Detection of Asphalt Pavement Cracks Based on Vision Transformer Improved YOLO V5
Sike Wang, Xueqin Chen, Qiao Dong
Abstract
Automatic and rapid detection of pavement cracks is one of the important tasks for the highway department. This study proposed an improved model of you only look once (YOLO) V5 integrated with the vision transformer (ViT) that can calculate the attention weights of image regions and form a new feature map with weights. The ViT module was added to the neck of YOLO V5 to improve the speed and accuracy of the model. 1944 asphalt pavement images were collected for testing. The test results showed that the proposed model obtained high accuracy and speed for longitudinal, transverse, and fatigue cracks and was capable of real-time detection. The ViT-improved YOLO V5m obtained 0.872 in mAP(0.5), and the detection time for a single image is 11.9 ms. The study also investigated the pavement crack detection performance of the models in a rainfall environment. All models did not have satisfying detection capabilities in rainfall conditions, but the developed YOLO V5 had better performance.