基于改进YOLOv3模型的热轧板卷表面缺陷检测技术研究  

Research on hot-rolled coil surface defects detection technology based on improved YOLOv3 model

作　　者：麻越 李月林 张琳 雷大伟 徐昊天 MA Yue;LI Yuelin;ZHANG Lin;LEI Dawei;XU Haotian(Automation Research and Design Institute of Metallurgical Industry Co.,Ltd.,Beijing 100071,China;State Key Laboratory of Metallurgical Intelligent Manufacturing System,Beijing 100071,China;HBIS Group Laoting Iron and Steel Co.,Ltd.,Tangshan 063600,China)

机构地区：[1]冶金自动化研究设计院有限公司,北京100071 [2]冶金智能制造系统全国重点实验室,北京100071 [3]河钢乐亭钢铁有限公司,河北唐山063600

出　　处：《冶金自动化》2024年第3期83-91,共9页Metallurgical Industry Automation

摘　　要：热轧板是钢铁产品生产的重要原材料,其表面缺陷直接影响着最终钢铁产品的质量,因此热轧板表面缺陷的检测技术对于控制产品质量有着重要的意义。为更准确、高效实现热轧卷板表面质量判定,降低唐钢热轧产线的人工成本,进而提高生产效率,研究热轧卷板表面缺陷自动检测技术势在必行。本文在YOLOv3模型基础上进行了改进:引入改进注意力模块SE-PRE,提高模型对缺陷区域的关注程度;增加检测分支扩大感受野,提高模型对小目标缺陷检测的能力;引入dropout检测头,有效控制了训练过拟合现象。同时在模型训练上,优化了学习率设置、损失函数及最优模型的选择策略;在数据上,利用数据增强技术扩充了数据量。通过现场大量数据实验,结果表明,本文的模型在检测精度和效率上均取得了较好的结果,对降低现场人工成本,提高产品质量具有重要意义。Hot-rolled coil is an important raw material for steel production,and its surface defects di-rectly affect the quality of the final steel products.Therefore,the detection technology of surface de-fects on hot-rolled coils is of significant importance for controlling product quality.In order to achieve more accurate and efficient determination of the surface quality of hot-rolled coils,reduce labor costs in Tangshan Iron and Steel's hot-rolling production line,and thus improve production efficiency,stud-ying automatic detection technology for surface defects on hot-rolled coils is imperative.This paper proposes improvements to the YOLOv3 model:the introduction of the SE-PRE improved atention module to increase the model's focus on defect areas;the addition of a detection branch to expand the receptive field,improving the model's ability to detect small target defects;and the introduction of a dropout detection head to effectively control overftting during training.Additionally,in terms of model training,the learning rate setting,loss function,and selection strategy for the optimal model were optimized.Data augmentation techniques is also used to expand the data volume.Through a large number of on-site data experiments,the results show that the model in this paper has achieved good results in both detection accuracy and efficiency.This is significant for reducing on-site labor costs and improving product quality.

关 键 词：热轧卷板 缺陷检测 数据增强 注意力模块 小目标检测 过拟合 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] TG335.11[金属学及工艺—金属压力加工]