Traffic Sign Recognition Research Base On Deep Learning

Traffic sign recognition is an important part of intelligent transportation.Because the drivers are easy to miss the traffic signs for the lack of attention,the automated traffic sign classification will assist the drivers to drive safely.However,the traffic sign recognition is challenged by varied factors,such as weather change,camera angle,occlusion etc.So the algorithms should be robust to these factors,and have characters of real-time.The traditional image recognition algorithms extract image features manually,then classify these features by choosing suitable classifier.The stage of artificial feature extraction is easy to cause the loss of some features,which leads to a decline in recognition performance.What's more,the quality of the images has a great effect on the recognition results when we extract features artificially.The traffic sign images taken in the real world always have poor quality.Compared to the traditional recognition algorithms,deep learning can realize the image recognition by automatic feature extraction.As a deep learning model,the convolutional neural network has made some remarkable achievements in the image recognition field.The paper analyzes the advantages and disadvantages of the convolutional neural network,and studies the relevant theory.The paper proposes new networks based on the convolutional neural network for traffic sign recognition.The paper's main work and research achievement is as follows:(1)Based on the convolutional neural network,the PCA Network can directly calculate the filters by the kernel principal component analysis in each convolutional layer.The PCANet avoids the repeated tuning and complicate manual intervention.However,the PCA method ignores the higher-order relationship between the data,which causes the extracted features are not optimal.As a nonlinear extension algorithm for linear PCA,KPCA can exact more excellent features.Therefore,the paper proposes the Kernel PCA Network for traffic sign recognition by inducing the kernel method.The Kernel PCA Network uses two-layer convolutional network to extract abstract features.After nonlinear mapping and pooling,the support vector machine is applied to the final classification.The results on GTSRB show the approach can achieve a high recognition rate.(2)Different from the traditional CNNs using the same pooling in each subsampling layer,the paper proposes a strategy of combining different pooling operations in our shallow network to achieve better performance.Because the recognition systems of local features and classifiers can become more competitive by carefully adjusting the pooling operations.Therefore,the shallow convolutional neural network based on the combined pooling method is proposed.The paper combines different pooling operations to improve the CNNs' recognition performance.In view of the real-time performance,the activation function ReLU is used to improve the computational efficiency.In addition,a linear layer with softmax loss is taken as the classifier.The results on GTSRB present that the network not only achieves a high recognition rate,but also guarantees real time performance.