Research And Application Of Road Disease Detection Model Improvement Method Based On YOLOv5

The total mileage of highways in China is 5.28 million kilometers.In many provinces,the road disease detection is carried out through the way that maintenance personnel observe the road conditions,and the workload of road disease detection is huge when the total mileage is so high.In addition,manual detection has many defects,such as high cost,low accuracy and traffic impact.With the maturity of depth learning,related algorithms have also been applied in the field of pavement disease detection.However,due to the high cost of road detection equipment and the instability of road network environment,many depth detection algorithm models cannot be widely used.And most of the target detection algorithms are not universal,and can not give consideration to both computational complexity and accuracy.In order to meet the requirements of both speed and accuracy in the practical application of road detection model,this thesis designs two road disease detection models,namely mobile terminal and server terminal.Complete a detection on the mobile terminal,filter out the road images without diseases,and upload the diseased and pending images to the server,reducing the amount of data transmission and server detection.After receiving the image,the server detects the pending image twice,which reduces the pressure on the server and the cost of centralized detection.It not only meets the performance requirements of real-time and accuracy,but also integrates the capabilities of edge computing and centralized computing.This thesis mainly completes the following work:(1)The road disease image dataset is established,and the road disease detection comparison experiment is conducted on three target detection models,YOLOv5,Center Net and Efficient Det,and the best performance YOLOv5 is selected as the benchmark model of this thesis.At the same time,the road image is preprocessed.The main methods include gray processing,image denoising,image enhancement.(2)Based on YOLOv5 model,the road disease detection models of mobile end and server end are designed respectively,and the models at both ends are optimized respectively.By replacing C3 module with CSP1＿X module,adding PANet for feature fusion,expanding input size,adding CA attention mechanism,data enhancement and other means to optimize the server model.The client model is optimized by using Mobile Net V3 to replace CSPParknet53,reduce Io U to improve recall,add CA attention mechanism,and enhance data.Finally,the effectiveness of the corresponding optimization methods of the two end models is verified through comparative experiments.(3)The road disease detection system is designed and implemented to apply the improved algorithm,and the detection effect diagram of UAV and road patrol vehicle and the overall test results of the system are given.The comparative analysis experiment shows that the detection accuracy of the service side road disease detection model proposed in this paper reaches 84.2%,which is 18.1% higher than YOLOv5 benchmark model.The detection accuracy of the mobile terminal road disease detection model proposed in this paper reaches 77.1%,which is11% higher than YOLOv5 benchmark model,and the detection speed is also greatly improved.The test results on the system show that the final detection accuracy,accuracy and missing rate of the road disease detection system implemented in this paper meet the industrial production standards,but the false detection rate needs to be further reduced.