Research Of Measured Road Damage Image Classification And Recognition Based On Improved YOLOv5s

Road transportation is an important means of national transportation in China.However,with the aging of the road surface and external forces,the performance of the road surface gradually declines,and ultimately the resulting road damage affects traffic safety and smooth driving.In this context,the rapid and accurate detection and identification of road damage has become an important research topic in the field of traffic safety.In recent years,deep learning has been widely used in road damage detection.However,due to the complex road background and diverse types of damage,current mainstream detection models still face great challenges in accurately and quickly identifying damaged road surfaces.Therefore,designing a fast and accurate road damage classification and recognition model has great research value.This paper aims to study how to improve the speed and accuracy of road damage recognition.In view of the problem that current road damage detection models have difficulty in quickly and accurately identifying damaged road surfaces,an improved OS-RFB-YOLOv5 s model based on the YOLOv5 s algorithm is proposed.Firstly,a lightweight and efficient feature extraction network,OSNet,in the backbone network.OSNet consists of OS modules and SS modules,with the SS module adding a channel attention mechanism on the basis of Shuffle Net V2 to better weight network channels.The OS module uses a 3×3 ordinary convolution to replace the 1×1 ordinary convolution and 3×3 depth convolution in the main branch of the SS module,thus improving the training speed of the model in the shallow layers of the network.Secondly,the Bottleneck structure of the C3 module in the neck network is improved,and an RFB-C3 module based on RFB(Receptive Field Block)is proposed,which uses a multi-branch structure and atrous convolution with different receptive fields to extract features and obtain a larger receptive field with fewer parameters.Finally,the position loss function of YOLOv5 s is optimized by changing the optimization of the ratio of the long and short sides to the bias of the long and short side parameters,which can effectively avoid the problem that the long and short side parameters of the predicted box cannot increase or decrease simultaneously.To verify the performance of the improved OS-RFB-YOLOv5 s model,model evaluation indicators can be used to evaluate performance of the proposed model on a road damage measurement dataset.The experimental results show that,by improving the backbone network,the m AP values of the original YOLOv5 s,the control model SS-YOLOv5 s,and the improved model OSYOLOv5 s are 77.4%,80.4%,and 81.7%,respectively.The m AP value of the model RFBYOLOv5 s,which improves the neck network of YOLOv5 s,is 80.5%.After optimizing the loss function,the m AP value is 78.5%.The improved model proposed in this paper,OS-RFBYOLOv5 s,achieves an m AP of 83.3% and an FPS of 90.4,which is 5.9% and 15 higher than the original model YOLOv5 s.The above experimental results demonstrate that the proposed improved model effectively improves the real-time detection accuracy of road surface damage.To verify the generalization ability of the proposed model,conducted experiments are conducted on the public dataset RDD2022,and the results showed that the Precision,Recall,F1-score,m AP,and FPS of our model are 78.3%,52.2%,0.62,80.8%,and 86.2,respectively,indicating that the algorithm model proposed in this paper has good generalization ability.