U2S-CNN 3-phase defect detection system, from anomaly detection to defect classification

Abstract

Abstract The manuscript proposes the new 3-step universal defect detection system U2S-CNN tuned with visual data containing gear wheel images. The main advantage of the system is the detection capability of even unknown patterns of defects occurring in datasets. The object detection and defect detection approaches differ significantly in the basic principle. The precisely specified objects or patterns are sought in object detection and in the case of defect detection, patterns of different shape, orientation, color, character, etc. are sought. The problem of searching unknown objects is solved by defining the correct areas on the controlled object by using an asymmetric autoencoder of our own design. Subsequently, the differences between the original and autoenconder generated image are produced. The differences are divided into clusters using the DBSCAN method. Based on the clusters, regions of interest are defined, which are then classified using the pretrained Xception network classifier fined tuned with our data. The result is a 3-phase system capable of focusing even on unknown defects not occurring in the dataset using the sequence of Unsupervised learning – Unsupervised learning – Supervised learning methods. The proposed system is inspired by similarly designed systems used for the detection of anomalies or tumors in MRI or CT images, where U-networks or autoencoders are used. From the point of view of the nature of the issue, these problems can be considered very similar.