计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (19): 235-242.DOI: 10.3778/j.issn.1002-8331.2010-0446

• 图形图像处理 • 上一篇    下一篇

改进Mask R-CNN算法的带钢表面缺陷检测

翁玉尚,肖金球,夏禹   

  1. 1.苏州科技大学 电子与信息工程学院,江苏 苏州 215009
    2.苏州市智能测控工程技术研究中心,江苏 苏州 215009
  • 出版日期:2021-10-01 发布日期:2021-09-29

Strip Surface Defect Detection Based on Improved Mask R-CNN Algorithm

WENG Yushang, XIAO Jinqiu, XIA Yu   

  1. 1.College of Electronics and Information Engineering, Suzhou University of Science and Technology, Suzhou, Jiangsu 215009, China
    2.Intelligent Measurement and Control Engineering Technology Research Center, Suzhou University of Science and Technology, Suzhou, Jiangsu 215009, China
  • Online:2021-10-01 Published:2021-09-29

摘要:

在带钢的生产过程中可能会因为生产工艺的问题导致带钢表面出现缺陷,传统的带钢表面检测方法存在检测速度慢、检测精度低等问题。在计算机深度学习快速发展的今天,为实现带钢表面缺陷快速有效的检测,提出改进的掩码区域卷积神经网络(Mask R-CNN)算法,使用[k]-means II聚类算法改进区域建议网络(RPN)锚框生成方法;同时调整Mask R-CNN模型的网络结构,去掉掩码分支,提高了模型的缺陷检测速度。实验在NEU-DET数据集的5种缺陷检测中将原算法的均值平均精度(mAP)从0.810?2提升到0.960?2,检测速度达到5.9?frame/s。并且能够实现对缺陷目标的检测和实例分割,以便研究人员观测缺陷的大小和形状,从而改进工艺。相比于目前其他深度学习的缺陷检测算法,更能满足带钢的生产检测要求。

关键词: 深度学习, 带钢表面缺陷检测, 锚框, 聚类算法, 掩码分支

Abstract:

In the production process of strip steel, the surface defects of strip steel may be caused by the problems of production process. The traditional surface detection methods of strip steel have the problems of slow detection speed and low detection accuracy. With the rapid development of computer deep learning, in order to achieve rapid and effective detection of strip surface defect, this paper proposes a new algorithm based on improved Mask Region Convolution Neural Network(Mask R-CNN), and uses [k]-means II clustering algorithm to improve the anchor frame generation method of Region Proposal Network(RPN). At the same time, it adjusts the network structure of Mask R-CNN, removes the mask branches, and improves the defect detection speed of the model. In the experiments, the mean Average Precision(mAP) of the original algorithm is improved from 0.810?2 to 0.960?2 in 5 kinds of defect detections of NEU-DET data set, and the detection speed reaches 5.9?frame/s. And it can detect the defect target and segment the case, so that researchers can observe the size and shape of the defect, so as to improve the process. Compared with other deep learning defect detection algorithms, it can better meet the requirements of strip production detection.

Key words: deep learning, strip surface defect detection, anchor box, clustering algorithm, mask branching