Traffic sign detection remains challenging because of dense small objects, complex backgrounds, occlusion, and limited edge resources. This letter proposes MMS-Net, integrating multi-scale semantic residual modeling, same-scale cross-semantic attention fusion, and a semantic-enhanced dynamic detection head for high-level semantic enhancement, discriminative feature fusion, and adaptive prediction. Under YOLO validation on TT100K, MMS-Net achieves 83.1% mAP@0.5 and 64.7% mAP@0.5:0.95 with 8.8 M parameters and 17.8 G FLOPs. Additional traffic-sign datasets show competitive accuracy at lower complexity than YOLOv12s, while Jetson Orin experiments support real-time deployment under the evaluated configuration.
MMS-Net: Multi-Granularity Semantic Enhancement for Efficient Traffic Sign Detection
Read on IEEE Xplore
