Subgrade micro-crack detection based on an attention-enhanced YOLO algorithm

Abstract

Under freeze–thaw cycles in seasonally frozen soil regions, road subgrade surfaces are prone to develop micro-cracks that gradually evolve into reticulate structures. Traditional detection methods rely on manual inspection, which is inefficient and incapable of identifying early-stage micro-cracks. This paper proposes an intelligent detection method for subgrade micro-cracks based on an attention-enhanced YOLO algorithm. First, freeze–thaw cycle tests are conducted to obtain real images of the subgrade surface crack development process, and a dataset containing micro-crack annotations is established with image preprocessing and augmentation. Then, a coordinate attention mechanism is embedded into the YOLO backbone network, and an improved network architecture along with a model training strategy is designed. Ablation experiment results show that the attention-enhanced YOLO algorithm achieves a mean average precision of 91.7% on the subgrade micro-crack detection task, which is 6.4 percentage points higher than the baseline model, while the miss rate is reduced to 8.2%. The improved algorithm exhibits excellent detection accuracy and robustness under complex freeze– thaw scenarios, providing a feasible engineering solution for non-contact intelligent inspection.