A New Method for Extracting Typical Diseases of Highway Asphalt Pavement in High-Resolution Images Based on YOLO Network

The exponential expansion of global highway infrastructure has intensified challenges in road inspection and maintenance operations, creating substantial demand for advanced automated interpretation systems. Current methodologies for detecting typical asphalt pavement distresses in high-resolution imagery demonstrate critical limitations in processing efficiency and detection accuracy, failing to meet contemporary requirements for highway maintenance. This study presents an innovative YOLO-based framework for automated detection of asphalt pavement pathologies in high-resolution images. Utilizing high-resolution pavement distress imagery acquired through specialized linear array camera systems as primary data sources, we implemented precise annotation of pavement deterioration patterns using LabelMe image annotation software. This process generated comprehensive data labels detailing spatial distributions and morphological characteristics of asphalt pavement pathologies. Through strategic application of image segmentation enhancement algorithms and standardized graphic segmentation techniques, we developed two specialized datasets: a Multi-Scale Multi-Disease (MSMD) dataset and a Single-Disease Semantic Segmentation (SDSS) dataset. The experimental design incorporated a Stratified-Shuffle-Splitting methodology to ensure non-repetitive, stratified random partitioning of datasets. This approach maintains feature distribution consistency between training and testing subsets while guaranteeing complete data segregation. Meanwhile, inter-rater reliability for the sample set was examined using the kappa coefficient. We conducted comparative analysis of five successive YOLO iterations (v8-v12), implementing progressive multi-scale incremental training protocols. Empirical results demonstrate that our incremental learning strategy enables continuous model performance enhancement, with YOLOv11 emerging as the optimal architecture for automated pavement distress detection. The YOLOv11 implementation achieved state-of-the-art performance metrics: an F1-score of 0.86 at confidence threshold 0.497, coupled with precision-recall metrics reaching 0.874 mAP@0.5. These results satisfy practical engineering requirements for automated pavement assessment, showing advantages in computational efficiency and detection accuracy. The proposed framework demonstrates significant potential for revolutionizing current pavement maintenance practices through AI-powered automated interpretation systems.