Research On Traffic Sign Detection And Recognition Based On Deep Learning

As the number of domestic cars continues to increase,traffic safety issues have become increasingly prominent.As a promising alternative,autonomous driving technology attracts much attention of researchers.Traffic sign detection and recognition is a critical part of this technology,which has been researched for many years until present.However,in the present stage,there are still some challenges.To deal with them,this article uses deep learning related technologies to explore underlying problems of traffic sign detection and recognition.Two network structures are designed,and their effectiveness are verified on public datasets.Inspired by the attention mechanism in the human visual system,this article proposes a group multi-scale attention network to address the problems involving traffic sign detection and recognition.The network includes a group multi-scale attention module and an adaptive pyramid convolution module.The group multi-scale attention module helps the network focus on some important regions of different scales,meanwhile suppresses interference information,so that the network can learn more pure features.The adaptive pyramid convolution module makes the network in a way learn optimal feature fusion mode,which can aggregate features from different scales to a certain scale adaptively,and exploit the effective information in the feature map.In addition,this article improves the loss function by introducing localization quality estimation into the classification loss,so that classification and localization can be connected.Finally,the performance of the proposed method is tested on several public datasets.The final experimental results demonstrate the effectiveness and superiority of the proposed method when compared with some state-of-the-art methods.In order to further improve the performance of traffic sign detection and recognition,this article starts from the nature of the detection and recognition tasks.Considering the significant differences between the two tasks,a two-stage network is proposed,which separately adopts a detection network and a recognition network to deal with two sub-problems.A feature aggregation module is designed and applied to the two sub-networks.The feature aggregation module not only has the function of traditional down-sampling,that is,reducing the size of the input feature map,but also can alleviate information loss as far as possible.On this basis,the cross-stage connection method is used to increase the diversity of the gradient information for the detection network.The spatial transformer network is adopted to correct the output,and then it is sent to the recognition network to obtain specific category information.The experimental results on the public dataset demonstrate that the performance of the twostage network proposed in this article is better than the current mainstream network.