YOLOv5 Traffic Sign Detection Incorporating Self-Attention Mechanisms
Yuanbin Zhu, bin Liu · 2025
This paper presents an enhanced YOLOv5 model for traffic sign detection by integrating a dual-branch channel attention mechanism (global semantic + 3×3 convolutional local texture) and spatial attention. The channel attention module employs parallel pathways to jointly enhance color semantics (e.g., yellow warning signs) and texture details (e.g., triangular borders, arrow contours), while spatial attention refines spatial localization of signs. On the CCTSDB2021 dataset, the improved model achieves 98.0% [email protected] (+1.5% vs. baseline), 97.0% recall for warning signs (+3.5%), and 92.3% detection rate for small targets (pixel ratio <0.1%, +3.8%). With a frame rate of 91 FPS, it maintains real-time performance while demonstrating robust detection in challenging scenarios, such as strong illumination, backlight, and rainy/foggy conditions. Experimental results validate that the dual-attention mechanism effectively suppresses background interference and improves feature discriminability for low-resolution and occluded signs.