AAH-YOLO: an adaptive attention-enhanced method for small-scale traffic sign detection in degraded imaging scenes

Abstract

Traffic sign detection under complex visual conditions remains a critical challenge for intelligent transportation imaging systems. Mainstream detection models have three key limitations: insufficient sensitivity to small-scale objects in captured images, poor adaptability to multiscale objects at varying imaging distances, and inadequate robustness to image quality degradation from adverse weather and illumination variations. To tackle these challenges, this paper proposes AAH-YOLO, an improved YOLOv8s-based deep learning detection method for high-precision traffic sign detection in complex, variable conditions. The method improves the network architecture in three ways. First, a spatial pyramid pooling fusion module integrated with the adaptive fine-grained channel attention mechanism is built to enhance critical feature capture via dynamic feature weight optimization. Second, an adaptive feature enhancement module is embedded into the C2f module to boost adaptability to complex traffic scenes and multiscale objects via multibranch dynamic fusion. Third, a high-resolution small object detection head is added to optimize localization of ultra-small objects (below 8×8 pixels) in long-distance traffic imaging. Verified on the augmented TT100K dataset with six imaging degradation types (including adverse weather, motion blur, noise, and illumination variations), experimental results show 89.3% mAP@0.5 (3.6% higher than YOLOv8s baseline) and 82.1 FPS inference speed, meeting real-time requirements. The proposed method provides reliable technical support for traffic sign perception in electronic imaging systems for intelligent transportation and autonomous driving.