华中科技大学学报（自然科学版）2024，Vol.52Issue(6)：32-38,7.DOI:10.13245/j.hust.240122
基于IDACL深度度量学习的零件表面缺陷检测
Surface defect detection of mechanical parts based on IDACL deep metric learning
摘要
Abstract
Aiming at the problems that the deep metric learning model applied to the surface defect detection of mechanical parts is easy to be interfered by noise labels,the training time is long,and the classification accuracy is not high,a deep metric learning method based on the improved deep attention center loss(IDACL)was proposed.First,overfitting to underfitting-net(O2U-Net)was used to clean the sample data to reduce the influence of noise samples on model training.Then,the parameters of the O2U-Net model were transferred to the deep metric learning model,and the sample centers were extracted as the initial class centers of deep attention center loss.Finally,the weight was set according to the distance between the sample points and the class center to optimize the loss function and improve the classification accuracy of the model.Experimental results of surface defects of intermediate shell parts show that the proposed method has faster training speed and higher detection accuracy than other methods.关键词
表面缺陷检测/深度度量学习/深度注意中心损失/O2U-Net模型/机械零件Key words
surface defect inspection/deep metric learning/deep attention center loss/O2U-Net model/mechanical part分类
机械制造引用本文复制引用
李可,储世伟,顾杰斐,宿磊,薛志钢..基于IDACL深度度量学习的零件表面缺陷检测[J].华中科技大学学报（自然科学版）,2024,52(6):32-38,7.基金项目
国家自然科学基金资助项目(52175096). （52175096）
