Highway traffic sign target detection based on MS-YOLOv8s

Aiming to address the inefficiency of existing target detection algorithms in detecting small-sized traffic signs on highways, this study proposes a small-target traffic sign detection algorithm named MS-YOLOv8s based on multi-scale feature fusion. Firstly, to improve the accuracy of small-target detection, an upsampling layer is added to the Neck part of YOLOv8, along with an improved bidirectional adaptive feature pyramid network (M-BiFPN), and the introduction of the P2 small target detection layer. This approach aims to fully utilize deep and shallow features for multi-size feature fusion and comprehensive feature information extraction. Secondly, to mitigate the issue of duplicate detection frames caused by Non-Maximum Suppression (NMS), a softening strategy improvement method called Soft-NMS is introduced. Soft-NMS gradually reduces the score of original frames as the overlap degree increases, effectively suppressing the generation of redundant frames while maintaining a high recall rate. Experimental results on the Tsinghua-Tencent 100K (TT100K) dataset demonstrate that the improved model achieves a mean average precision (mAP) of 83.7%, which is 2.32% higher than that of YOLOv8s. Additionally, the number of parameters is approximately \(10.14 \times {10}^6\), representing a 4.53% increase over YOLOv8s, thus achieving the objective of detecting fewer parameters with higher precision.