基于改进YOLOv7的热轧板材表面缺陷检测研究  

Research on Surface Defect Detection of Hot-Rolled Sheet Based on Improved YOLOv7

作　　者：孙铁强 麻文建 宋超 肖鹏程 SUN Tieqiang;MA Wenjian;SONG Chao;XIAO Pengcheng(College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China;Hebei Provincial Key Laboratory of Industrial Intelligent Perception,North China University of Science and Technology,Tangshan 063210,China;School of Metallurgy and Energy,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 [2]华北理工大学河北省工业智能感知重点实验室,唐山063210 [3]华北理工大学冶金与能源学院,唐山063210

出　　处：《组合机床与自动化加工技术》2024年第6期146-150,共5页Modular Machine Tool & Automatic Manufacturing Technique

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：为提高热轧板材表面检测的速度以及检测的精度,提出一种改进的YOLOv7-BRS目标检测算法。首先,对YOLOv7中ELAN结构进行改进,提出一种新型计算模块BRConv,使用深度可分离卷积并添加多分支的跳跃连接方式来减小模型复杂度,实现模型轻量化并提高检测速度;其次,设计了一种新型多尺度识别的注意力机制,拥有不同的感受野,进一步提高模型对不同尺度重要特征的提取能力;最后,对损失函数进行改进,引入角度损失概念,重新定义了惩罚指标,提升模型训练时的收敛速度以及准确性。实验表明,改进后的模型体积减小了36%,在NEU-DET数据集上mAP提高了7.3%,FPS提高了14.4。相比于目前主流算法,检测精度和速度都有显著提高,并且体积更小,能够有效完成板材表面缺陷检测任务。In order to improve the speed and accuracy of surface detection of hot-rolled sheet,an improved YOLOv7-BRS object detection algorithm is proposed.Firstly,the ELAN structure in YOLOv7 is improved,and a new computing module BRConv is proposed,which uses deep separable convolution and adds multi-branch hop connection to reduce the complexity of the model,realize model lightweight and improve the detection speed.Secondly,a new attention mechanism for multi-scale recognition is designed,which has different receptive fields,which further improves the model's ability to extract important features at different scales.Finally,the loss function is improved,the concept of angle loss is introduced,the penalty index is redefined,and the convergence speed and accuracy during model training are improved.Experiments show that the improved model volume is reduced by 36%,mAP is increased by 7.3%,and FPS is increased by 14.4 on the NEU-DET dataset.Compared with the current mainstream algorithms,the detection accuracy and speed are significantly improved,and the volume is smaller,which can effectively complete the task of detecting surface defects on the surface of the plate.

关 键 词：缺陷检测 YOLOv7 模型轻量化 注意力机制 损失函数 

分 类 号：TH165[机械工程—机械制造及自动化] TG66[金属学及工艺—金属切削加工及机床]