Research On Traffic Sign Detection And Recognition Based On Convolutional Neural Network

In recent years,Internet technology and deep learning technology have been developed to varying degrees,and more and more people have applied related technologies to practical application scenarios.Intelligent Traffic System（ITS）is to combine and apply advanced technologies such as Internet technology and artificial intelligence to transportation,so as to make transportation more intelligent.As an important research direction in ITS,the detection and recognition of traffic signs has extremely practical significance in assisting driving and obtaining road condition information.In practical application scenarios,how to detect and identify traffic signs in road streetscape more timely and accurately has been the focus and difficulty of the study.After many research by scholars in China and abroad,compared with traditional detection and recognition methods,the method based on convolutional neural network has a good processing speed.It has obvious advantages in both recognition accuracy and real-time performance,and has good research value.In view of the shortcomings of the current YOLOv5 network,such as too many model parameters in the training process and too many missed detections and false detections in detecting small targets,this paper is based on the YOLOv5 network.The following aspects of research and related improvements:Firstly,the advantages and disadvantages of the four YOLOv5 networks are analyzed.The YOLOv5 x network in YOLOv5 is used as the basic network model,and the light-weight Ghost Bottleneck network（ghost network）is introduced to replace the original Bottleneck CSP network structure.The nonlinear transformation of some characteristics of the original network is changed into a linear transformation,which decreases the number of parameters and calculation of the model,and improves the feature extraction part（Backbone part）of YOLOv5 x.Then,the feature compression method is used as the improvement point to improve the channel attention.Firstly,the global pooling and average pooling operations are performed on the features in the channel dimension,and the features obtained by the respective operations are spliced in the channel dimension.Finally,the channel attention mechanism after the improved feature compression method is introduced into the Backbone to ensure that the effective features of small target are enhanced during feature extraction and the interference of irrelevant features is suppressed.The experimental results show that the improved YOLOv5 x network reduces the model volume,and at the same time,it can realize the rapid detection and classification of 43 types of traffic signs in the GTSDB data set,with good accuracy.However,there is still room for improvement for the problems of missed detection of small targets and error recognition between similar signs.Secondly,a single fine classification network EfficientNet network is improved.CBAM（Convolutional Block Attention Module）attention mechanism is introduced before each downsampling of EfficientNet network to improve the network′s attention to important feature maps and important spaces of traffic signs.On the basis of the original classification accuracy,it has been significan tly improved.Finally,based on the original improved YOLOv5 network,the improved EfficientNet network is cascaded to detect and classify traffic signs.Firstly,the YOLOv5 network is used for rapid detection of traffic signs.The detected traffic signs are roughly classified into only of the four categories（indicator,warning,prohibition and other）,and then the detected four categories of traffic signs are input into the cascade EfficientNet network for specific fine classification of 43 categories.Compared with the previous single YOLOv5 network,the cascade network has a slight increase in the model volume,but the detection accuracy of the street traffic sign dataset has been significantly improved,which can realize the rapid detection and classification of traffic signs.