Research On Tire Trace Image Identification Method Based On Siaseme Network

Tire tracks have obvious stability and specificity,and can be used as one of the main features to identify automobile types and obtain automobile information.Therefore,in the investigation of traffic accident crime scene,the corresponding automobile types can be traced according to tire tracks,which is of great significance for restoring the crime scene and judging the accident responsibility.The similarity estimation of vehicle tire trace images is a key scientific problem in traffic accident investigation.By collecting tire trace images on the spot and comparing them with the images stored in the database,the vehicle involved in traffic accidents can be quickly locked.In this dissertation,aiming at the similarity estimation of tire trace images,the discriminant features of tire trace images are extracted by depth neural network technology,and the similarity calculation model of tire trace images is established.The main work completed includes:(1)A tire trace image data set was constructed,including 784 images of 34 vehicles.On-the-spot tire trace images of different vehicle types on different road surfaces are photographed,and the trace image areas are manually cut,so as to remove the interference of irrelevant content.In order to enhance the image characteristics of tire trace image data set,image preprocessing techniques such as contrast enhancement,image denoising and sharpening are used.In order to alleviate the risk of over-fitting caused by fewer image samples,the data volume is increased to 10 times by data enhancement methods such as image translation,rotation and graying.(2)The performance of classical convolutional network model in tire trace image similarity estimation task is compared and analyzed.Four convolutional neural network models,Alex Net,VGG16 Net,Res Net18 and Mobile Net＿V2,are selected and applied to the similarity estimation task of tire trace images.For each model,the loss function and optimization algorithm suitable for tire trace image modeling are analyzed,and the advantages and disadvantages of the four models are given,which provides a theoretical basis for subsequent experimental verification.(3)Aiming at the problem of insufficient generalization ability of classical convolution network when the sample size is small,a similarity estimation method of tire trace images based on twin network framework is proposed.Siamese network framework is more suitable for multi-classification and small sample data modeling tasks.Four convolutional neural network models,such as Alex Net,VGG16 Net,Res Net18 and Mobile Net＿V2,are used as the backbone networks of Siamese networks to construct the optimal tire trace image similarity estimation model.The experimental results show that the accuracy of Alex Net model can reach 80.04%,VGG16 Net model can reach 97.65%,Res Net18 model can reach 98.50%,and Mobile Net＿V2 model can reach 98.57%.Aiming at the problem of similarity estimation and recognition of tire trace images,this dissertation designs a siaseme network model based on four backbone networks(Alex Net,VGG16 Net,Res Net18 and Mobile Net＿V2),and tests and compares it on the self-built tire trace data set.The experimental results show that the siaseme network model based on Mobile Net＿V2 has the highest accuracy and can be used as an effective method to estimate and identify the similarity of tire trace images.Especially in the case of insufficient sample data,this method has important reference value.This work provides a new idea and method for tire trace image analysis,and also provides a reference for image similarity estimation and recognition in other fields.