Design And Implementation Of Vehicle Recognition Based On Deep Neural Network

With the rapid increase of the number of vehicles in China,intelligent transportation system is becoming more and more popular.As a kind of basic attribute of vehicle,accurate identification and statistic of vehicle type is very important to the construction of intelligent traffic system.This thesis expounds the research significance and development status of vehicle type classification technology,and summarizes the basic theory of deep learning.This thesis briefly introduces the hardware and software vehicle type classification technology,and focuses on the computer version technology based on deep neural network algorithm,which has the advantages of low deployment cost,high accuracy and strong learning ability.For the disadvantage of existing classification methods,such as too large granularity,this thesis proposes a classification model that integrates the overall features with some key local features.The research data set is mainly for the front face of the vehicle captured by the surveillance camera.In order to make full use of vehicle front face image for the task,the following work.In order to establish a model database suitable for this article,this article integrates the vehicle data sets of PKU Vehicle ID,BIT Vehicle and various types of vehicle images crawled from the Internet using web crawlers,which ensures the richness of data sources used in this paper.And in order to further enhance the contrast of the darker images,a histogram equalization process is performed on the images whose brightness in the data set is lower than the threshold.The final model database contains a total of 10,432 images from 228 categories vehicle samples such as jeep-guides,BMW-X1,and BYD F3-2007-2011.In addition,five key components such as rear-view mirror,vehicle lamp,license plate,vehicle logo,and front windshield are defined as the strong semantic information part of the front face image of the vehicle,and corresponding detection datasets are made.For the problem that the original SSD(Single Shot Multi Box Detector)has insufficient utilization of the network feature maps,which leads to the problem of low accuracy when detecting small objects,an improved detection model for detection of vehicle facial critical parts.This moedl utilizes feature pyramid network as the base network to extract feature map, better integrates the low-level and high-rise features of neural network,improves the detection ability of SSD algorithm for smaller target such as vehicle parts,and improves the detection rate of vehicle components by 5% compared with the original SSD model.The commonly used image classification network is analyzed,like Alex Net,VGGNet,and Res Net.And they and traditional image classification method were trained with the datasets in this article.Experimental results show that the accuracy of the method based on deep neural network far outperforms traditional image classification method.The Res Net classification accuracy is the highest,reaching 86.8%.On this basis,the reason why Alex Net's accuracy rate is low is studied,and after improvement,the accuracy of the classification task on the model is increased by 5.1%.Then,the improved Alex Net and the modified Le Net were used to extract the features of the entire vehicle picture and key components respectively,and eventually formed a joint feature.Use the generated joint features for classification.Ultimately 2% more accurate than previous Res Net.Based on the above research results of vehicle classification,the paper designs a model classification platform based on vue.js and flask Open source framework,utilizes Caffe open source depth learning library,and develops a set of beautiful and Easy-to-use interface software in Linux environment.