摘要
钢材表面缺陷是影响钢材质量的因素之一。现有的缺陷检测方法难以在提供少量标注样本的前提下实现钢材缺陷的准确检测。针对这一问题,提出了一种改进的缺陷半监督检测方法Mean Teacher-BA。首先,提出了一种双级注意力机制(bi-level attention module,BAM),以增加模型对复杂缺陷的关注和检测能力;其次,对主干网络Resnet50提取的特征图像应用BAM模块,进一步提升对缺陷图像的特征提取能力;最后,提出了一种全局自适应损失策略,以改善在训练期间生成的伪标注质量。实验结果表明,Mean Teacher-BA在5%、10%、20%、100%的标注比例条件下比原模型Mean Teacher分别提升了1.0%、1.2%、1.5%和0.9%,相比较于其他全监督和半监督网络性能更好,更适合于工业生产的部署应用。
Surface defects are easy to occur in the process of steel production,and the existing defect detection methods can′t achieve accurate detection of steel defects under the premise of providing a few labels.To solve this problem,an improved semi-supervised defect detection method called Mean Teacher-BA is proposed in this paper.Firstly,a bi-level attention module(BAM),is proposed to increase the model′s ability to pay attention to and detect complex defects.Secondly,BAM module is applied to the feature images extracted from backbone network Resnet50 to further improve the feature extraction capability of defect images.Finally,a global adaptive loss strategy is proposed to improve the quality of pseudo-tags generated during training.The experimental results show that the Mean Teacher-BA proposed in this paper is 1.0%,1.2%,1.5%and 0.9%higher than the original Mean Teacher model respectively under the conditions of 5%,10%,20%and 100%.Compared with other supervised and semi-supervised networks,the proposed Mean Teacher-BA is more suitable for deployment and application in industrial production.
作者
王晓宾
沈飞翔
张强
陈成军
WANG Xiaobin;SHEN Feixiang;ZHANG Qiang;CHEN Chengjun(School of Mechanical and Automotive Engineering,Qingdao University of Technology,Qingdao 266520,China;Wabco China Co.,Ltd.,Qingdao 266555,China)
出处
《组合机床与自动化加工技术》
北大核心
2024年第12期85-90,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家自然科学基金项目(52175471)
山东省自然科学基金项目(ZR2021MF110)。
