基于残差池化自编码机的无监督异常分割  

Unsupervised anomaly segmentation based on residual max-pooling autoencoder

作　　者：杨胜雄 陈莹 YANG Shengxiong;CHEN Ying(Key Laboratory of Advanced Process Control for Light Industry,Ministry of Education,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学轻工过程先进控制教育部重点实验室,江苏无锡214122

出　　处：《计算机集成制造系统》2024年第6期2069-2079,共11页Computer Integrated Manufacturing Systems

基　　金：国家自然科学基金资助项目(61573168,62173160)。

摘　　要：为了解决异常缺陷面积大小不一及背景干扰问题,基于残差池化自编码机提出一种无监督异常分割模型,检测和分割物体表面的异常缺陷。所提模型设计并使用残差池化模块,增加传统模型对异常的逆向重建能力,使正常和异常之间的分水岭更加明显,解决了大面积缺陷分割不完全的问题;在异常评分阶段引入高斯平滑函数,使模型具有鲁棒性,减小了背景对模型的干扰。在公开的MVTEC AD数据集上进行验证,结果表明,所提模型在图片级别检测精度上达到95.6%、像素级别检测精度上达到96.5%、区域级别的检测精度上达到91.7%,证明了所提模型的有效性,将该模型与其他异常分割方法进行比较,证明了模型的优越性。To solve the problem of different defect areas and interference of background,an unsupervised anomaly segmentation model was proposed based on residual max-pooling autoencoder machine to detect and segment abnormal defects on object surface.The residual max-pooling module was designed and used in the proposed model to solve the problem of incomplete segmentation of large area defects,resulting in the ability of the traditional model to reverse the anomaly reconstruction was enhanced,and the watershed between normal and abnormal was more obvious.Gaussian smoothing function was introduced in the abnormal scoring stage,resulting in the model robust was enhanced and the interference of background to the model was reducing.On the MVTEC AD data set of simulated industry,the detection accuracy of image level was 95.6%,pixel level was 96.5% and region level was 91.7%,which proved the validity of the proposed model.By comparing with other anomaly segmentation methods,the superiority of the proposed model was validated.

关 键 词：异常分割 无监督 残差池化模块 高斯平滑 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TH74[自动化与计算机技术—计算机科学与技术]