Research On Unsupervised Industrial Image Anomaly Detection Algorithm Based On Deep Learnin

Industrial anomaly detection is the classification and location of areas that deviate from the normal appearance.At present,the supervised anomaly detection algorithm based on deep learning has achieved great success,but there are still problems such as heavy workload of data labeling and insufficient abnormal samples.Therefore,unsupervised learning method is introduced into industrial detection to improve the effect of anomaly detection and location.Although the unsupervised anomaly detection algorithm made up for the shortcomings of the supervised anomaly detection algorithm,they still have the following limitations: 1)When solving anomaly detection tasks based on image inpainting model,convolutional neural network(CNN)is usually used to construct the inpainting network,which makes the model’s inpainting ability weak and leads to the low detection performance of the algorithm.2)When solving anomaly detection tasks based on knowledge distillation model,most teachers and students(T-S)network models are built with similar or the same symmetric structure,which makes the difference of abnormal expression between T-S networks not high,resulting in a high miss rate of the algorithm.In view of this,this paper studies the unsupervised industrial image anomaly detection algorithm based on deep learning,and the main contents are as follows:(1)Anomaly detection algorithm based on image inapinting: Aiming at the problem of insufficient inpainting ability of CNN’s image inpainting model,a hybrid model based on Ushaped Swin Transformer and CNN(MSTUnet)is proposed.Firstly,in order to solve the problem of insufficient abnormal samples in the training phase,the abnormal simulation and mask strategy were applied to the normal samples to generate simulated abnormal samples.Secondly,the powerful global learning ability of Swin Transformer is used to inpaint the mask area to achieve the purpose of improving the performance of anomaly detection.Finally,the convolution-based Unet network was used to realize end-to-end anomaly detection and localization.Experimental results on MVTec AD dataset show that MSTUnet achieves excellent anomaly detection and localization performance.(2)Anomaly detection algorithm based on knowledge distillation: Aiming at the problem that the symmetric knowledge distillation model has low difference in anomaly expression,a T-S network model based on asymmetric multi-scale feature matching(ATSN)is proposed.ATSN is composed of a teacher network and an encoder-decoder structure U-shaped student network.Firstly,a lightweight teacher network was used to guide the encoder to fine-grained learn the feature distribution of normal samples,so as to improve the correctness of knowledge transfer.Secondly,the decoder was used to reconstruct the deep feature information in the encoder,so as to increase the diversity of the T-S network for abnormal expression.Finally,the multi-scale feature fusion module(FFM)was used to fuse the multi-scale feature information in the encoder and decoder to further increase the difference of abnormal expression,thereby reducing the missed detection rate of the algorithm.Experiments on the MVTec AD dataset show that the ATSN algorithm outperforms the most advanced anomaly detection algorithm based on knowledge distillation,which proves the superiority of the detection performance of the algorithm.