Research On Anomaly Detection Algorithm Of Industrial Image Based On Deep Learning

Abnormal samples refer to samples that significantly deviate from normal conditions,and anomaly detection refers to the process of detecting abnormal samples from all data samples.Anomaly detection technology can be used in various fields,and has broad application prospects in network intrusion detection,fraud detection and industrial anomaly detection.Especially in the field of industrial image anomaly detection,abnormal samples often contain abnormal information with potential safety hazards,which often lead to greater economic losses and life threats.Therefore,anomaly detection algorithm in industrial scenes has higher research value.At present,there are two cases of anomaly in the industrial field: one is the defect fault of the instance-level sample,and the other is the intrusion of the unknown category sample.In order to cope with the challenges of the above two abnormal situations,this thesis focuses on unsupervised anomaly detection methods and researches on instance-level anomaly detection and unknown category anomaly detection.The main work includes the following two aspects:(1)For the instance-level anomaly detection task,this thesis proposes an instance-level image anomaly detection algorithm based on discrete-continuous feature coupling.The algorithm first transforms the latent features into continuous features and discrete features,namely block descriptive features and block hash features.Hash feature has the binarization characteristics,which can avoid under-sampling of latent space,thereby the problem of traditional reconstruction can be effectively solved.Based on the coupling relationship of discrete-continuous features,the graph shrinkage method is used to establish the block similarity matrix which constructs the association between hash features and description features.Then the inter-block reconstruction method is proposed,which can ensure the high-quality reconstruction of the image and solve the problem of low-quality interference.Extensive experiments conducted on the MVTec AD dataset show that the proposed method can effectively detect defective and faulty instance samples,which is superior to the current instance-level anomaly detection methods in index accuracy.(2)For the unknown category anomaly detection task,this thesis proposes an unknown category image anomaly detection algorithm based on category pattern mining.The algorithm first uses a clustering algorithm to explore sample correlation and proposes pseudo-label loss to learn the category information of the sample.Then the algorithm proposes a self-information mining module to learn the data pattern of samples of known categories.In addition,this thesis further proposes a three-tuple information learning strategy,which expands the inter-class distance by reducing the mutual information between negative sample pairs while tightens the intra-class distance by increasing the mutual information between positive sample pairs.This process jointly optimizes the learning of category information and the learning of data patterns of known categories.Validation experiments are carried out on the CIFAR-10 dataset.The experimental results show that the algorithm in this thesis can effectively detect samples of unknown categories and has a better effect on preventing unknown category intrusions,which is better than the current unknown category anomaly detection methods in index accuracy.