Two-phase Learning for Classifying Road Damage using YOLO Architecture

The rapid growth in transportation demands emphasizes the urgent need for effective maintenance of road infrastructure, particularly in regions with substantial road damage. Traditional methods of manual inspection and multifunctional vehicles are often inefficient, underscoring the necessity for advanced solutions. This study introduces a deep learning-based approach using YOLOv8 to enhance road damage identification, focusing on various damage types including longitudinal cracks, transverse cracks, alligator cracks, and potholes. The dataset, consisting of 4,020 images collected from Kaggle, Google Street View, and the Ministry of Public Works and Housing, was meticulously annotated and underwent comprehensive preprocessing. Techniques included auto-orientation, resizing, contrast adjustment with adaptive equalization and CLAHE, as well as augmentation strategies such as horizontal flip, rotation, exposure, blur, noise, zoom, and brightness adjustment. The YOLOv8 models, both small and medium, were trained sequentially: initially on two damage types to establish a performance baseline, followed by training on all four types to refine the model. The study also explored three testing scenarios: variations in image resolution (416, 640, and 800 pixels), hyperparameter tuning (batch sizes of 16 and 32 and learning rates of 0.001 and 0.00001), and optimization through cascading training on extended datasets. Performance analysis demonstrated that the Adam optimizer significantly outperforms the SGD optimizer, achieving an F1-Score of 0.61 and a mean Average Precision (mAP) of 0.60. This research underscores the impact of advanced preprocessing and augmentation methods in improving model accuracy and presents the YOLOv8-based approach as a promising solution for efficient road damage detection, providing valuable insights for future research and practical applications in road maintenance and infrastructure improvement.