Research On Pavement Defect Image Detection Method Based On Deep Learning

Pavement diseases not only affect the appearance of pavement and driving comfort,but also cause potential harm to road life and traffic safety.The detection of road surface defects is of great significance for pavement maintenance.Therefore,it is of great practical value and research significance to realize the rapid intelligent detection of pavement defects.Deep learning can be used to quickly identify the characteristics of road surface defects.In this thesis,the technology is used to complete the rapid intelligent detection of five types of defects: cracks,broken plates,potholes,repairs and pavement dotted line.The main research work is as follows:(1)According to the definition of pavement defects in " Highway Performance Assessment Standards ",four kinds of asphalt and concrete pavement damages are distinguished through the description of phenomenon,and the dividing line is regarded as misidentification defect.An image acquisition device was used to obtain photos and videos of pavement defects.After screening and labeling,D1,a dataset of pavement damages with2,600 images and 7070 defects,was obtained for training and verification of deep learning algorithm.In order to enhance the recognition ability of the model for dark brightness and motive-blurred images,data set D1 was extended to data set D2 by Python image enhancement algorithm,with a total of 6180 images and 16436 defects.(2)In order to realize rapid and intelligent detection of pavement defects,YOLOv5,a one-stage target detection algorithm with a fast detection speed in deep learning,was improved,its original Euclidean distance was replaced as sample distance to re-cluster this dataset anchor,and CBAM attention mechanism was introduced after the second,third and fourth convolution layers in the Head part.Thus,the detection accuracy of the algorithm is improved.The test results show that the average accuracy of the YOLOv5-A model on the D1 verification set is 92.7%,1.7% higher than that of the original model,and the detection speed is 0.025 seconds per image.The identification confidence of the YOLOv5-AC model on cracks and potholes is 89% and 90%,45% and 5% higher than that of the original model.In D2 dataset,the confidence for dark image detection is increased by 7%.The data show that the method achieves high detection accuracy and meets the requirement of real-time intelligent front-end detection of pavement damage.(3)To understand the neural network judgment process and algorithm feedback,the heat map visualization of the detection image was carried out by using the activation diagram CAM to explore the focus of the model and excavate more details beyond the final detection results.Based on Py Qt5,the detection effect of improved algorithms such as YOLOv5-AC on the photos and videos of pavement damage was displayed in the form of an interface,which was closer to the practical engineering application.