Lightweight Road Traffic Sign Detection Based on Ghost-Net

The autonomous driving system relies heavily on the detection of traffic signs. In this paper, we propose an enhanced lightweight model based on YOLOv8 to address the issues of excessive model parameters, high computational cost, and inadequate small-target detection performance in traffic sign identification tasks. The model parameters and computational expenses are greatly decreased by substituting traditional convolutional layers with GhostNet modules in the back-bone network. The redundant P5 large-object detection head is eliminated to maximize com-putational efficiency, and a specialized P2 small-object detection layer is added to improve high-resolution feature extraction capabilities. Additionally, the feature fusion step incorporates a parameter-free SimAM attention method to dynamically improve feature weight allocation, increasing the accuracy of detection. The enhanced model achieves a higher mean Average Precision (mAP@0.5) of 89.1% compared to the baseline model's 87.6%, while reducing pa-rameters from 11.1M to 3.6M (67.57% reduction) and computational costs from 28.5 GFLOPs to 20.9 GFLOPs (26.67% reduction), according to experimental results on the TT100K dataset.