| Traffic sign recognition is a key technology in intelligent transportation systems,and its accuracy is directly related to the safety of intelligent transportation.Traditional traffic sign recognition methods can no longer meet the high-performance requirements of traffic sign recognition,while traffic sign recognition methods based on deep learning have advantages such as good recognition effect and faster speed.However,traffic sign images in natural scenes have problems such as complex backgrounds,susceptibility to adverse weather,lighting conditions,white noise,and deformation of traffic sign images due to poor shooting angles.As a result,existing traffic sign recognition algorithms based on deep learning in natural scenes have shortcomings such as low accuracy,weak generalization performance,and low robustness.Based on this,this thesis proposes two new traffic sign recognition models based on deep learning methods from the perspectives of severe weather and complex scenes.(1)Traffic sign recognition model based on DCV-YOLOXAiming at the problem that traffic sign image information is susceptible to extreme adverse weather,such as fog and rain,this thesis proposes a new DCV-YOLOX traffic sign recognition model.Firstly,the model improves the data enhancement method at the input of YOLOX network to enhance the generalization performance of the model;Secondly,a Dense RFB-I receptive field enhancement module is added to its backbone network to enrich context information and improve feature extraction capabilities;Thirdly,a CS-CA attention mechanism module is added to the neck network to suppress interference information and enhance the effect of feature fusion;Finally,at the network prediction end,the use of a Varifocal Loss function to focus on positive sample information improves model accuracy and accelerates model convergence.The experimental results show that the m AP50 and m AP50:5:95 indicators of the DCV-YOLOX model on the TT100 K dataset reach91.0% and 73.3%,respectively,with a reasoning speed of 10.5 ms/frame.(2)Traffic sign recognition model based on Swin-Bi FE-YOLOv5Aiming at the problem that traffic sign image information in complex scenes is easily affected by lighting,white noise,and deformation of traffic sign images due to poor shooting angles,this thesis proposes a new Swin-Bi FE-YOLOv5 traffic sign recognition model.This model is based on the YOLOv5 model,incorporating the strong ability of the Swin Transformer model to extract global information,and using the Swin Transformer Block to replace the last CX module in its backbone network;Secondly,in the neck network,Bi FPN structure is used to replace the original FPN-PAN structure to enhance the effect of feature fusion;Finally,the Focal-EIo U loss function is used at the network prediction end to avoid penalty item failures and focus on difficult samples,further improving the accuracy of the model.The experimental results show that the m AP50 and m AP50:5:95 indicators of the Swin-Bi FE-YOLOv5 model on the CCTSDB dataset reach 97.4% and 84.9%,respectively,with a reasoning speed of only 8.1 ms/frame.The two new traffic sign recognition models established in this thesis can effectively solve the problems of low accuracy,weak generalization performance,and low robustness in traffic sign recognition in natural scenes,providing a good solution for high-performance traffic sign recognition. |
