Research On Model Identification Application Based On Convolutional Neural Network

New technologies such as machine learning,deep learning and natural language processing are the driving forces behind the development of all aspects of modern society,whether in terms of content recommendation on e-commerce sites or in the filtering and analysis of large amounts of network information data,which play an indispensable role.Model identification has broad application prospects,in the process of model identification,the accuracy of vehicle positioning and the accuracy of vehicle model classification are important factors to determine the quality of model identification.Most of the existing model recognition algorithms are the identification of different brand models of vehicles,which are lack of more detailed identification of different models of the same brand and the accuracy of vehicle positioning.Based on the convolutional neural network,this thesis studies the model identification:vehicle positioning and model classification,and presents a model identification network model based on Faster RCNN,which realizes the fine-grained model recognition and improves the accuracy of model identification.The main research content is as follows:(1)Data augmentation can increase the generalization ability and robustness of the trained model while increasing the amount of training data.This article uses Stanford University's public car data set Stanford Cars Dataset as the basic data set using network reptiles,geometric transformation,Gaussian noise,affine transformation to expand the underlying data set,and then pre-processing and labeling the data obtained,forming a new automotive data set Cars-extend is used for network model training.(2)The existing target detection RCNN,Fast RCNN and Faster RCNN algorithms are analyzed and studied.The faster RCNN algorithm is used as the basis for vehicle positioning in model identification,and two pre-trained models on the ImageNet dataset,VGG16 and ResNet50,were extracted from CNN as features of Faster RCNN,which were trained and tested on the new automotive dataset Stanford Cars-extend,the tests showed that the vehicle with the ResNet50 network model was more accurate in positioning.(3)Based on the use of ResNet50 network model as the feature extraction CNN of Faster RCNN,the InceptionV3 network model is used to classify the models,which constitutes the model identification network model of Faster RCNN and ResNet50 and InceptionV3.On the new car data set Stanford Cars-extend,the ResNet50 and GoogLeNet Inceptionv3 pre-training models on the ImageNet dataset were trained and tested.The experimental results show that the model classification effect is better with GoogLeNet Inceptionv3,and the accuracy of model recognition is 93%.