TireNet: A high recall rate method for practical application of tire defect type classification

Abstract

Tire enterprises often use X-ray photographs to inspect tire defects, which is still in the stage of manual inspection. Existing methods have achieved a lot of success in tire defect detection, but have not yet reached the application level requirements. It is hard for unsupervised methods to handle the complex specs and patterns of tires, while the supervised methods suffered from a lack of enough data. In this paper, an end-to-end method (TireNet) is proposed for the practical application of automatic tire defect detection with X-ray images. Object detection models are introduced as the baselines, it is found that the recall rate (missed detection rate) cannot meet the application-level needs of enterprises, since tire defects are quite different from general objects. Inspired by the features of tire X-ray images, which are periodic, the Siamese network is used in the new model as part of the downstream classifier to capture the defective features. Additionally, 120,000 labeled tire images (including qualified tires) were set up and collected as datasets. The experimental results of this method reach a higher recall rate compared to Faster R-CNN, SSD, and YOLO in the labeled datasets. At present, the model has been successfully deployed to the quality inspection department of enterprises and partly replaced artificial jobs. The application results show that the missing rate of TireNet is 0.17%, which is lower than that of the artificial 2.4%. With the help of a deployed assistant detection system, the tire defect detection rate increased by more than 100%, and the tire refund rate decreased by approximately 20%.