Research and Comparison of Intelligent Detection Methods of Pavement Distress Based on Deep Data Augmentation

To solve the time-consuming problem of manual pavement distress detection and the possible low detection accuracy problem due to unbalanced sample dataset, a method of deep data augmentation was employed to enhance the dataset of high-definition road images taken by smartphones. The results after the data augmentation were evaluated and tested by using two different target detection algorithms. The main research contents of the paper included. First, considering the limitations of experimental conditions and acquisition environment, a deep data augmentation method was employed by combining WGAN-GP and Poisson transfer algorithm, which supplemented and balanced the training sample data by generating road pothole images under different lighting conditions. Then, multiple target detection algorithm variants of Yolo(Yolov5s, Yolov5m, Yolov5l, and Yolov5x) and Faster R-CNN algorithm were introduced, and the accuracy and efficiency of various target detection algorithms after applying the data augmentation were compared through experiments. Experimental results on the Japanese open road detection dataset show that the average improvement of P, R and F1 of five detection algorithms is 2.8%, 4.0% and 3.6%, respectively, after using the deep data augmentation method. Among the five detection algorithms, Yolov5l achieved the highest F1 value, reaching 60.9%. If conditions are suitable, such as in the test set with moderate light conditions, the F1 value of Yolov5l algorithm can reach 68.7%.