Research On Key Technology Of Road Surface Defect Detection Systems

At the same time of the rapid development of China’s road industry,road maintenance tasks are also increasingly heavy.China’s current road development has gradually transitioned from the construction phase to the maintenance phase.The maintenance of roads in China has become a arduous and urgent task for the stable development of China’s road industry.However,at present,the detection of road surface defects in China is still using manual detection method.This detection method is not only time-consuming and laborious,low in efficiency,subjective,and high missed detection rate also affects traffic.Therefore,the research on detection and identification technology of road surface defects is especially urgent.Therefore,this paper studies the road surface defect detection and recognition technology.At present,most methods for detecting and identifying road surface defects are traditional machine learning methods,the implementation of such methods is cumbersome and requires manual design to extract features,resulting in poor recognition and time consuming issues.Deep learning methods can integrate feature extraction and classification,realize automatic extraction of features,and has strong data representation ability.Therefore,this paper applies deep learning methods to road surface defect detection and recognition.This paper studies the road surface defect detection and recognition technology based on deep learning,and determines the overall design scheme of the road surface defect detection system.According to the requirements of road surface defect detection,the image acquisition module,Data set building module,defect recognition positioning and display module,preservation module and graphical user interface are completed.The main research contents are as follows:① In view of the use of video and continuous shooting in the road surface image acquisition in this paper,it is necessary to convert the video into an image.At the same time,the image size is too large and the hardware calculation pressure is too large.Based on this,the video framing and image blocking algorithm are designed.For the over-fitting problem,three kinds of data enhancement methods,such as mirror processing,rotation processing and adding a small amount of Gaussian noise,are designed to expand the training dataset in this paper.Finally,the experiments in this paper significantly relieved the memory pressure and did not show significantover-fitting problems.② Aiming at the problem of image classification and recognition of insufficient training dataset,this paper uses the migration learning method and model optimization algorithm to optimize the training of deep learning network model.The VGG-16,Inception-v3,and ResNet-50 pre-training models use fine-tune all layers of models and Convnet as the feature extractor methods to train and test the dataset in this paper,and compare the test results.In contrast,the VGG-16 model and fine-tune all layers of model combination method achieved a recognition accuracy of 98.9%.Based on this,the SGD,Adagrad,Adamelta,Momentum,Rmsprop and Adam model optimization algorithms are used to optimize the training process of the model,and the results are compared and analyzed.Finally,the VGG-16 model and the fine-tune model all layers and the SGD optimization algorithm combination method achieve the highest recognition accuracy of 99.1%,which proves the experimental effect of the proposed method in the road surface defect detection and recognition task.③ In response to the actual needs of road maintenance management,the graphical user interface of the road surface defect detection system was designed using Pyqt5.The system function,road surface defect identification and performance test of the road surface defect detection system prove the feasibility and use value of the system designed in the road surface defect detection.