Research On Road Defect Detection Algorithm Based On YOLOv5

With the rapid increase of highway mileage in our country,in order to ensure the smooth and safe traffic,the detection of road defects is more and more high.Manual testing is not only inefficient and costly,but also unable to meet the growing demand.Therefore,the research and development of automatic road defect detection device is of great significance to improve the efficiency of road defect detection.For detection devices,algorithm is the core.Road detection algorithm based on deep learning has received great attention at present.Existing network models generally have the problems of large model and large number of parameters,which should not be deployed on mobile terminal devices with low computing power.In order to solve the lightweight problem of the network,based on the YOLOv5 l detection network,the backbone network and neck network of the YOLOv5 l detection network were improved.Under the premise of not reducing the detection effect and convergence speed,the number of model parameters was significantly reduced,which could effectively improve the real-time performance of the algorithm and reduce the computing power requirements of the device.Specific work is as follows:(1)In view of the complex structure,insufficient reasoning speed and large network model of YOLOv5 l network,the convolutional module and C3 module of YOLOv5 l backbone network were replaced by Ghost Net module,and the G-YOLOv5 l detection network was proposed.It can obtain the characteristic information obtained by convolution with YOLOv5 l model with fewer parameters,thus realizing the lightweight of the model.In addition,in order to improve the feature extraction capability of G-YOLOv5 l network,CA attention mechanism is integrated into the feature extraction network,so as to focus on the feature map channels useful for the current task.The experimental results of suppressing the feature channels that have little effect on the current task show that under the condition that the detection effect and convergence speed are comparable to that of the YOLOv5 l network,The model parameters are greatly reduced.(2)In order to further improve the accuracy of lightweight target detection algorithm GYOLOv5 l network,a B-YOLOv5 l algorithm is proposed.The BiFPN is used to replace the original feature pyramid network(FPN)to improve the structure of the neck network of GYOLOv5l and improve the feature fusion capability of the model.Meanwhile,Transformer and attention mechanism are added respectively at the head and tail of the neck network to make the network pay more attention to the area of interest.Thus,the detection performance of the network model with low resolution and small targets can be improved.The new algorithm can ensure the lightweight and improve the detection accuracy of the network.Through training and testing on the common data set RDD-2020,the number of model parameters of the improved algorithm decreases significantly when the detection accuracy and convergence speed are the same,which can effectively reduce the requirement of equipment computing power.