Traffic Sign Recognition Algorithm Research Based On Improved Convolutional Neural Network

With the rapid development of China's economy,with a car for many families is not difficult,but at the same time the increasing traffic accidents to the community has brought great harm.In order to avoid traffic accidents,in the highway,provincial highway,dangerous road sections and so will set the corresponding traffic signs to guide the driver according to the provisions of safe driving,but often due to the driver's own reasons,failed to pay attention to the road side of the traffic signs,Resulting in traffic accidents.The development of traffic sign recognition system can effectively use the means of technology to solve this problem,but the real scene is extremely complex,for the computer identification system,the different light intensity,occlusion and so will bring difficulties to the detection and classification,these factors Sign recognition and even car-assisted driving development.In order to solve the above problems,this paper summarizes the research on the related fields at home and abroad,in the face of the deformation and defects of the signs in the image,combined with the recent rapid development of machine learning support vector machine and convolution neural network brand recognition aspects of the algorithm research,this paper is divided into two stages,the specific content is as follows:The first stage is the research of sign detection algorithm.In recent years,MSER(Maximally Stable Extremal Regions)has been widely used in license plate recognition,text recognition and other fields,and its affine invariance,stability and other characteristics make it in these areas has made excellent achievements,SVM in the image classification,target recognition,especially the two classification problems on the outstanding performance,This paper combines the advantages of the two,proposed a stable based on SVM and MSER traffic sign detection algorithm.The algorithm first performs HSV color segmentation preprocessing on the image,then performs MSER region detection on the segmented image,filters the region by rule,and finally uses the SVM model to determine whether each region contains a sign.Based on a large number of experiments,the MSER parameters were determined by statistical analysis.On the basis of the German Traffic Sign Detection Benchmark(GTSDB)and the German Traffic Sign Recognition Benchmark(GTSRB)data set,The SVM training set and the SVM model are developed.The experimental results show that compared with other algorithms,the detection algorithm has higher accuracy and recall rate,and it has better stability in complex and changing scenes,and the calculation efficiency is higher,which satisfies the requirement of sign recognition system.The second stage is the algorithm of traffic sign classification.This paper presents a signature classification algorithm based on improved lenet-5 network.The traditional lenet-5 network has no strict requirements on the input form,the convolution kernel initialization method and so on.The convolution kernel initialization method adopts the most primitive random initialization,and the network model is very time-consuming in the training and classification stages.In this paper,the lenet-5 network convolution kernel initialization method is set to average Gabor core initialization,remove the traditional lenet-5 network C5 layer,replace the Sigmoid with the excitation function Re LU,and compare the real-time and accurate characteristics required by the traffic sign recognition algorithm.The output layer's Softmax classifier is replaced by a faster,lightweight SVM classifier.The experimental results show that the improved lenet-5 network is much better than the traditional network in the training convergence speed and classification time,and it has the advantages of high GTSRB data set with high representation in the field of sign recognition.Of the classification accuracy,to meet the signs of real-time accurate identification system requirements.In this paper,based on the requirements of accurate,real-time,stable and other requirements of the sign recognition system,combined with the research status quo at home and abroad,designed to meet the requirements of the sign detection algorithm and sign classification algorithm,respectively,in the GTSDB and GTSRB data set The ideal effect.