Research on Bayesian optimization-based YOLOv5 method for pavement distress detection

To address the inefficiency of hyperparameter tuning and poor scene adaptability of the YOLOv5 model in asphalt pavement distress detection, this study proposes an automated hyperparameter optimization method based on Bayesian optimization. First, a cross-regional dataset, Road-D, is constructed, comprising 5,800 annotated images of six distress categories under varying illumination and wear conditions. Second, a multi-objective optimization function integrating detection accuracy and inference efficiency is designed. A Gaussian process-based probabilistic surrogate model is established, and parameter sensitivity analysis is employed to compress the search space dimensionality, enabling adaptive hyperparameter optimization. Finally, a parallelizable Bayesian optimization framework is developed to conduct a global search for optimal hyperparameter combinations in the YOLOv5 model. Experimental results demonstrate that the optimized model achieves a mean Average Precision ([email protected]) of 88.1% on the Road-D test set, a 6.2% improvement over the baseline model, with a per-image inference time of 15.3 ms. Compared to traditional grid search, Bayesian optimization reduces tuning time by 65%. This method significantly enhances the accuracy-efficiency trade-off in pavement distress detection, providing reliable technical support for intelligent road maintenance.