Research On Vehicle Recognition Algorithm Based On Deep Learning

With the rapid development of economy,people’s material living standard has been improved significantly,and people’s demand for vehicles is also increasing.However,at the same time,traffic environment problems and traffic management technology problems are becoming increasingly prominent.In many aspects,the traditional traffic management mode can not meet the needs of the changing situation.In this context,the role of intelligent transportation system is particularly prominent.Vehicle recognition system is the core content of intelligent transportation research field,and it is also the research hotspot of image processing,computer vision,pattern recognition and other interactive fields.As an indispensable part of intelligent transportation,vehicle identification plays an important role in high-speed toll system,driverless and other fields.As an indispensable means of transportation in daily life,vehicles have many brands,so it is of great significance to identify them.Vehicle type recognition system is an important application of pattern recognition technology and computer vision in intelligent transportation system(ITS).It is widely used in highway vehicle video monitoring,parking lot management,and electronic police.Therefore,in-depth study of vehicle identification system not only has high theoretical value,but also has high economic and practical value.In this paper,the related algorithms of vehicle recognition are deeply studied,several effective improved algorithms are proposed,and the vehicle recognition system is developed.(1)This paper studies the model recognition algorithm based on deep learning.The research object adopts the model data set published by Stanford,which includes a variety of model images with fine-grained model labels.Because there are many types of vehicles in the dataset,but the number of images in each category is small,it is not easy to learn enough features for classification.Firstly,this paper studies the common data enhancement methods and expands the data set.The original data set and enhanced data set are integrated and preprocessed,which are divided into training set and test set for training.(2)For the problem of complex and changeable vehicle background in multi-directional environment,this paper proposes a SSD target detection framework based on resnet-50.The framework is used to detect and locate the vehicle target in the image,which reduces the interference of other environmental factors on vehicle recognition.(3)In this paper,a bilinear perception network is proposed for vehicle recognition.Firstly,different end-to-end convolutional neural network models are created,and the knowledge learned from many data is used to solve the problem that there are only a few labeled sample data in the data set through transfer learning.Through training and learning,effective vehicle features are extracted and classified.Through comparative experiments to select the model,analyze and study the benefits of different training modes and different models in the accuracy of classification.Then,we choose the best perception net model to optimize.Finally,the recognition accuracy is improved from 85.7% to 92.3%,which verifies the usefulness of the optimization...