Research On Road Surface Defect Detection Based On Deep Learnin

Road surfaces are subjected to various degrees of defects such as cracks and potholes due to frequent use and the influence of natural or human factors.If these road defects are not repaired in time,they can affect the smoothness and safety of vehicle travel,and may even cause traffic accidents.Road defect detection technology can be used to timely identify road defects so that corresponding maintenance measures can be taken promptly.Traditional automated road defect detection technologies have disadvantages such as slow detection speed or accuracy,making it difficult to effectively detect road defects.With the outstanding performance of deep learning in the field of computer vision in recent years,the development of deep learning-based road defect detection technology has also been rapid.Nevertheless,there still exists the problem of low detection accuracy in road defect detection tasks,especially under complex road conditions where road defect shapes are complex and sizes vary.The difficulty of detection is further increased by natural weather conditions such as lighting,rain,and snow.Based on the analysis of relevant literature,this paper conducts indepth research on deep learning-based road defect detection algorithms,with the following main research contents:(1)In order to improve the detection accuracy of YOLOv5 in road defect detection,firstly,to address the problem of imbalanced positive and negative samples in road defects,the positive sample selection area of YOLOv5 is expanded to obtain more positive samples that can enhance the gradient during network training.Secondly,the new loss function SIo U is used in YOLOv5 to reduce the possibility of prediction box wandering during the training process,which improves the convergence speed of the network when training on the road surface dataset while improving accuracy.Finally,to further enhance the network’s ability to capture road defect features,CBAM attention mechanism is introduced,and the spatial information capturing capability in CBAM attention mechanism is strengthened by using three serial 5x5 convolution kernels to capture more road defect feature information,thereby further improving the detection accuracy of YOLOv5.(2)In response to the problem of poor crack detection in road defect detection,this research builds upon previous work(1)and uses larger convolution kernels.Larger convolution kernels have a larger receptive field which allows them to better capture detailed features such as the shape and edges of cracks.By using a large convolution kernel structure,the performance of the model is improved.Firstly,structural reparameterization is used in the large convolution kernel to improve the detection accuracy of the model while ensuring the speed of model detection.Secondly,depthwise separable convolution is used to reduce the number of network parameters caused by the large convolution kernels.Additionally,experiments were conducted on a self-built road defect detection dataset with other algorithms.This paper proposes a road defect detection model that has better robustness for road defect detection tasks.(3)The study of YOLOv5 in pavement defect detection has enabled the development of a pavement defect detection system,which mainly implements the detection of pavement defects in the browser and the storage of the detection results.This paper conducted a series of studies on YOLOv5,and the proposed improvements were validated through experiments on the RDD2020 dataset.Ultimately,the model proposed in this paper improved m AP by 3.2% relative to YOLOv5,and had relatively good robustness.Additionally,based on this research,a road defect detection system was implemented.