MSCRep-YOLO: A new method for traffic sign recognition

This paper primarily addresses the challenges of small-scale traffic signs and real-time requirements in traffic sign recognition. We propose an improved YOLOv8n-based algorithm that incorporates a small object detection layer, a multi-scale convolutional attention mechanism, and a lightweight RepViT architecture, effectively enhancing detection capabilities for smallsized traffic signs while maintaining real-time performance. Additionally, the MPDIoU loss function replaces the Complete Intersection over Union loss function to further improve boundary regression performance. Tested on the TT100K dataset, the model achieves an 6.3% increase in precision, a 7% increase in recall, and an 7% increase in mean average precision, with only 78.82% of the original model’s parameters. These results demonstrate the model’s practical value in real-time traffic sign recognition tasks, especially in handling small-scale signs.