Research On Traffic Sign Recognition Based On Deep Learning

With the rapid development of the economic situation in the 21 st century,cars have become a means of transportation for daily travel.People want to travel in a safe and fast road environment,so intelligent transportation system(ITS)came into being.Traffic signs are used to express the information of vehicles and pedestrians,such as instructions,directions,warnings and prohibitions.The functions of traffic signs include guiding drivers to drive safely and adjusting the traffic flow on the road.It also has the function of regulation and control for pedestrians and other non motor vehicles,and is of great significance for the smooth passage of roads.Whether the traffic signs can be correctly detected and identified is the key to the realization of its.Because the traffic signs in the natural road scene are exposed in the field for a long time,it is easy to be affected by bad weather,tree occlusion,light refraction and other factors,which brings some difficulties to detection and recognition.However,intelligent transportation system requires high real-time,accuracy and robustness of traffic information,so when considering the actual running speed of the algorithm,we should also consider the robustness of the algorithm,which puts forward higher requirements for traffic sign detection and recognition algorithm.This paper has carried out related research on the disturbing factors that are prone to the detection and recognition of traffic signs in natural road scenes.Combining the research results of relevant theories at home and abroad,traffic sign defuzzification based on multi-scale residuals and traffic sign detection and recognition algorithm based on improved RetinaNet are proposed.The main work of this article is as follows:1.Aiming at the optimization problem of over fitting or training model in the field of deep learning,the application of residual network in the field of image deblurring is proposed;in order to improve the feature extraction ability of network without increasing the depth of network,the application of multi-scale convolution unit level network structure in image deblurring is proposed;the deepening of network is regarded as one of model optimization Direction.The combination of the two,this paper will consider the deepening and optimization of the network model to further remove the image blur.The experimental results on tt100 k traffic sign data set show that the proposed algorithm based on multi-scale residuals has stronger restoration ability,more robust to image and its fuzzy type,noise level,and can obtain higher peak signal-to-noise ratio(PSNR)and structural similarity(SSIM).2.Aiming at the problem of small target detection of traffic signs,a traffic sign detection and recognition algorithm based on improved RetinaNet is proposed.Using FPN network can obtain different scale information of original image,which is helpful to detect traffic signs of different scales;combining high-level semantic information with low-level feature information to detect traffic signs of small targets,which improves the detection accuracy of small targets;using focal Loss function is applied to traffic sign detection task,which can solve the problem of difficult classification of similar traffic signs.This experiment trains on the data set of tsinghua-tencent-100 k,and tests the traffic signs in the actual scene.The experimental results show that the convergence time of the model is reduced by 17.6% on the basis of no loss of average detection accuracy.Figure [26] table [8] reference [51]...