摘 要:针对传统算法检测钢材表面缺陷(如开裂、斑块、划痕等)精准度较低的问题,提出一种基于分割与分类的两段式深度学习网络。该网络是专为表面缺陷的检测、分割以及分类而设计的。第一阶段利用 YOLOv5 算法对钢材表面的缺陷进行定位、分割;第二阶段使用 EfficientNet 网络对钢材表面的六种缺陷类型进行分类。实验结果表明,相较于传统的 YOLOv5 算法,该方法的平均精准度提高了 16%,适合用于钢材表面缺陷检测。
关键词:深度学习;YOLOv5;缺陷检测;钢材表面
DOI:10.19850/j.cnki.2096-4706.2023.03.034
基金项目:湖南省教育厅重点项目(19A446);邵阳学院研究生科研创新项目(CX2022SY051)
中图分类号:TP391.4;TG115 文献标识码:A 文章编号:2096-4706(2023)03-0147-04
Steel Surface Defect Detection Method Based on Double Network
XIE Lianghui 1,2, ZHAO Chenglin1,2
(1.Hunan Provincial Key Laboratory of Southwest Hunan Rural Informatization Service, Shaoyang 422000, China; 2.School of Information Engineering, Shaoyang University, Shaoyang 422000, China)
Abstract: In order to solve the problem of low accuracy of traditional algorithms for detecting steel surface defects such as cracks, patches and scratches, a two-stage deep learning network based on segmentation and classification is proposed in this paper, the network is designed for the detection, segmentation and classification of surface defects. In the first stage, uses YOLOv5 algorithm to locate and segment the defect on the steel surface. In the second stage, uses the EfficientNet network to classify six types of defects on steel surfaces. The experimental results show that the average accuracy of this method is improved by 16% compared with the traditional YOLOv5 algorithm, and it is suitable for steel surface defect detection.
Keywords: deep learning; YOLOv5; defect detection; steel surface
参考文献:
[1] 刘坤,张晓怿,陈海永 . 带钢表面质量在线检测系统研究与设计 [J]. 计算机测量与控制,2015,23(6):1881-1884.
[2] 刘坤,陈宁纪,张晓怿 . 实时带钢表面质量检测算法研究与应用 [J]. 机械设计与制造,2015(7):74-76+79.
[3] 徐凤云 . 基于神经网络的钢材表面缺陷快速检测 [D]. 南京:南京理工大学,2009.
[4] 吴平川,路同浚,王炎 . 钢板表面缺陷的无损检测技术与应用 [J]. 无损检测,2000(7):312-315.
[5] REN S Q,HE K M,GIRSHICK R,et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks [J]. IEEE transactions on pattern analysis and machine intelligence,2017,39(6):1137-1149.
[6] 刘洋 . 基于深度学习的工业钢材瑕疵检测研究 [D]. 合肥:安徽建筑大学,2021.
[7] 李一鸣,王潇 . 基于 YOLOv5s 模型的轧钢表面缺陷检测[J]. 制造业自动化,2021,43(11):117-119.
作者简介:谢良辉(1996—),男,汉族,河南濮阳人,硕士研究生在读,研究方向:机器视觉。