华东师范大学学报(自然科学版) ›› 2020, Vol. 2020 ›› Issue (6): 105-114.doi: 10.3969/j.issn.1000-5641.201921021

• 计算机科学 • 上一篇    下一篇

基于深度学习的铝材表面缺陷检测

张旭1, 黄定江1,2   

  1. 1. 华东理工大学 理学院, 上海 200237;
    2. 华东师范大学 数据科学与工程学院, 上海 200062
  • 收稿日期:2019-08-27 发布日期:2020-12-01
  • 通讯作者: 黄定江, 男, 研究员, 博士生导师, 研究领域为机器学习与人工智能. E-mail: djhuang@dase.ecnu.edu.cn E-mail:djhuang@dase.ecnu.edu.cn
  • 基金资助:
    国家自然科学基金(U1711262, 11501204)

Defect detection on aluminum surfaces based on deep learning

ZHANG Xu1, HUANG Dingjiang1,2   

  1. 1. School of Science, East China University of Science and Technology, Shanghai 200237, China;
    2. School of Data Science and Engineering, East China Normal University, Shanghai 200062, China
  • Received:2019-08-27 Published:2020-12-01

摘要: 随着信息技术在工业制造领域的深入应用, 工业制造大数据研究正成为实现智能制造、帮助政府指导制造企业转型升级的重要参考依据. 在传统的钢铁、铝材等金属制造行业, 更是存在生产方式粗放、生产工艺简单等问题. 因此, 迫切需要利用人工智能等新一代信息技术来改善生产流程,提高生产效率. 在使用铝材时, 必须检查铝材表面. 现有的铝材表面缺陷检测受限于传统人工肉眼检查, 十分费力, 或基于传统的机器视觉算法, 识别率不高, 通常不能及时准确地判断出表面瑕疵. 为解决这些问题,利用深度学习来进行铝材表面缺陷检测: 首先运用两大目标检测算法Faster R-CNN(Region-CNN(Convolutional Neural Networks))和YOLOv3对制作的铝材缺陷数据集进行检测; 然后基于YOLOv3算法进行改进, 提升铝材表面很小缺陷的检测效果. 在广东工业智造大数据创新大赛提供的“铝型材瑕疵识别”数据集上进行了实验验证, 实验结果显示, 改进算法的平均精度均值(mean Average Precision, mAP)比YOLOv3算法高3.4%, 比Faster R-CNN算法高1.8%.

关键词: 铝材缺陷, 检测, 机器视觉, 深度学习

Abstract: With the application of information technology to industrial manufacturing, big data research in industrial manufacturing has become an important basis for realizing intelligent manufacturing and helping the government to guide the transformation and upgrade of manufacturing enterprises. In traditional metal manufacturing industries, like steel and aluminum, there are challenges, such as uncalibrated and rudimentary production techniques. Therefore, it is urgent to improve production processes and production efficiency by using newer generation information technology, such as artificial intelligence. The surface of aluminum, for example, must be inspected during the manufacturing process. Existing surface defect detection techniques for aluminum materials is limited to traditional manual naked eye inspection-which is very laborious-or traditional machine vision algorithms-which have a low rate of detection. Hence, surface defects cannot be accurately judged in a timely fashion. This paper studies the use of deep learning for surface defect detection of aluminum materials. Firstly, two major target detection algorithms, Faster R-CNN and YOLOv3, are used to detect the defects on a set of aluminum materials. Then, based on the YOLOv3 algorithm, the detection ability of small defects on aluminum surfaces was improved. We conducted experiments on the “Aluminum Profile Defect Recognition” dataset provided by the Guangdong Industrial Intelligent Manufacturing Big Data Innovation Competition. The experimental results showed that the mean average precision (mAP) of the improved algorithm is 3.4% higher than that of YOLOv 3 and 1.8% higher than that of Faster R-CNN.

Key words: aluminum defect, detection, machine vision, deep learning

中图分类号: