Convolutional Neural Network Study On The Recognition Of Traffic Signs

Traffic sign recognition in natural environment is the essential requirements of unmanned technology.In recent years,Baidu,Amazon,Google,Facebook and other companies pay more attention to the unmanned technology.A few years ago,researchers usually use color,shape and scale to express the features of traffic signs.This paper makes full use of the convolutional neural network fired up in recent five years to the recognition of traffic signs.Through the network optimization,experiments,adjusting the network structure and parameters,the final design of a traffic sign recognition network model has outstanding performance,good accuracy,high efficiency and portability.To solve the classification problem using convolutional neural network theory,a large amount of data is usually used.The paper is based on the German traffic sign data set(GTSRB),the image processing effect best Caffe is used as a framework,and the network model and parameters suitable for traffic signs identification are designed.The research work of this paper mainly includes the following aspects:(1)based on the Siamese network model,we propose a double linear network.We first do not share all the characteristics of the volume and the weight of deposit pool of Siamese network.With the continuous extraction of image features,the weight is shared,so as to improve the accuracy of traffic signs.(2)in order to further optimize the structure of network and improve traffic sign recognition accuracy and portability,this paper uses the 2-channel network structure,and the input images are updated from the single channel to double channel.This method can not only improve the generalization ability of the network,but also improve the feature extraction ability of image expression,and brings more convenient means of identification in the traffic sign recognition because of insufficient the amount of data.(3)at the end of this paper,using the application of inception new structure as the basic element,the Ubuntu14.04 as operating system,the Caffe framework as the platform,and GoogleNet as the framework of the overall network model,a high precision TrafficNet network model is designed.We mainly improve the network node hierarchy structure and incentive function,and the speed of network training is greatly accelerated.For the 43 types of traffic signs,99.8%correct classification rate is achieved for a total of 54324 pictures.