Research And Application Of Vehicle Type Recognition Based On Deep Learning

With the development of deep learning and computer vision technology,image recognition technology has made great progress in recent years.At present,good results have been achieved on the classification of vehicle types,but the classification categories stay on the coarse-grained categories,while the resear-ch on fine-grained vehicle recognition is less,and the accuracy is not high enough.It is of great significance to study how to classify the fine-grained categories correctly and improve the recognition accuracy.This paper studies the vehicle recognition method based on deep learning,using Stanford open data set of vehicle types as research object.The dataset includes a variety of vehicle type images with fine-grained labels.Since there are many types of vehicles and few images for single category in the dataset,it is difficult to learn enough features for classification.Therefore,common data enhancement method and crawler method are studied to expand the dataset.At the same time,generative adversarial network(GAN)is studied,learning the distribution of vehicle images by using deep learning and expanding the dataset further by generating images according to the distribution.The original dataset and the enhanced dataset are intergrated and preprocessed,and divided into training set and test set for deep learning training.Secondly,a bilinear Inception CNN(BI-CNN)with hierarchical labels is proposed for vehicle recognition targets.This paper firstly establishes different end-to-end convolutional neural network(CNN),and uses knowledge learned from large amouts of data through transfer learning to solve the problem of small number of labeled sample data in the dataset.Through training and learning,effective vehicle features are extracted and used for classification.Through comparison experiment,the effects of different models and different training modes on classification accuracy are studied and analyzed.The optimal InceptionNet is selected for further optimization.The optimization of network focuses on structure and hierarchical label.The classification accuracy rate is improved from 8 7.8%to 91.5%by the optimized network,which verified the effectiveness of optimization.Finally,based on the CNN model obtained from training,a vehicle type recognition system is designed and established.The system can read images selected by users,use learned CNN to extract effective features of input images and obtain the prediction of vehicle classification according to the features.The prediction results are displayed with a user-friendly interface.By testing and verification,the system is capable of classification and recognition of fine-grained vehicle types.