Research On Traffic Sign Recognition Technology Based On Neural Networks

Traffic signs are used to guide drivers in traffic scenarios.Real-time recognition technology of traffic signs is essential for advanced driving assistance systems to provide drivers with information about the road ahead.However,different weather conditions such as strong or weak light can affect the accuracy of the recognition of traffic signs.To address the low precision issue of YOLOv5 in multi-class classification,an improved model for traffic sign recognition based on a revised YOLOv5 s is proposed in this article:Firstly,the Backbone and Detection Head of YOLOv5 s were optimized.The Transformer module was used in the backbone network to increase the receptive field and better extract network features.Meanwhile,the Swin Transformer module was introduced in the Head Detection for mobile window support and reduced computational complexity.Secondly,functions were modified,and better interpolation algorithms were applied using bilinear interpolation instead of nearest neighbor interpolation to improve the quality of scaled images.F-EIOU non-maximum suppression was used for optimization to improve detection efficiency and accuracy.Finally,based on the above main work,experiments were conducted to compare the modifications in each part.Finally,the improved YOLOv5 s algorithm was compared with YOLOv5 and YOLOv3 algorithms.The results show that compared with YOLOv3 and YOLOv5 algorithms,the improved model has higher accuracy in detection.Under natural conditions,the m AP values for traffic sign detection were 0.838,0.799,and 0.6573,respectively.The experiment shows that the use of improved YOLOv5 s increases the average precision by 3.9%,meeting the requirements for real-time detection.