Abnormity detection based on fusion feature distribution learning and image reconstruction

Unsupervised learning is the main research direction in the field of industrial product defect detection at present,and it is mainly divided into two types of methods: reconstruction based and feature based methods. The former constructs content-aware mappings to map abnormal regions into normal regions and detect defects through residual images,focusing on overall performance of the images. The latter uses high-level semantic features to achieve positioning exceptions and pay more attention to image detail presentation. According to the advantages and disadvantages of the two methods,a method is proposed based on the fusion of characteristics and reconstruction, which effectively combines advantages of the two methods to complement their shortcomings and realize unified end-to-end learning and reasoning. A reconstructed model is trained firstly,then a normalized flow model is adopted to fully learn the probability distribution of high probability data of input normal samples, and it is integrated with the reconstructed model to effectively improve the accuracy of defect detection and defect positioning of the reconstructed model. On the widely used MVTec AD data set, the average image level AUROC of the proposed fusion model reaches 98. 7 degrees o, the average pixel-level AUROC reaches 94. 2 degrees o,in particular,an increase of 3. 3 degrees o compared to a single reconstruction model. The convergence model of characteristics and proposed reconstruction network has significantly improved the shortcomings of defect positioning in the reconstruction network part, which makes experimental results more accurate.