Research And Application Of Vehicle Recognition Method Based On Deep Learning

With the rapid development of science and technology,Chinese is gradually entering the ranks of innovation-oriented countries,at the same time,the quality of our lives has been improved,and we care more about the traffic travel situation,as a result,vehicles are becoming more and more popular as mobile phones which are an indispensable part of people's daily life.In recent years,the number of private cars has been increasing,and its growth rate is far more than urban road construction.This result has caused a series of chain reactions.For example,the increasingly serious traffic jams,traffic accident happens,false card deck car,hit-and-run and highway toll have stunted economic development and advancement and the society improvement of a region.After the research and development of a certain period of time,in our country the technologies of vehicle type identification and vehicle license plate recognition have reached a certain height,the vehicle can be roughly divided into small car(car category),car(van and SUV,etc.)and large vehicles(buses,trucks,passenger cars,etc.)and the technologies can solve the problem of license plate recognition simple scenarios,such as residential entrance and parking lot scene.But for vehicle specific models(such as Audi A4L)and the actual situation of vehicle license plate distortion,light conversion and low resolution,it can not be able to accurately identify the specific vehicle type and license plate number.In response to this problem,the Intelligent Transportation System(ITS)has emerged.Among them,fine-grained vehicle type recognition and license plate recognition are important parts of the intelligent transportation system.Deep learning can automatically extract features from original image data.It does not need to manually select features based on experience,and the features learned automatically are very good,so that it has a good application prospect in fine-grained vehicle type recognition and license plate recognition.In this paper studies four core technologies with vehicle image and license plate image which include vehicle detection.vehicle type recognition,license plate detection and license plate recognition.Finally,vehicle comprehensive information detection system is developed and integrated.The main work is as follows:1.Aiming at the problem of vehicle detection,this paper proposes a target detection method based on deep learning for vehicle detection,and adaptively modifies the model.This method can determine whether the image contains the vehicle and two big problems of location of the vehicle,at the same time in order to make the robustness of the trained model has a better image of the real scene of the training data model of a total of more than 44000 tagged vehicle image,and each image contains one or more car.Experiments on the test set show that the detection and location accuracy of the algorithm is 92.79%.2.For existing vehicle recognition methods,it is difficult to identify vehicle manufacturers,models and production age information accurately,so a vehicle recognition method based on deep learning is proposed.At the same time,this method is implemented with convolutional neural network.Then a series of experimental comparisons are used to select the best neural network--IncptionAB-Full.The experimental data has 223 car models and 38,998 images.At the same time,make the image data augmented.Finally,The accuracy of the vehicle type identification is improved to 96.65%in the test set.3.Then,on the accurately positioned vehicle image,the target detection method based on deep learning is used again to locate the license plate,and at the same time,the partial structure of the model is modified to make it more suitable for license plate detection.Next,the license plate location method with mathematical morphology and the license plate location method based on color features are combined.And the two license plate location methods are compared on the test data set.The experimental results show that the method based on deep learning target detection has more advantages in license plate location.4.The license plate is corrected with the method of Radon,and whole license plate is identified with the target detection method again.At the same time,the traditional license plate recognition process is used to perform the character segmentation and recognition of license plates.Finally,the advantages and disadvantages of the two routes are compared through experiments to provide the necessary relevant information for the subsequent system integration.5.Based on the previous research,the development and integration of vehicle comprehensive information detection system is used with Python language.With the trained model structure,the system human-computer interaction interface is designed and developed,at the same time,the CPU and GPU modes,as well as single data and multi data processing,are also taken into account.Finally,the key points of this work and the integration and development of vehicle integrated information detection system are analyzed and summarized,also some problems of the system are pointed out,and the further research plan for further research is prospected.