Research On Traffic Sign Detection And Recognition Based On Convolutional Neural Network

With the development of science and technology,the rapid growth of the number of people owning cars,many urban traffic congestion has become a common phenomenon,lead to the traffic safety problem increasingly serious,caused the social big losses.In this case,the unmanned was more references in the 13 th during the two sessions,and secretary of the ministry MiaoWei also mentioned that with the advent of 5G,unmanned also entered People’s Daily life in the future.Image detection and identification of traffic signs is an important part of the unmanned system.It helps drivers or autonomous vehicles by identifying traffic signs on the road.So how to quickly and accurately detect and identify the traffic signs in the natural scene requires further discussion and research.In this paper,the problem of slow detection speed is caused by the long time required to generate part of the traffic sign image in the candidate region.This paper presents a Faster R-CNN method,this method can be generated through regional Network(Region Proposal Network,RPN)to extract candidate area,and use of RPN Network fast convolution neural Network to extract candidate area of the study,by two Network after convolution sharing layer characteristics,so as to achieve fast and efficient detection.In the training model,the RPN network and the Fast R-CNN network were adjusted by using the GTSDB dataset.In order to verify the traffic sign detection algorithm based on Faster R-CNN effectiveness,this paper based on Caffe framework to build Faster R-CNN model,for the python programming language,the final result compared with classical methods,this paper gives the traffic sign detection method is more intelligent.The image features of traffic signs can not fully reflect the low recognition rate caused by the original image feature.Are presented in this paper a convolution based on improved neural network method,the original network can automatically extract the image feature,the image pixel matrix in the form of the input to the network,through the forward and reverse transmission process of network training.During the training period,the GTSRB data training set was used to train the network to learn the images in various situations.When the loss function was completed,the training was completed and the network parameters were saved.During the testing phase,the parameters of the training phase are used to identify the traffic sign images.In orderto make network have better effect,the following improvements were made in this paper,the network model: 1.Amend the convolution and pooling layer alternate combination for input layer-convolution convolution-pooling layer-convolution-convolution layer-pooling layer,make full extraction of the feature;2.Change the activation function Sigmoid to ReLU to effectively transfer the gradient;3.Use Max Pooling method to replace Mean Pooling method in the Pooling layer;4.Join the Dropout operation after the full connection layer to train for the network.In order to quickly and effectively implement the design scheme,this paper implemented the python language in the deep learning platform Tensorflow,and the final recognition rate reached 96%.Compared with the convolution neural network before improvement,this method is greatly improved.