Pavement Damage Detection Based on Improved YOLOv5 Model

Pavement damage not only significantly impacts the load-bearing capacity of roads but also poses a serious threat to vehicle safety. The existing pavement recognition methods usually have the problem of missing detection and misdetection in some areas where the pavement damage characteristics are not obvious. In order to solve these problems, an improved network model algorithm called FDD-YOLO is proposed. In this method, firstly the FasterNet network was selected to replace the feature extraction network in the original YOLOv5s model to reduce the number of Giga Floating-point Operations Per Second(GFLOPs) and parameters of the model. Then, the lightweight CARAFE operator is utilized as the up-sampling operator, using its feature content sensing recombination algorithm to minimize feature loss during up-sampling. Finally, deformable convolution is used to enable more comprehensive extraction of image feature information by the model under larger receptive field conditions. Experimental results show that the proposed FDD-YOLO model can effectively identify road surface damage. Compared with other models, the GFLOPs of the model is reduced by 54.43%, the number of parameters is reduced by 51.99%, and the mean Average Precision is increased by 1.1%, which can meet the basic requirements of pavement damage detection.