Traffic Sign Recognition Based On Deep Learning

Traffic sign recognition is not only an important part of intelligent driving,but also a challenging task in computer vision.With the rapid development of artificial intelligence technology and related hardware,the mainstream traffic sign recognition technology has changed from the traditional image processing technology to the deep learning method based on convolution neural network.This thesis focuses on the traffic sign recognition method based on convolution neural network.Aiming at the data distribution and visual characteristics of traffic signs in natural scenes,this thesis improves the traffic sign recognition effect from two aspects of network structure and feature extraction.The main works of this thesis are as follows.(1)A domestic traffic sign dataset named TS-507 under natural scenes is constructed,whitch includs 507 categories of domestic main traffic signs.It has more than 200000 traffic sign pictures coming from real road driving scenes.The dataset is open for download.(2)In order to solve the problem of long tail distribution of traffic sign data in natural scenes and the problem of fine-grained recognition with small difference between classes and large difference within classes,this thesis proposes a new bilateral destruction reconstruction network structure named BDCL,which can effectively improve the recognition accuracy and the recognition effect of tail data.(3)An efficient,lightweight and universal attention mechanism module BAM is proposed.In order to improve the feature extraction ability of the backbone network for traffic sign image,we design a spatial and channel joint sub module based on attention mechanism.The module can effectively improve the attention of the feature extraction network to the discriminative regions in the image without introducing additional parameters,and the module can be easily combined with various feature extraction networks.In order to verify the effectiveness of the proposed method in traffic sign recognition,all the methods are tested on the established dataset and other open source datasets,and good results are achieved.