基于半监督语义分割的钢板表面缺陷检测方法  

A semi-supervised semantic segmentation method for steel plates surface defects detection

作　　者：吴昆鹏 石杰[1,2] 杨朝霖 邓能辉[1,2] WU Kunpeng;SHI Jie;YANG Chaolin;DENG Nenghui(National Engineering Technology Research Center of Flat Rolling Equipment,University of Science and Technology Beijing,Beijing 100083,China;Design and Research Institute Co.,Ltd.,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学国家板带生产先进装备工程技术研究中心,北京100083 [2]北京科技大学设计研究院有限公司,北京100083

出　　处：《冶金自动化》2023年第4期93-99,共7页Metallurgical Industry Automation

基　　金：广西科技重大专项(AA22068080)。

摘　　要：表面缺陷的有效检测是保证钢材表面质量的重要控制手段。现有检测方法多依赖于深度学习进行模型构建,需要大量的标记样本进行学习,延长了检测系统投用进度,因此提出了一种基于半监督语义分割方法来检测中厚板表面缺陷。首先在增加区域权重分支的双边分割网络(bilateral segmentation network,BiSeNet V2)上训练教师模型,再结合分类器辅助判断机制进行未标记样本的伪标签构建,最后将多样的背景图像和缺陷前景合成扩增样本数据集,训练得到学生模型。试验结果表明,该方法能够依赖少量的标签样本将平均识别率提高至90.5%,平均漏检率降低至2.2%,在实际应用中取得了很好的效果。The effective detection of surface defects is an important control means to ensure the steel surface quality.Most of the existing detection methods rely on deep learning for model construction,which requires a large number of labeled samples for learning,thus prolonging the use progress of the detection system.Therefore,a semi-supervised semantic segmentation method was proposed to detect the surface defects of medium and thick plates.Firstly,the teacher model was trained on BiSeNet V2 with increased area weight branches,and then the pseudo label construction of unlabeled samples was carried out in combination with the auxiliary judgment mechanism of the classifier.Finally,a variety of background images and defect prospects were combined into an expanded sample dataset to train the student model.The experimental results show that this method can rely on a small number of label samples to improve the average recognition rate to 90.5%,and reduce the average miss detection rate to 2.2%,which has achieved good results in practical applications.

关 键 词：钢板 缺陷检测 半监督学习 语义分割 BiSeNet V2 

分 类 号：TG142[一般工业技术—材料科学与工程] TG115[金属学及工艺—金属材料] TP391.41[金属学及工艺—金属学]