Research On Image Anomaly Detection Method

Image anomaly detection is widely used in industry,agriculture,medicine,security and other fields,and has important application value.The goal of image anomaly detection is to build an anomaly detection model,and then discover abnormal image samples that deviate from the normal pattern.Traditional one-class classifier is widely used in various anomaly detection situations,but its anomaly detection performance is limited by the representation of input data.Therefore,this thesis studies how to improve the anomaly detection performance of traditional one-class classifier as follows:(1)Considering that the performance of traditional one-class classifier is limited by the representation of input data,this thesis proposes an image anomaly detection framework based on the existing anomaly detection methods combining autoencoder and one-class support vector machine,which is one-class classifier based on single-level image features.This framework includes two anomaly detection methods: one-class support vector machine based on single-level image features(SFOCSVM),and kernel density estimation based on single-level image features(SFKDE).The main feature of this framework is that the compressed representations(bottleneck features)of image samples are extracted based on convolutional autoencoder.The framework extracts compressed representations(bottleneck features)of image samples based on convolutional autoencoder,which provides high-quality data representations for the training of one-class classifiers,thereby effectively improving the anomaly detection performance of one-class classifiers.(2)This thesis improves the one-class classifier based on single-level image features,and proposes a new anomaly detection framework,which is an ensemble one-class classifier based on multi-level image semantic features.This framework includes two anomaly detection methods: one-class support vector machine based on multi-level image features(MFOCSVM),and kernel density estimation based on multi-level image features(MFKDE).Different from the one-class classifier based on single-level image features,the ensemble one-class classifier based on multi-level image semantic features uses convolutional autoencoder to extract different levels of image semantic features,then train multiple base classifiers,and finally make the final decision by voting.The experimental results on image anomaly detection benchmark datasets(MNIST,Fashion-MNIST,MVTec AD)show that the one-class classifier based on single-level image features effectively improves the anomaly detection performance of traditional one-class classifiers.The ensemble one-class classifier based on multi-level image semantic features effectively improves the anomaly detection performance of one-class classifiers based on single-level image features,and the two anomaly detection frameworks outperform the comparison methods.