基于小样本学习的磁瓦表面缺陷检测  

Surface Defect Detection on Magnetic Tile Based on Few-shot Learning

作　　者：王春泽 林容泰 朱拾东 赵艺 于赐龙 Wang Chunze

机构地区：[1]联通数字科技有限公司5G应用创新中心,北京100032 [2]深圳大学机电与控制工程学院,广东深圳518000 [3]深圳青鼎装备科技有限公司研发中心,广东深圳518000 [4]哈尔滨工业大学(深圳)教务处,广东深圳518000 [5]华东理工常熟研究院有限公司研发部,江苏常熟215500

出　　处：《工业控制计算机》2025年第4期12-13,16,共3页Industrial Control Computer

基　　金：江苏省卓越博士后计划项目(2023ZB009);中国联通2023年核心技术攻关项目(Z9223000JS4339)。

摘　　要：磁瓦片是永磁电动机的核心组件,其质量对电动机的性能和使用寿命至关重要。针对传统缺陷检测方法受限于缺陷样本少、缺陷细小、背景复杂,难以达到较高精度和效率的问题,提出一种基于小样本学习的缺陷检测方法。该方法采用正负样本采样策略,结合对比学习技术,突出目标特征的类内相似度和类间差异性,通过设计对比相似度损失函数监督训练,有效解决小样本场景下两阶段网络的候选框置信度低和目标特征表示不足的问题。所提出的方法在不同厂家所生产的磁瓦片样本进行验证,实验结果表明,该方法可对磁瓦片的崩缺、裂纹、多层面、倒角异常、污渍5类常见缺陷进行快速、准确的检测和分类。In response to the problem that traditional magnetic tile defect detection methods are limited by few defect samples,small defects,and complex backgrounds,making it difficult to achieve higher accuracy and efficiency.This paper proposes a defect detection method based on few-shot learning.This method adopts a positive and negative sample sampling strategy,combined with contrastive learning technology,to highlight the intra-class similarity and inter-class difference of target features.By designing a contrastive similarity loss function for supervised training,it effectively addressed the low confidence of candidate box in two-stage networks and inadequate representation of target features in small sample scenarios.The proposed method is verified on magnetic tile samples produced by different manufacturers.The experimental results show that this method enables rapid,accurate detection and classification of five common defects on magnetic tiles:chipping,cracks,multi-layering,abnormal chamfering,and stains.

关 键 词：磁瓦片 小样本 正负采样 对比学习 表面缺陷检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]