Research on Tire Defect Detection Based on Improved YOLOv5

An improved YOLOv5s-based method is proposed to address the challenge of accurately recognizing tire defects in this paper. In detecting process, the accuracy of YOLOv5s is low because the up-sampling module of the original YOLOv5s algorithm causes the network to lose information due to insufficient consideration of the distance between pixels. In this method, the up-sample operator of YOLOv5s is replaced by the carafe up-sampling operator. By adding the up-sampling kernel and the dot product operation of adjacent pixels in the input feature map, the spatial perception of up-sampling is strengthened, and finally the performance of segmentation task and the accuracy of image processing task are improved. The experimental verification is carried out on a complex datasets including four kinds of tire defects. Compared with the original YOLOv5 algorithm, the recognition accuracy is significantly improved. Specifically, under real-time performance constraints, the proposed method improves the recognition accuracy of mAPO.5 dataset by 2.9%. The results indicate that the proposed method enables accurate recognition of tire defects in complex backgrounds. Furthermore, it demonstrates better recognition precision than YOLOv5s.