Automatic Traffic Sign Recognition Algorithm Based on Attention Mechanism and YOLOv4

Abstract

Image recognition, a key technique in deep learning within the realm of computer vision, has found extensive application in the transportation sector in recent years. However, traditional image recognition technologies suffer from low efficiency and weak analytical capabilities. This research proposes a traffic sign recognition model that embeds a reconstructed squeeze and excitation network channel attention mechanism into the You Only Look Once Version 4 framework. Specifically, depthwise separable convolution is adopted to reconstruct the attention module, and a soft threshold denoising module is integrated before multi-scale feature fusion. The model also utilizes a soft threshold denoising module for feature extraction of complex semantic information. Experimental results show that when the attention mechanism fusion algorithm iterates five times, the accuracy reaches 98.5%. The highest recognition accuracy, prediction recall rate, and harmonic mean of recall rate are 96.35%, 95.88%, and 95.12%, respectively. The evaluation of the fusion model shows that the model has the highest recognition accuracy of 0.98 for different types of traffic signs. Compared with the highest accuracy of 0.89 of Faster Region-based Convolutional Neural Network and the highest accuracy of 0.90 of Field-Programmable Gate Array, the research method has significantly higher recognition accuracy. These results suggest that the improved traffic sign recognition model can effectively identify real-world road traffic signs for autonomous vehicles, with excellent feature-capturing performance. This research contributes to the future development of road traffic and autonomous driving fields.