A comparative analysis of computer vision techniques used for the Classification and Detection of Pavement Surface Distresses in Pavement Management Systems

Abstract

This dissertation presents a comparative analysis of computer vision techniques for the classification and detection of pavement surface distresses, focusing on their application within pavement management systems (PMS). The study evaluates four deep learning models—MobileNetV3, ResNet50, VGG-16, and YOLOv8—based on their accuracy, precision, recall, F1-score, training time, and inference time. The results indicate that MobileNetV3, with its shorter training and inference times, is suitable for real-time applications, enhancing the frequency of pavement condition monitoring. ResNet50, despite its higher computational demands, provides high-accuracy detection, making it ideal for detailed inspections. VGG-16 offers a balance between accuracy and efficiency, suitable for routine assessments. YOLOv8 excels in real-time detection, making it ideal for immediate on-site inspections. The integration of these models into PMS can improve the quality and frequency of pavement condition data, supporting more informed and timely maintenance decisions. The study concludes with recommendations for the deployment of these models based on specific PMS needs, highlighting their potential to enhance the sustainability and safety of road infrastructures.