Synergising Heterogeneous Features and Guided Attention for Small Traffic Sign Detection in Complex Scenes

ABSTRACT Traffic sign detection is a pivotal technology in the Intelligent Transportation System. Existing approaches suffer from low accuracy in complex scenes and high miss rate for small objects. To this end, this paper proposes a novel approach called ETSD to detect traffic signs based on enhanced features and attention. First, a heterogeneous feature enhancement block is designed to strengthen the network's capacity to capture fine‐grained features of small objects in both channel and spatial dimensions. Second, to resolve the challenges of local information degradation and global semantics ambiguity in deep networks, we introduce the selective guided attention to strengthen global semantics through 3D‐permutation modelling while preserving local details via neuron modelling. Finally, to further enhance the detection performance for small objects, we constructed a dedicated detection layer that extracts low‐level features of small objects and fuses them with deep network features. ETSD is rigorously evaluated on four benchmark datasets (CCTSDB2021, TT100K, GTSDB and RoadSign). It consistently surpasses YOLO11n, achieving a notable gain of 5.2% in mAP50 and 4.5% in mAP50:95 on CCTSDB2021. Moreover, its mAP50:95 is improved by 4.5%, 2.0% and 3.5% on TT100K, RoadSign and GTSDB, proving its effectiveness and robustness.