Autonomous Synthetic Data Generation for Asphalt Pavement Crack Segmentation Using Generative Models

Abstract

This study addresses the challenge of limited annotated datasets in pavement crack analysis by integrating synthetic data generation with automated segmentation models. The Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) was utilized to generate realistic synthetic pavement images, and the U-Net model was used to enhance segmentation accuracy. Current datasets, such as Crack500, lack sufficient size and diversity, limiting model performance. By augmenting the dataset with synthetic images in increments of 20% to 2000%, segmentation performance improved significantly. The superior model, trained with a 2000% increase in synthetic data, achieved an Intersection over Union (IoU) score of 0.92, demonstrating the effectiveness of data augmentation in improving segmentation. Its robustness powers were shown through validation on myriad real-world datasets, indicating good performance across different crack patterns and imaging necessities. These findings highlight the potential of combining generative and segmentation models to address data scarcity in pavement management.