特异小样本工业产品表面缺陷检测方法研究  

Research on the detection method for special small-sample defects in industrial products

作　　者：郑李明[1] 许天赐 高浩然 李庆华 胡晨光 窦智 Zheng Liming;Xu Tianci;Gao Haoran;Li QingHua;Hu Chenguang;Dou Zhi(School of Mechanical and Electrical Engineering,Jinling Institute of Technology,Nanjing 211169,China;School of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China;Laiwu Iron and Steel Group Yinshan Section Steel Co.,Ltd.,Jinan 271104,China)

机构地区：[1]金陵科技学院机电工程学院,南京211169 [2]河南师范大学计算机与信息工程学院,河南新乡453007 [3]莱芜钢铁集团银山型钢铁有限公司,济南271104

出　　处：《河南师范大学学报（自然科学版）》2024年第6期88-96,共9页Journal of Henan Normal University(Natural Science Edition)

基　　金：国家自然科学基金(1904123,61901160);山东钢铁股份有限公司科技创新项目(GF2023014B)。

摘　　要：基于机器视觉的工业产品表面缺陷检测设备和系统大量应用在工业制造领域,目前其难点在于工业检测数据的采集,由于训练样本缺失导致深度学习网络模型无法有效训练.为解决上述问题,首先,提出一种基于不规则掩码的伤痕样本生成算法,改善了钢板表面缺陷检测任务中特异小样本数据集正负样本不均衡的情况;然后,在YOLOv8主干网络引入MHSA多头自注意力,提高对钢板表面缺陷的关注度;最后,使用SIoU替换原损失函数,增强网络模型的定位能力,提高检测的准确性.基于热轧钢板表面缺陷检测问题的实验结果表明,该方法能够有效解决特异小样本工业探伤的具体问题.Machine vision-based industrial product surface defect detection equipment and systems are widely used in the industrial manufacturing field.Currently,the main difficulty lies in the collection of industrial inspection data and the inability of deep learning network models to be effectively trained due to the lack of training samples.To solve these problems,firstly,this paper proposes a scar sample generation algorithm based on irregular masks to improve the imbalance of positive and negative samples in the special small sample dataset for steel plate surface defect detection task;then,the MHSA multi-head self-attention is introduced into the YOLOv8 backbone network to enhance the attention to steel plate surface defects;finally,the SIoU loss function is used to replace the original loss function to enhance the network model's localization ability and improve detection accuracy.The experimental results on the hot rolled steel plate surface defect detection problem based on this method show that can be effectively solved.

关 键 词：深度学习 目标检测 YOLOv8 注意力机制 数据增强 特异小样 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]