Study On Vehicle Logo Recognition Method Of Transfer Learning And Depth Based On Convolutional Neural Network

Vehicle-logo as a key feature of the vehicle,combined with plate recognition and vehicle color recognition can greatly improve the reliability of vehicle identification which plays an important role in scenarios of identifying vehicle information like security check and fake plate vehicles recognition.Traditional identification methods require manual extraction,which cannot automatically combine classifier with feature extractor.Its process of tedious and most of them are shallow structure algorithms,which are insufficient in generalization of complex classification.In the actual test,we found that this problem is a fine grained classification problem.There is a large similarity between different types of vehicle-logo while there is a large difference between different samples of the same type.Under the complex circumstances,the small difference of vehicle-logos is easily covered,which makes it difficult to classify vehicle-logos.Deep convolutional neural network can achieve better results in image classification,but there are problems itself.For example,training a deep convolutional neural network requires a large number of samples.Based on the deep convolutional neural network,the main research contents include:(1)Aiming at the problem that a good network model cannot be trained with small data sets,this thesis combines transfer learning and deep convolution neural network.We transfer the feature extractor of pre-training network model ResNet,fine-tune top-level parameters and replace the top-level classifier for SVM in order to obtain a primary model for vehicle-logo recognition,SVM-Res Net.(2)To improve the recognition accuracy of fine grained classification of vehicle-logo and vehicle-logo image under complex circumstances,the thesis combined bilinear feature and SVM-Res Net model,considered the second order statistical properties of the convolution layer output feature mapping,to obtain a classification algorithm with higher recognition accuracy,Bil-SVM-Res Net.(3)Considering the computing performance of the terminal equipment,the thesis compressed the algorithm in(2)to obtain the compressed model CLR-SVM-Res Net,which reduces non-essential parameter calculation and space complexity of the algorithm and also can maintain a good recognition accuracy.(4)The algorithm proposed in this paper is verified experimentally,the spatial complexity and recognition precision of the algorithm are analyzed,and the robustness of the algorithm CLR-SVM-Res Net was tested by means of noise addition and brightness adjustment.Combining transfer learning with deep convolution neural network can not only effectively use the powerful generalization ability of deep convolutional neural network,but also avoid designing complex models and time-consuming training.Experiment shows that the proposed algorithm Bil-SVM-Res Net and CLR-SVM-Res Net have high accuracy and robustness in fine grained classification of vehicle-logo with small sample data sets,with an accuracy of above 98%.And the model of algorithm CLR-SVM-ResNet is smaller so as not to be limited by terminal equipment,which is meaningful.