基于反向知识蒸馏的工件异常检测算法研究  

Research on workpiece anomaly detection algorithm based oninverse knowledge distillation

作　　者：张晓涌 王黎明[1,2] 李璇 韩星程[1,2] ZHANG Xiaoyong;WANG Liming;LI Xuan;HAN Xingcheng(College of Information and Communication Engineering,North University of China,Taiyuan 030051,China;Shanxi Key Laboratory of Signal Capturing and Processing(North University of China),Taiyuan 030051,China)

机构地区：[1]中北大学信息与通信工程学院,太原030051 [2]信息探测与处理山西省重点实验室(中北大学),太原030051

出　　处：《激光杂志》2024年第8期92-97,共6页Laser Journal

基　　金：国家自然科学青年基金(No.62203405);山西省应用基础研究计划项目(No.20210302124545);动态测试技术国家重点实验室开放研究基金(No.2022-SYSJJ-08);山西省青年科技研究基金,基金号(No.201901D211250);2022年山西省应用基础研究计划项目(No.202203021212123)。

摘　　要：工件异常检测是工业生产中极其关键的一环,由于异常样本数量少,随机性大,有监督学习不能完全学习到所有的异常类型,存在模型稳定性差的问题,针对上述问题,研究了一种基于反向知识蒸馏的无监督工件异常检测算法,利用ResNet网络结构设计的教师模型和学生模型作为主干网络,教师模型真实地提取图像特征,学生模型根据先验知识重构图像,采取逆向结构扩大异常状况的特异性;中间加入记忆模块和掩码注意力模块,提取出样本的多维特征信息,避免遗漏图像中的细节信息;记忆模块之后添加的掩码注意力机制,将图像的多维度、多层次特征整合起来,进一步提升了检测的精确度。在两个公开工业异常检测数据集上进行实验的结果表明,所提算法能有效地定位细小异常,且相比普通知识蒸馏算法AUC提升了5%~7%。Workpiece anomaly detection is a key link in production.Due to the small number of abnormal samples and large randomness,supervised learning can not fully learn all types of anomalies,and there exists the problem of poor model stability.In order to solve the above problems,this paper studies an unsupervised workpiece anomaly detection algorithm based on reverse knowledge distillation,and uses the teacher model and student model designed by ResNet network structure as the backbone network.The teacher model truly extracts the image features,the student model reconstructs the image according to the prior knowledge,and adopts the reverse structure to expand the specificity of the abnormal condition.A memory module and a mask attention module are added to extract the multi-dimensional feature information of the sample to avoid missing the details of the image;after the memory module,a mask attention mechanism is added to integrate the multi-dimensional and multi-level features of the image.the accuracy of detection is further improved.The experimental results on two open industrial anomaly detection data sets show that the proposed algorithm is 5%higher than the general knowledge distillation algorithm AUC by 7%,and the effect of locating subtle anomalies is better.

关 键 词：异常检测 知识蒸馏 注意力机制 记忆模块 深度学习 机器视觉。 

分 类 号：TN209[电子电信—物理电子学]