Abstract: Because of the complexity of image background,different defect characteristics and lack of standard defect data set,the hot rolled steel end defects are difficult to automatically identify with traditional visual methods.The deep learning-based technical architecture of defect detection for hot-rolled steel was presented,and the key techniques such as image acquisition,preprocessing and segmentation,deep neural network algorithm design and defect feature database construction were given.A deep learning-based steel roll end defect detection system was designed.The test showed that the defect detection rate of this system was more than 90%,which solved the industrial problem of long-term dependence on manual naked eye recognition of end surface defects of hot rolling steel,and realizes the online,automatic and accurate detection of end surface defects.A complete set of end defect detection technology and equipment were formed,which were used in many hot rolling lines in the steel industry.The application results showed that the detection system was stable and reliable.

Key words: steel roll, end defects, deep learning, identification

摘要： 针对热轧钢卷端面图像背景复杂、缺陷特征各异、缺乏标准缺陷数据集,用传统视觉方法难以自动识别问题,提出了基于深度学习的热轧钢卷端面缺陷检测的技术架构,给出了该框架下的图像采集、预处理与分割、深度神经网络算法设计和缺陷特征数据库搭建等关键技术,最后开发了基于深度学习的钢卷端面缺陷检测系统,试验表明该系统的典型缺陷检出率在90%以上,解决了热轧钢卷端面缺陷长期依赖人工肉眼识别的行业性难题,实现了端面缺陷的在线、自动、精准检测,所形成的成套端面缺陷检测技术和装备,在钢铁行业多条热轧产线稳定、可靠地投用运行。

关键词: 钢卷, 端部缺陷, 深度学习, 识别