Traffic Sign Recognition Based On Bilinear Convolutional Neural Network

There are different colors,shapes and pattern combinations in road traffic signs,which convey instructions,warnings and prohibitions of road information to drivers.As an important part of advanced assistant driving system and automatic driving technology,traffic sign detection and recognition can help drivers quickly obtain road information and improve driving safety.In order to detect and recognize the various practical traffic sign with small target and low resolution,two parts of modules of detection and recognition are researched in this paper.With the knowledge of deep learning and image processing,the paper studies the problems existed in traffic sign detection and recognition,and the processes are as follows:In order to improve the accuracy of traffic sign segmentation,image preprocessing is carried out first,in which a joint low-light enhancement and denoising method.This method can suppress the image noise and retain the edge details of traffic sign while enhancing the traffic sign image.As a result,the problems are resolved,such as weak contrast of traffic signs image caused by the changes of light and weather,and fuzzy details prompted by traditional images preprocessing,thus the accuracy of traffic sign detection is improved.In the traffic sign detection stage,coarse detection of traffic sign is applied at first to obtain possible areas of traffic signs by using the color probability model based on YCb Cr space.Then,HOG features of traffic sign areas are extracted and reduced dimensions with t-SNE algorithm.The feature vectors obtained after dimensionality reduction is used to train SVM classifier and accurately judge the candidate areas segmented in the previous step.Therefore,non-traffic sign areas are removed and the detection accuracy is improved.Experiments demonstrate that these procedures have wide adaptability,strong stability and robustness.In the traffic sign recognition stage,aiming at the over-fitting problem that happens easily when convolutional neural network is used to classify small samples,application of CNN-SVM model in the traffic sign recognition field is discussed.Traditional CNN-SVM model extracts features from network and traffic signs are classified by using SVM in the output layer.Based on this model,an improved method that conducts normalization processing for the features extracted by CNN network structure is put forward in this paper.The feature mapping model established through the inner layer of CNN,after the transmitted features being normalized,has a good feature presentation ability in traffic sign classification tasks,improves the SVM classification performance,and shows better classification accuracy and generalization performance.Finally,bilinear CNN for feature extraction is used in the improved CNN-SVM model,which can enhance the ability of feature interaction and solve the problem of low accuracy in small target traffic sign recognition.The bilinear CNN is improved by virtue of the feature validity difference extracted from different convolutional layers,so the network is guided to learn the feature with discriminants.Ultimately,an accuracy of 98.93% is achieved in the specific classification task of traffic sign dataset created in this paper,by using the model proposed in this stage,and the recognition accuracy of small traffic signs in low resolution is increased.