基于改进YOLOv7的连铸坯表面缺陷检测算法  

Surface defect detection algorithm for continuous casting billets based on improved YOLOv7

作　　者：马智华 曾凯 陈波 肖鹏程 朱立光 

机构地区：[1]华北理工大学电气工程学院,河北唐山063210 [2]华北理工大学冶金与能源学院,河北唐山063210 [3]河北省高品质钢连铸工程技术协同创新中心,河北唐山063000 [4]河北科技大学材料科学与工程学院,河北石家庄050018

出　　处：《中国冶金》2024年第7期101-112,共12页China Metallurgy

基　　金：国家自然科学基金资助项目(51904107);中央引导地方科技发展资金资助项目(236Z1017G);河北省博士研究生创新资助项目(CXZZBS2021096);唐山市市级科技计划资助项目(22130220G)。

摘　　要：针对连铸生产过程中铸坯表面缺陷检测准确率不高、检测速度慢等问题,提出一种基于改进YOLOv7-Tiny的铸坯表面缺陷检测算法。首先,提出一种结合分布移位卷积的自适应高效层聚合网络,增强对不同尺度特征信息的提取能力,提高模型泛化能力和计算效率;其次,在特征融合部分引入协调注意力模块,增强通道和位置信息感知,提高对重要特征的捕获能力,加强特征融合;同时,引入基于最小点距离的MPDIoU损失,解决现有损失函数局限性,增强回归结果准确性并加速模型收敛;最后,根据采集的实际生产中的铸坯缺陷图像,构建铸坯缺陷数据集并进行验证。试验结果表明,改进YOLOv7算法相较于基础网络有较大提升,在参数量保持不变的情况下,计算复杂度降低约35%,准确率提升4.4个百分点,平均精度均值(mAP)提高2.8个百分点,检测速度为130帧/s,能满足连铸生产现场对缺陷的实时检测要求,同时在NEU-DET公开数据集上体现出较强的泛化性。本研究提出的铸坯表面缺陷检测算法为提高缺陷检测精度与优化检测流程提供了技术支撑。To address the challenges of low accuracy and slow detection speed in identifying surface defects of casting billets during continuous casting production,an enhanced surface defect detection algorithm based on improved YOLOv7-Tiny was presented.Firstly,adaptive and efficient layer aggregation network coupled with distributed shift convolution was introduced to augment the ability of feature information extraction at different scales,thereby enhancing model generalization and computational efficiency.Secondly,coordinated attention module was incorporated into the feature fusion component to amplify the perception of channel and position information,thereby improving the capturing ability of crucial features and strengthening feature fusion.Simultaneously,the Minimum Point Distance IoU(MPDIoU)loss was introduced to overcome limitations in existing loss functions,which could boost the accuracy of regression results and expedite model convergence.Lastly,utilizing actual production images of continuous casting billets defects,a dataset was constructed and verified.Experimental results reveal that it has substantial improvements for improved YOLOv7algorithm compared to the baseline network.The improved YOLOv7algorithm achieves 35%reduction in computational complexity,4.4percent point increase in accuracy,and 2.8percent point increase in average precision(mAP)while maintaining the same number of parameters.With detection speed of 130FPS,it meets real-time defect detection requirements at continuous casting production sites and demonstrates robust generalization on the NEU-DET public dataset.The defect detection algorithm proposed in this study offers technical support for enhancing the accuracy of defect detection and optimizing the inspection process.

关 键 词：铸坯表面缺陷 目标检测 YOLOv7-Tiny 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] TF777[自动化与计算机技术—计算机科学与技术]