基于孪生网络的带钢表面周期性缺陷检测方法  

Method for periodic defect detection of strip surface based on siamese network

作　　者：吴昆鹏 石杰 WU Kun-peng;SHI Jie(Institute of Engineering Technology,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学工程技术研究院,北京100083

出　　处：《冶金自动化》2020年第6期93-98,共6页Metallurgical Industry Automation

摘　　要：周期性缺陷作为热轧带钢生产过程中最常见、最重要的缺陷,其检测不能简单通过分类模型进行识别。首先从周期性缺陷的产生原因出发,总结出周期性缺陷的三大主要特征,即横向位置一致性、缺陷图像相似性和缺陷周期统一性;然后提出基于孪生网络的图像相似性检测方法,可通过三点实现周期性缺陷的快速判别;最后利用等距离差法拟合出缺陷的最佳周期,拟合精度可达±10 mm。将本文方法应用到实际生产现场中,周期性缺陷的检出率可达95%以上。As the most common and important defect in hot rolled strip production process,the detection of periodic defects cannot be simply identified by classification model.Starting from the causes of periodic defects,three main characteristics of periodic defects were summarized,namely,horizontal position consistency,defect image similarity and defect period uniformity.Then,a method of image similarity detection based on siamese network was proposed,which can quickly identify periodic defects through three points.Finally,the optimal period of the defect was fitted by the method of equidistance difference,and the fitting accuracy was up to±10 mm.The detection rate of periodic defects can reach more than 95%when the method was applied to actual production.

关 键 词：缺陷检测 孪生网络 周期拟合 图像相似性 等距离差法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TG335.56[自动化与计算机技术—计算机科学与技术] TP18[金属学及工艺—金属压力加工]