A Data-Efficient Deep Learning Paradigm for Crack Segmentation Using Generative AI

High manual annotation costs and poor generalizability often limit the robustness of deep learning models for pavement crack segmentation. This study addresses these challenges by developing and validating a diffusion-based generative artificial intelligence framework to support data-efficient crack segmentation models for infrastructure inspection. The proposed framework formalizes the generation and utilization of synthetic, pixel-level annotated crack data and is empirically evaluated across multiple learning paradigms. To implement and validate the framework, an annotated synthetic dataset (AI500) was generated using Google’s Imagen 4 Ultra and designed to capture diverse crack morphologies, lighting conditions, and surface textures. The utility of the synthetic data was systematically assessed through three experimental strategies: synthetic-only training, hybrid dataset training, and sim-to-real transfer. Model performance was evaluated using three deep learning architectures and five real-world datasets. Results show that synthetic-only models provide a robust, transferable performance baseline, achieving cross-domain generalization comparable to the target-only baseline (TOB) for models trained on real-world datasets. Augmenting AI500 with only 10% real-world data yielded performance comparable to TOB. Furthermore, sim-to-real transfer using the same 10% fraction achieved near parity with TOB. Cost analysis revealed the potential for double-digit savings without compromising model robustness. Additionally, the development of G3P-500 using Gemini 3 Pro demonstrated superior photorealism and diversity, requiring minimal human verification, with only 3.2% of pixels added to generated annotations. These findings demonstrate that high-quality synthetic data within a structured generative framework can substantially reduce annotation requirements while enabling accurate, generalizable pavement crack segmentation models. However, further validation remains necessary in more challenging environments.