| Summary: | Amid the era of intelligent construction and inspection, traditional object detection models like YOLOv8 struggle in bridge defect detection due to high computational complexity and limited speed. To address this, the lightweight SATH–YOLO model was proposed in this paper. First, the Star Block from StarNet was used to build the STNC2f module, enriching semantic information and improving multi-scale feature fusion while reducing parameters and computation. Second, the SPPF module was replaced with an AIFI module to capture finer-grained local features, improving feature-fusion precision and adaptability in complex scenarios. Lastly, a lightweight TDMDH detection head with shared convolution and dynamic feature selection further reduced computational costs. With the SATH–YOLO model, parameter count, computation, and model size were reduced significantly by 39.9%, 8.6%, and 36.2%, respectively. Meanwhile, the average detection precision was not impacted but improved by 1%, which meets the demands of edge devices and resource-constrained environments.
|
|---|
