Vehicle Logo Recognition Using Tree-based Convolution Neural Network

Along with the automobile industry continues to rise,the car has begin to enter the ordin ary families.the number of vehicles increasing brings great challenges to the management of r oads and vehicle identification technology has become a hotspot of current research.The logo contains the makers of information,not easy to replace,in vehicle identification plays an important role.In vehicle recognition system,image acquisition through the camera,the ima ge acquisition in the natural environment may be ambiguous,distortion,deformation,occlusio n etc,for feature extraction and image recognition inefficient.For the current logo recognition technology mostly uses traditional features,such as SIFT,HOG,combined with the traditional classification and recognition effect is not very ideal,and the complexity of large,long time,is not conducive to the increasing amount of recognition for vehicle status,application value relatively low.However,deep learning technology,which is more simple and efficient vehicle logo recognition and stacke multiple hidden layers,able to simulate more complex functions and has a strong ability to learn the main features of the data set.a typical model of deep learning convolutional neural network,through the simulation of the human brain operating mechanism,can automatically learn the image features.Convolutional neural network with weight sharing,sub sampling and othe r characteristics,making the extraction of the characteristics of translation,distortion invarian ce,can extract features logo more significantly efficient feature.compared with the traditional method,The current logo recognition not like face recognition,handwriting recognition with open training data set and test data set,can only be collected small sample size by the experimenter himself.while for small sample data set,convolutional neural network training will appear over fitting phenomenon.In this paper,a artificial neural network and convolution neural network are introduced.Based on the characteristics of the current vehicle mark recognition and the characteristics of vehicle image acquisition,a new convolution neural network model is proposed,and is proposed a solution to the phenomenon of overfitting in small sample vehicle train.The main work of this paper is as follows:Firstly,according to the current vehicle logo recognition method,combined with the char acteristics of convolutional neural network,puts forward the convolution neural network mod el for tree structure.Compared with the traditional convolution network structure,this model has not a single kind of convolution kernel,but has a variety of convolution kernels.Experiments show that the recognition of tree-based convolution network model is better than the traditional vehicle logo method recognition.Secondly,According to the data of the vehicle logo small sample set and the convolution al neural network training method,we use three methods of convolution network training.Firs t,the expansion of the sample data set,data enhancement operation,geometry changes of sam ple data.Second,the network structure adjustment,in the small sample data set,using a small network structure,the convolution kernel,network size,network structure adjustment.Third,transfer learning,using existing knowledge to study the current data set,based on the existing public data sets of training and learning,and then the network parameters in the logo on the d ata set by the model has some fine-tuning,to improve the recognition rate.