Research on application of lightweight AF-YOLO model in traffic sign recognition

To address the current challenges in traffic sign detection, including low accuracy, large model size, and slow speed, a lightweight network model, AF-YOLO, is proposed. The model is built upon the YOLOv5 architecture. Firstly, the network backbone is reconstructed using the C3-SE module, which combines deep separable convolution with the SE channel attention mechanism. The former reduces the model's parameter count and computational burden, while the latter enhances its expressiveness and generalization capability. Additionally, an Asymptotic Feature Pyramid Network (AFPN) is introduced to improve feature fusion capabilities. The Wise-IoU loss function is adopted in place of the original CIOU loss function, and category weights are incorporated to improve detection performance on datasets with imbalanced class distributions. The TT100K dataset is then processed, selecting 42 categories with relatively stable image numbers. Experimental results indicate that, in comparison to the original YOLOv5s model, the proposed approach decreases the parameter count by 21.79% and reduces computational complexity by 3.73%, improves average precision by 1%, and increases detection speed by 16 frames per second, achieving a well-balanced trade-off between accuracy, model size, and speed.