Research On Traffic Sign Recognition Based On Deep Learning

With the rapid development of economy and society,urban traffic has also become developed,under this background,driving has become an inevitable choice in people’s daily life.However,it also leads to more and more traffic accidents.In this case,intelligent transportation system has emerged,which changes the disadvantage that driving can only identify traffic signs by naked eyes,it promotes the development of machine vision technology,on the other hand,it provides a guarantee for the safe driving of drivers,which is of great significance to the harmony and stability of the society and people’s happy life.Traffic sign detection and recognition is an essential part of intelligent transportation system.This paper focuses on the detection and positioning methods and means of traffic signs,consult a lot of domestic and foreign literature,and fully analysis the existing problems of traffic signs,such as environmental factors,driver factors,factors of traffic signs itself,etc.,a traffic sign detection and recognition system is made with less time,high accuracy,and can cope with all kinds of complications environment,the research methods are detailed as follows:Traffic sign advance processing: for image tilt,rotate and translate the image.For image redundant information,cut the image.For different image size,zoom the image,select local mean method to reduce,select cubic spline method to enlarge.For image brightened by strong light and occlusion effects darken the image,RGB image is transformed into NTSC image by spatial transformation,brightness component Y is separated,adaptive histogram equalization with limited contrast is selected to process Y component,after processing Y component,I component and Q component are fused together to get NTSC image,RGB image is restored by spatial transformation.Wiener filter is selected to process image in view of the impact of noise.Select threshold segmentation to process the image.Image de fogging: aiming at the situation of fuzzy traffic signs in foggy conditions,this paper puts forward the method of improving the dark channel de fogging.On the one hand,it improves the calculation method of light illumination intensity value,on the other hand,the image is sub sampled,it can save time and improve efficiency,in order to reduce the damage of the high brightness part of the image to the image restoration effect,the double-sided filter is used to process the transmittance image,and the color and edge features of the image are better preserved.Traffic sign detection and positioning: improved fast is proposed R-CNN image location method.Construct convolution neural network model,increase the width of the model through parallel convolution kernel,fuse the HOG and LBP features of the original image with the image features extracted from multi-scale,NMS is used to form a location frame.The ticket counting detection frame mechanism is used to extract the feature maps processed by different convolution kernels,the candidate frames are input to RPN to get the candidate boxes,and the candidate boxes with more than 60% coincidence degree are selected,multiple candidate boxes are weighted and summed,finally,the location frame is fine tuned and the classification probability is obtained.Traffic sign recognition: a method of traffic sign detection based on SIFT feature fusion of convolutional neural network is proposed.By determining the extreme value,determining the key position,describing the key points,constructing a visual dictionary to process the shift feature of the image,obtaining the shallow features of the image.Improving the convolutional neural network Alexnet the structure,multi-layer convolution and pooling operation,the introduction of the Droupout layer to prevent over fitting,the introduction of BN layer to modify the weight,through the full connection layer to form the feature vector,get the deep features.Finally combine the two feature layers to form the feature collage layer,put into the softmax classifier for classification,achieve accurate identification of traffic signs.The experimental results show that compared with the domestic and foreign research methods,this method can effectively remove the light and fog,and achieve good results.The recall rate of traffic sign positioning algorithm is 97.4%,the accuracy rate is 98.4%,andachieve good results.The accuracy of traffic sign classification algorithm is 99.2%,which has achieved good results.