Improved Algorithm for Small Target Detection of Traffic Signs on YOLOv5s

Aiming to address the issues of small target size, complex background, and low recognition accuracy of traffic signs in road scenes, an improved YOLOv5s algorithm is proposed. The algorithm introduces a Coordinate Attention in the feature fusion network, enhancing the model's ability to integrate spatial coordinate information and preserve a wider range of positional information. The Efficient Decoupled Heads are used to replace the Detection Heads in YOLOv5s, with an additional layer for small object detection, improving the model's generalization ability, robustness, and utilization of small object information. The Normalized Wasserstein Distance (NWD) is introduced to mitigate the sensitivity of IoU to small target position deviations. The algorithm is trained on the TT100K dataset with data augmentation and enhancement. Experimental results demonstrate that the improved algorithm achieves a 4.4% increase in mAP@0.5 compared to the original algorithm.