face of steel coil defect detection based on improved YOLOv7

作　　者：孙铁强 秦愿伟 宋超 肖鹏程 SUN Tieqiang;QIN Yuanwei;SONG Chao;XIAO Pengcheng(College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学人工智能学院,唐山063210

出　　处：《现代制造工程》2024年第7期117-125,25,共10页Modern Manufacturing Engineering

基　　金：河北省“三三三人才工程”资助项目(A202102002);2023年唐山市重点研发项目(23140204A)。

摘　　要：针对钢卷端面纹理复杂、缺陷较小,以及YOLOv7算法识别速度较慢、小目标检测率较低等问题,提出一种改进的YOLOv7目标检测算法。对YOLOv7算法骨干网络中的ELAN结构进行改进,通过加入PConv卷积层设计了一种新的结构,以减少模型复杂度,提高模型的检测速度。由于小目标检测中容易出现漏检的现象,设计一种新的注意力机制CSCA,提高网络对小尺度目标的敏感度。在此基础上,使用WIoUv2损失函数替换原YOLOv7算法网络中的CIoU损失函数来优化损失函数,提高网络的鲁棒性。在自制的钢卷端面缺陷数据集上进行试验,结果表明改进的YOLOv7算法的mAP@0.5提升了4.1%,FPS提升了10.84 f/s,检测效果优于原算法。An improved YOLOv7 object detection algorithm was proposed to address the issues of complex texture and small defects on the end face of steel coils,as well as slow recognition speed and low detection rate of small targets using YOLOv7 algorithm.The ELAN structure in the YOLOv7 algorithm backbone network was improved by adding a PConv convolutional layer to design a new structure to reduce the model complexity and improve the model detection speed.Due to the tendency to miss detection in small object detection,a new attention mechanism CSCA was designed to improve the network′s sensitivity to small-scale objects.On this basis,WIoUv2 loss function was used to replace the CIoU loss function in the original YOLOv7 algorithm network to optimize the loss function and improve the robustness of the network.Experiments were conducted on a self-made dataset of steel coil end face defects,and the results showed that the improved YOLOv7 algorithm improved mAP@0.5 by 4.1%and FPS by 10.84 f/s,with better detection performance than the original algorithm.

关 键 词：钢卷端面缺陷检测 YOLOv7算法 注意力机制 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]