Asphalt Pavement Distress Detection Based On Deep Neural Network

With the rapid development of China’s national economy,especially the transportation industry,highway mileage has continued to increase.As of the end of 2018,the total mileage of national highways has reached 4,486,500 kilometers,of which a large number of backbone highways are about to enter the overhaul and maintenance period,and road disease detection is one of the primary goals of highway maintenance procedures.Considering the large total mileage of China’s highways and the uneven population and economic development in various regions,it is difficult to comprehensively use manual visual inspection or high-precision professional equipment for road evaluation.In order to solve the above problems at a lower cost,based on the 2D pavement image data captured by digital cameras and the excellent Single Shot Multi Box Detector algorithm in the field of object detection,a neural network model was designed and implemented to detect several common diseases on the road surface,including Horizontal / longitudinal fissures,reticular fissures,road potholes / bulges,blurred road markings,etc.By analyzing the causes of detection errors in the existing research,three optimization strategies are proposed for the target detection algorithm.First,Cycle GAN is used to expand the image of road disease types with fewer samples in the training data set to solve the imbalanced data set categories.Secondly,using training,prediction,and re-annotation methods to improve the quality of data annotation for some inaccurate or ambiguous images in the training data set.Finally,introduce Focal Loss as the calculation of category loss in SSD object detection algorithm for improving the recall rate of the pavement disease detection model.The model was tested using the public Road Damage Dataset,and the validity of the method was verified.Experiments show that the proposed object detection model has better performance than other models,and especially solves the problem of extremely low recall of some categories due to insufficient training samples.With no serious loss in the average accuracy rate of each category of targets,the average recall rate increased from 0.48 to 0.59.