基于分布建模的磁瓦表面缺陷检测  

Distribution Modeling Based Surface Defect Detection for Magnetic Tiles

作　　者：郭宏[1] 李欣伦 张德华 畅晨吕 焦士轩 GUO Hong;LI Xinlun;ZHANG Dehua;CHANG Chenlyu;JIAO Shixuan(School of Mechanical Engineering,Taiyuan University of Science and Technology,Taiyuan 030024,China;Shanxi Pingyang Heavy Industry Machinery Co.,Ltd.,Linfen 043003,China)

机构地区：[1]太原科技大学机械工程学院,太原030024 [2]山西平阳重工机械有限责任公司,临汾043003

出　　处：《组合机床与自动化加工技术》2025年第4期113-117,共5页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：山西省重点研发项目(202102150401009)。

摘　　要：传统视觉检测方法对不同光照条件下的磁瓦表面缺陷检测精度较低,针对这一问题,提出一种基于分布建模的算法用于分割磁瓦表面缺陷。首先,算法模型使用预训练模型与蒸馏学习训练策略,实现学生网络对磁瓦数据的特征提取;其次,在学生网络中插入CA注意力模块和自注意力模块,以增强特征对正常数据的表述;最后,提出一种自适应特征融合,改善特征数据结构,增加分布的泛化性。实验结果表明,改进后的算法在磁瓦数据集上的像素级AUROC为92.8%、PRO为87.1%,均优于RIAD、PaDiM等缺陷分割算法。Traditional visual inspection methods have low accuracy in detecting surface defects of magnetic tiles under different lighting conditions,to address this problem,a distribution modeling based algorithm is proposed for segmenting surface defects of magnetic tiles.Firstly,the algorithm model uses a pre-training model with a distillation learning training strategy to achieve feature extraction from magnetic tile data by the student network;secondly,a CA attention module and a self-attention module are inserted into the student network to enhance the representation of features on normal data;finally,an adaptive feature fusion is proposed to improve the feature data structure and increase the generalization of the distribution.The experimental results show that the improved algorithm has a pixel-level AUROC of 92.8%and a PRO of 87.1%on the magnetic tile dataset,which are better than the defect segmentation algorithms such as RIAD and PaDiM.

关 键 词：磁瓦 缺陷检测 分布建模 注意力机制 

分 类 号：TH161[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]