Research On Vehicle Type Recognition Based On Convolutional Neural Network

With national incomes rising,the purchase of motor vehicles has become the first choice in the pursuit of a more convenient life.In order to attract consumers,businesses have also introduced more and more diversified cars.The surge in the number of cars has made intelligent transport systems increasingly important.Vehicle identification is a crucial link in intelligent transportation system.The key of vehicle identification is feature extraction.In the traditional algorithm,the prior knowledge of the corresponding domain is often needed to extract effective features,so the researchers propose the features specific to their own domain.However,if these features are directly applied to large-scale image recognition,the effect is often not very good,and it needs a lot of time to design and tune.With the development of convolutional neural network,researchers began to study learning features instead of design features.The automatic extraction of multi-level features by convolutional neural network satisfies the demand of automatic learning features.Experiments show that the performance of convolutional neural network is much better than the traditional feature extraction algorithm.At the same time,the emergence of massive image data and the improvement of computer hardware also bring great opportunities for the large-scale training of convolutional neural network.In recent years,people mainly focus on the recognition between some large categories of models by using convolutional neural network,such as SUV,car,truck,etc.,which has a high degree of differentiation and easy feature extraction.The research results of the convolutional neural network in the recognition of these vehicle models have produced practical results.However,the research investment of convolutional neural network in fine-grained image recognition is not enough,and the social benefits are not enough.Fine-grained image recognition is also known as subclass image recognition.There are slight differences between subclasses of cars,which are more difficult to distinguish in visual images,so this subclass image recognition needs more powerful feature extractor.The main work of this paper is to conduct in-depth research on fine-grained vehicle identification to improve the precision of fine-grained vehicle identification.In this paper,the precision of fine-grained vehicle identification is improved mainly from two directions of dataset and convolutional neural network structure.In terms of data sets,the candidate box information and SRGAN algorithm were used to expand the dataset,and PCA data preprocessing was conducted to speed up the training.The results show that the candidate box information is used to expand the data set to increase the multi-scale information of the input image,and the recognition accuracy is improved.SRGAN can be used to improve the resolution of the data set.Since the resolution of the data set in this paper is high,the improvement effect is not obvious.However,this method can expand the data set,prevent overfitting,and has great help in reducing the hardware acquisition facilities of the data set.In terms of network structure,it is found from the experimental comparison of VGG,Inception and ResNet that multi-scale information is more effective for the identification of fine-grained vehicle models,and the key to extract multi-scale information is not to deepen the depth of the network,but to expand the width of the network.This structure can effectively extract multi-scale information.In addition,in this paper,faster-rcnn was used to detect and identify vehicle models through manual labeling to remove the interference of background information,and the experiment showed that the recognition accuracy could be improved.