Improving Pavement Distress segmentation with Diffusion-Based Generative AI

Abstract

The lack of diverse and sufficiently labeled datasets in pavement management limits the performance and generalization of segmentation and detection models. Existing datasets often fail to capture real-world variability, such as different pavement distress types, leading to models that struggle with unseen scenarios and rare distress patterns. This study addresses these challenges by applying diffusion-based generative AI to mitigate data scarcity. Using a Denoising Diffusion Probabilistic Model (DDPM), synthetic images were generated to augment datasets, significantly enhancing model robustness and accuracy. The synthetic images achieved an SSIM of 0.92 and a diversity score of 0.85, resulting in notable improvements in the U- Net segmentation model, including an IoU increase from 0.81 to 0.93, an F1 Score rise from 0.85 to 0.95, and pixel-level accuracy improvement from 0.89 to 0.95. These findings underscore the potential of generative AI to support scalable and robust digital twin solutions for pavement management systems.