Unsupervised defect detection and positioning method based on multilevel feature reconstruction

Abstract

The invention relates to an unsupervised defect detection and positioning method based on multilevel feature reconstruction. The method comprises the following steps: acquiring a non-abnormal image of a product and abnormal images of different types of defects; extracting a multi-scale feature group by using a non-abnormal image input feature extraction network; inputting the feature of the highest dimension into a reconstruction network to reconstruct new feature groups corresponding to different scales layer by layer; constructing a loss function training reconstruction network for the two feature groups; and in the test stage, an anomaly graph and an anomaly score are calculated according to the difference condition of the two feature groups and used for judging anomaly and positioning a defect area. According to the method, the feature information difference of the non-abnormal image and the abnormal image in different dimensions is effectively utilized, and the defect area of the product can be detected, so that manual labeling is avoided, and the product quality detection efficiency is improved.

Description

Technical field [0001] The present invention relates to the field of computer visual, and more particularly to a non-supervised defect detection and positioning method based on multi-level feature reconstruction. Background technique [0002] Defect detection has been wide and important in industrial production. In industrial production, it is possible to effectively guarantee the good product of the product in industrial production. In practical scenes, due to the characteristics of defective samples, the number is very rare relative to the number of normal samples, and the type of defect has uncertainty, the size of the appearance, the position is random, so it is not possible to predict in advance These defects are characterized. In addition, under the actual industrial production scenario, there is a certain requirement for the real time of the defect detection algorithm. [0003] With the continuous development of deep learning in the field of target detection, defect detect...

AI Technical Summary

Problems solved by technology

In order to locate the defect area, constructing an anomaly map of the size of the input image can intuitively reflect the prediction results of the network model for image defects, but directly reconstructing the normal image may face the problem of too strong generalization ability, thus re- Construct defect images, in addition, image block-level positioning does not meet the needs of locating defect areas of different scales

Images