| In recent years,with the completion of China’s expressway planning and construction targets and the continuous development of expressway traffic network,the maintenance of expressway has become increasingly important.How to improve the informatization,specialization and intelligence level of highway maintenance through technical means is one of the urgent problems to be solved in the field of highway maintenance.Expressway asphalt pavement distress detection and classification technology is an important basis to realize intelligent road maintenance.The difficulty of its work is that it is difficult to detect and classify the microcracks and various types of highway pavement distress.In view of this,this paper carried out research on expressway asphalt pavement dietress perception method based on two-dimensional linear array image,and the research results and innovations obtained are as follows:Firstly,FR-PDD,YOLOv5s-PDD and SSD-PDD network models for expressway asphalt pavement distress detection were constructed based on FR-RCNN,Yolov5 s and SSD convolutional neural networks.In addition,model training and comparative experiments were carried out.Experimental results show that the average detection and classification accuracy m AP of FR-PDD,YOLOv5s-PDD and SSD-PDD are all better than 92%.Among them,YOLOv5s-PDD network model has the best performance,and its average recognition accuracy m AP reaches 98.1%.Secondly,FCN-D121-PDS,FCN-D201-PDS,DL-D-PDS and DL-M-PDS are constructed based on FCN and Deeplabv3+ semantic segmentation neural networks for expressway asphalt pavement distress segmentation.The pixel-level manual annotation was carried out for the high highway asphalt pavement distress image data set,and the model training and comparison experiment were carried out.The experimental results show that the segmentation accuracy of DL-M-PDS network model is 98.0% and the average crossover ratio is 77.5% superior to the other three types.Two segmentation network models,DL-M1-PDS and DL-M2-PDS,were improved based on the optimization of DL-M-PDS network model.The comparative experimental results show that DL-M2-PDS segmentation model is superior to DL-M1-PDS,with segmentation accuracy of 98.3% and average crossover ratio of 78.3%.Finally,FCNN-PDP-FR,FCNN-PDP-YOLOv5 s and FCNN-PDP-SSD fusion network model construction methods are proposed for expressway asphalt pavement distress perception.DL-M2-PDS network model was used to segment the original pavement distress image to obtain the binary image,and the segmentation image was superimposed on the original image to build the high highway asphalt pavement distress data set for fusion model training.Three fusion network models,FCNN-PDP-FR,FCNN-PDP-YOLOv5 s and FCNN-PDP-SSD,are trained and optimized,and comparative experiments are carried out.Experimental results show that the average detection and classification accuracy m AP of the three fusion network perception models are better than that of the single network model.Among them,the FCNN-PDP-YOLOv5 s fusion perception model has better perception accuracy than 98% for horizontal crack,longitudinal crack,net crack,pit slot and other types of distress,and the average accuracy is 99.2%. |
