Generative adversarial network for real-time identification and pixel-level annotation of highway pavement distresses

Abstract

Efficient analysis of road pavement distresses is crucial for infrastructure management and safety. Traditional methods are labor-intensive, and recent deep-learning approaches face challenges such as overlapping bounding boxes and poor pixel localization. This paper presents PaveGAN, a real-time method to identify and annotate pavement distresses using generative adversarial networks. While trained on several loss functions, PaveGAN achieved its best mean absolute percentage errors of 1.49%, 1.15%, and 3.26% for the mean squared error, Huber loss, and structured similarity index, respectively, at the pixel level. With intersection over union scores exceeding 90% for several distresses, and operating at 89.15 frames per second, PaveGAN is suitable for real-time applications. Independent evaluations show its annotation accuracy aligns closely with human experts, achieving 84% accuracy versus 86% for manual annotations. By automating pixel-level annotation in widely used formats such as COCO and LabelMe, PaveGAN provides a scalable, cost-effective solution for pavement distress monitoring.