Traffic Sign Image Recognition Method Based On Convolution Neural Network

As an important part of future intelligent transportation system,traffic sign recognition plays an important role in assisting driving and ensuring the safety of driver's life and property.In the real traffic environment,because of the influence of various factors,such as image blurring,partial occlusion and noise interference,the quality of the traffic sign image is not high,which brings great difficulties to the accurate identification of traffic signs.Therefore,the use of convolution neural network has the characteristics of strong learning and classification.It is of great value to study the recognition rate of traffic sign recognition,and therefore it is of great application value.When convolution neural network is used to solve classification and recognition,a large number of data sets are needed.Due to the lack of traffic sign data set,the German traffic sign data set(GTSRB,German Traffic Sign Detection Benchmark)and the(BTSC,Belgium Traffic Sign Classification Dataset)are selected in this paper to train the network.A clear and efficient framework of depth learning framework Caffe is used as an experimental platform also a network model suitable for traffic sign recognition is designed,and the most effective super parameter of traffic sign recognition is obtained through training.The main contents of this paper are as follows:(1)In terms of network structure and parameter setting,a parallel multi-scale deep fusion convolution neural network structure named M_S_F Net(Multi-Scale_F Net)is built.First of all,the idea of modularization of network structure is used and the theory of deep integration of parallel network to confirm the idea of network construction.After that,a comprehensive M_S_F Module network module is proposed based on the comprehensive multi-scale ? asymmetric convolution method and the advantage of batch normalization(BN).Finally,modules are arranged in a basic stack network.The parameter settings of the above work are set according to the source network AlexNet of the basic network.(2)In this paper,the convolution layer used in the network M_S_F Net is replaced by the basic module except the first layer.In order to verify the arrangement of the M_S_F Net structure and the effectiveness of the asymmetric convolution kernel,a computer simulationexperiment is carried out by using traffic sign images data set(GTSRB)with large amount of data.After 60000 iterations and multiple super parameters tuning,98.05% recognition rate is achieved.At the same time,the M_S_F Net network is compared with M_S Net(Multi-scale Net)and M_S+M_S_F Net(The network structure composed of M_S_F module and M_S module.),and the experimental results show that the recognition rate of the traffic sign image of M_S_F Net structure is obviously higher than that of the other two networks,which shows the correctness and validity of the M_S_F Net structure.(3)In this section,a light convolution neural network structure MyLeNet is designed,and the computer simulation experiment is carried out with the traffic sign data set(BTSC)of small amount of data.The good image recognition effect is obtained,and the recognition accuracy is up to 98.30%.On the BTSC dataset,the recognition rate of traffic sign images is higher than that of M_S_F Net and other two network structures.It shows that the lightweight convolution network structure MyLeNet designed in this paper can achieve high recognition performance when the number of pictures in the dataset is small.(4)At last,the author collects 10 types of traffic signs images in China,including traffic signs taken randomly on the streets of Guilin,and the noise is added to some of the images.After that,the image is input to M_S_F Net network model for verification.The results show that the M_S_F Net model can identify the traffic sign images accurately and it has good robustness.