印刷电路板缺陷持续检测与定位方法研究  

Research on continual detection and localization method for printed circuit board defect

作　　者：杨奡飞 续欣莹 谢刚 刘华平[2] YANG Aofei;XU Xinying;XIE Gang;LIU Huaping(College of Electrical and Power Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China)

机构地区：[1]太原理工大学电气与动力工程学院,山西太原030024 [2]清华大学计算机科学与技术系,北京100084

出　　处：《智能系统学报》2025年第1期219-229,共11页CAAI Transactions on Intelligent Systems

基　　金：山西省回国留学人员科研项目(2021-046);山西省自然科学基金项目(202103021224056);山西省科技合作交流专项(202104041101030).

摘　　要：针对目前缺陷检测与定位方法只能对特定类型的缺陷进行检测,而不能连续地学习检测不同类型缺陷的问题,提出了一种基于反向蒸馏模型的缺陷检测与定位方法。该方法以反向蒸馏模型为基础模型,对模型中间层输出的特征图以及一分类嵌入表示进行池化蒸馏,使得模型能够在连续的任务序列上不断地学习新的检测任务,从而达到持续学习的能力。在4个连续的印刷电路板(printed circuit board,PCB)缺陷检测与定位任务上进行实验,实验结果表明该方法的性能优于对比方法,能够满足工业生产场景的应用需求,在抑制对旧任务样本的检测能力的遗忘的同时,能够保持学习检测新任务的能力。Existing defect detection and localization methods can only detect fixed types of defects and cannot meet the continual defect detection requirements in real application scenarios.To address this issue,this paper proposes a defect detection and localization method based on the reverse distillation model.This method uses the reverse distillation model as the basis model and performs pooling distillation on the feature maps from the middle layers of the model and the one-class classification embedding representation.So that the model can continually train new detection tasks without forgetting previous tasks.Experimental results on four printed circuit board defect detection and localization tasks show that this method can meet the requirements of industrial applications,and it outperforms other methods.It maintains the ability to learn and detect new tasks while suppressing the trend of forgetting the ability to detect samples of previous tasks.

关 键 词：缺陷检测 缺陷定位 持续学习 深度学习 无监督学习 反向蒸馏 一分类嵌入 池化蒸馏 印刷电路板 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]