Research On Anomaly Detection Method Based On Hierarchical Reconstruction Network

With the development of machine learning,anomaly detection has been successfully applied to medical anomaly detection,intrusion detection,food supervision and other fields.However,as new kinds of anomalies are easy to appear in the industrial assembly line,and the cost of manual identification and labeling is too large,we hope to have a more effective unsupervised anomaly detection method.Existing unsupervised anomaly detection mostly uses a single model,which cannot achieve stable performance for all categories.In view of the above problems,this paper carries out research from the perspectives of network structure,feature fusion and anomaly construction.The main work is as follows:(1)An anomaly detection method based on hierarchical reconstruction is proposed,and the feature fusion module designed completes the fusion from high-level features to low-level features to improve the quality of feature reconstruction.We completed the reconstruction of the two features at the same time in the training stage through different categories responding to different features at the top and bottom,so that each category could independently choose the most suitable feature as the reconstruction result in the testing stage,so as to improve the indicators of anomaly detection and anomaly segmentation.In addition,we explore the most suitable fusion form for our model through experimental comparison,so as to further improve the reconstruction effect of underlying features.Experimental results also verify that texture categories respond better to high-level features containing more detailed information,while object categories tend to favor low-level features(2)Based on SKNET,a feature fusion module that can dynamically generate channel weights for the two features is proposed.Considering the negative impact of the feature fusion module from high-level to low-level on the reconstruction of high-level features,SSIM loss is introduced as the additional supervision of high-level features.In the feature fusion module,the high-level features are further extracted by multi-scale,so as to achieve better fusion with the low-level features.The feature fusion under different parameter settings is compared and the optimal settings are selected for the model.(3)For the existing cutpaste-scar anomaly construction method,an improved method guided by attention diagram is proposed,and a corresponding classifier is designed for our autoencoder structure to help the reconstruction network to accomplish its task better.In the training stage,we set the reconstruction targets of both the original image and the augmented image as normal samples,expecting that the autoencoder could better learn the features of the normal samples.Meanwhile,we combined the reconstructed features with the original features as the input of the classifier,so as to improve the overall performance of the network.Compared with the original method,our model greatly improves the anomaly segmentation and detection indexes on MVTec,and at the same time achieves the current best results in the field of anomaly detection using exception construction.