Research On Small Traffic Sign Recognition Technology Based On Deep Neural Network

Traffic signs contain information on the road conditions and traffic regulations of the roads and are a non-negligible factor in the driving of motor vehicles.Driving according to traffic signs can greatly avoid safety accidents and avoid traffic jams.Because the farther the object in the picture,the smaller it looks,improving the recognition accuracy of small traffic signs enables drivers to know in advance the road conditions in order to safely drive the vehicle.There are two main difficulties in identifying micro traffic signs: In one aspect,since traffic signs are mostly outdoors,there are problems such as different light intensity,blocking,deformation,and incompleteness,which are not conducive to recognition;on the other hand,micro-signage The proportion of the entire picture is small,and the number of occupied pixels is also small,which leads to a small amount of information carried,and it is difficult to extract relatively good features,which also makes it difficult to identify.This thesis summarizes the existing algorithms for identifying micro traffic signs at home and abroad and proposes a micro traffic signs recognition algorithm based on deep neural networks.The algorithm is divided into two phases: the first phase is the detection phase.The algorithm in this thesis can intelligently integrate the resolution information and context information of the picture into the neural network.Through this network,it is possible to extract candidate regions that may be traffic signs in the input image,and in order to avoid double counting,this thesis proposes a template selection strategy of “Type_AB”.The second stage is the classification stage.The main task of this stage is to classify the proposal regions in the first stage.This thesis integrates Inception V3 and Xception.Combining voting techniques to classify candidate regions.Finally,this paper analyzes the influence of the depth of neural network on the detection of small landmarks,and analyzes the convergence speed of this algorithm.Under the premise of using the TT100 K data set,the accuracy of the algorithm proposed in this thesis is higher than the algorithm proposed by Ren,Zhu,Yan et al.in the range of(0,400] resolution.In the range of(96,400] resolution,the recall is higher than those of the algorithms proposed by Ren,Zhu,Yan,etc..In the range of(32,96],(0,400] resolution,the recall is lower than the algorithm proposed by Yan et al..The algorithm proposed in this thesis can identify some traffic signs that are not marked manually in the TT100 K data set..