首页|基于YOLOX的类增量印刷电路板缺陷检测方法

基于YOLOX的类增量印刷电路板缺陷检测方法

扫码查看
为了应对更加实际的增量式印刷电路板缺陷检测场景,本文将知识蒸馏与YOLOX相结合,提出了一种基于YOLOX的类增量印刷电路板缺陷检测方法.在只使用新训练数据的情况下,模型能够检测出所有学过的缺陷类型.通过对模型的输出特征和中间特征使用知识蒸馏来促进旧缺陷类别知识的传递,使得学生模型能够有效保留教师模型在旧缺陷类别上的检测性能.实验结果表明,本文方法能够显著缓解增量学习过程中的灾难性遗忘问题,在两阶段增量场景下,模型对所有缺陷的平均检测精度为 88.5%,参数量为 25.3×106,检测速度为 39.8 f/s,便于工业设备部署的同时,可以满足增量式检测场景下印刷电路板(printed circuit board,PCB)质检的检测精度和检测速度要求.
Class-incremental printed circuit board defect detection method based on YOLOX
To cope with more practical incremental printed circuit board detection scenarios,by combining the know-ledge distillation with the YOLOX,this paper proposes a class-incremental Printed Circuit Board defect detection meth-od based on YOLOX.The model can detect all learned defect types when only new training data is used.The transfer of knowledge about old defect categories is facilitated by using knowledge distillation for the model's output features and intermediate features,enabling the student model to effectively retain the detection performance of the teacher model on old defect categories.The experimental results show that the method in this paper can significantly alleviate the cata-strophic forgetting problem during the incremental learning process.Under the two-stage incremental scenario,the mod-el has a mean average precision of 88.5%for all defects,a parameter size of 25.3 M,and an inspection speed of 39.8 f/s,which facilitates the deployment of industrial equipment and at the same time,it can satisfy the detection accuracy of printed circuit board(PCB)quality inspection and the inspection speed requirement in incremental detection scenarios.

deep learningprinted circuit boardclass-incrementalincremental learningdefect detectionobject detec-tiondynamic detectionknowledge distillationcatastrophic forgetting

吴瑞林、葛泉波、刘华平

展开 >

南京信息工程大学 自动化学院,江苏 南京 210044

清华大学 计算机科学与技术系,北京 100084

深度学习 印刷电路板 类增量 增量学习 缺陷检测 目标检测 动态检测 知识蒸馏 灾难性遗忘

江苏高校"青蓝工程"项目

R2023Q07

2024

智能系统学报
中国人工智能学会 哈尔滨工程大学

智能系统学报

CSTPCD北大核心
影响因子:0.672
ISSN:1673-4785
年,卷(期):2024.19(4)
  • 10