Research On Pavement Damage Detection Model Based On Deep Learning

In recent years,China’s road construction still maintains an upward trend,China is the world’s longest road mileage of the country,road traffic greatly facilitates the travel of the people,and to maintain a good road condition needs to do a good job of road maintenance work,how to save manpower and material resources in the case of road damage detection and positioning is one of the key issues.With the gradual development of related technologies,the shortcomings of traditional road detection methods have slowly emerged.In recent years,deep learning technology and computer vision technology have developed rapidly and have been applied in various fields,which can greatly improve the processing efficiency of image tasks,and for the problem of pavement damage detection,this paper proposes a pavement damage detection model based on the relevant theoretical knowledge of deep learning.The main research contents are as follows:(1)It studies the current situation of pavement damage detection at home and abroad,analyzes the various types and characteristics of pavement damage,and integrates the lightweight convolutional neural network Mobile Netv3 into the YOLOv3 network based on the image characteristics of the current pavement damage,replaces the standard convolution,realizes a lightweight network model,retains the multi-scale feature fusion structure,reduces the amount of parameters compared with the original model,and accelerates the detection speed of the model.(2)In view of the YOLOv3-Mobile Netv3 network model,the disadvantages of the loss function of the original model are analyzed.In order to improve the efficiency of pavement damage detection,it is proposed to optimize the loss function to combine the focus loss with the DIOU loss,which improves the model accuracy without introducing too many parameters.In addition,in view of the insensitivity to small-scale targets in the road damage detection task,the feature fusion structure of the network is improved,and the small-scale fusion structure is deleted,which further improves the calculation speed of the model.(3)In view of the lightweight of the model,a method of combining BN layer channel pruning and interlayer pruning with scaling factors is proposed,which not only avoids the gradient descent problem of the network in the training process,but also retains the role of the scaling factor,and completes the compression of the network model without increasing the amount of computation.The amount of weight parameters of the network model is greatly reduced,and excessive loss of accuracy of the model is avoided.Finally,the training test of the model was completed with the pavement damage dataset made in this paper,and then several comparative experiments were carried out to verify the effect of optimization of each part.After comprehensive comparison between the model proposed in this paper and the original model,it shows that the model proposed and optimized in this paper has a smaller volume and faster computing speed,reduces the hardware requirements of the algorithm,and has better adaptability to small computing devices.