Application Research Of Vehicle Type Recognition Based On Deep Learning

With the improvement of people’s living standards in recent years,the number of cars has increased rapidly.The problems such as frequent traffic accidents and difficulty in vehicle management are becoming increasingly prominent.In order to solve these problems,many experts and scholars are committed to researching intelligent transportation systems.The detection and identification of car models is the basis of modern intelligent transportation systems,which can be applied to traffic flow statistics,high-speed vehicle classification and toll collection,and traffic safety management,Therefore,research on car model recognition has significant practical importance in real life.In the current era of intelligent transportation,there are some urgent problems that need to be solved in the detection and recognition technology of automobile models.Firstly,the difficulty in identifying car models is increasing due to the increasing similarity between different brands and different series of the same brand.Secondly,the accuracy of identification is low due to the different shooting angles and image noise in the car model dataset.Thirdly,most car model identification algorithms require manual marking of certain feature points of the vehicle,resulting in high labor costs.To address these issues,this thesis proposes a car model identification method based on weakly supervised attention and car face alignment,and designs a "timely" car dataset for the current domestic market.Based on this,a car model identification system has been developed.The main work of the thesis is as follow:(1)In order to solve the problem of low accuracy in vehicle recognition caused by the similarity between car models,different shooting angles,and image noise in real-world scenarios,this thesis proposes a car model recognition model CNN-cf3 based on weakly supervised attention and car face alignment.The algorithm uses a three-layer convolutional neural network to detect vehicles in images,replacing manual annotations with detection boxes and five feature points.The five vehicle feature points are used to align the appearance of the vehicle,solving the problem caused by differences in shooting angles and improving recognition accuracy.Then,a weakly supervised attention network is used to pay special attention to the area around the feature points provided by the vehicle detection,generating attention maps and extracting more discriminative local features to address the problem of reduced recognition accuracy due to similar vehicle models.Finally,image enhancement guided by attention mechanisms is used to suppress background noise and extract more detailed features.To verify the effectiveness of the proposed algorithm,it was tested on a dataset of real-world vehicle images.The experimental results show that the CNN-cf3 network model has higher recognition accuracy and stronger robustness compared to other network models(2)In order to improve the applicability of the vehicle model recognition algorithm in the current real-world usage environment of domestic vehicles,this thesis has established a new automobile model dataset,Ccars-150,with "timeliness",which is used for algorithm training and recognition Firstly,a web crawler based on the Scrapy framework was designed,which collected images of 150 commonly seen domestic car models.Then,through data augmentation,a total of 163,400 images were obtained.In order to facilitate comparative experiments,the selected images were labeled using Label Img to create the car dataset Ccars-150.Using VGG-16 network model,Res Net-50 network model,and self-built CNN-cf3 network model respectively,comparative tests were conducted on classic datasets Compcars,VMMRdb,Stanford Cars,and self-built dataset Ccars-150.The experimental results indicate that three network models have a high accuracy in recognition on the Ccars-150 dataset,and the data has "timeliness",which can be better applied to the design of vehicle recognition systems.(3)Designed and realized a vehicle type recognition system.This thesis base on the Pycharm platform to design and implement a complete vehicle detection and recognition system,which integrates image processing,vehicle detection,model recognition and visual display,the system has the functions of account registration,system login,model initialization,vehicle identification and data recording.The user can use the UI interface to identify the vehicle in the picture and video.