Unsupervised anomaly segmentation based on residual max-pooling autoencoder
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 abnor-mal 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 obvi-ous.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 su-periority of the proposed model was validated.
NETL
NSTL
万方数据
