Dugdale-GAN: Physical Dugdale Model Integrated Generative Adversarial Network for High-quality Crack Image Generation

Vision-based crack detection systems face significant challenges in training data scarcity and imbalance. While Generative Adversarial Networks (GANs) offer potential solutions, existing methods often generate physically implausible patterns. We propose Dugdale-GAN, a novel framework that integrates the physical Dugdale model with GANs to generate realistic crack patterns. Our approach incorporates material-specific parameters and features a dual-generator and dual-discriminator architecture to ensure both physical consistency and visual quality. We also introduce the Diversity Quality Indicator (DQI), a comprehensive metric for evaluating both image quality and pattern diversity. Experimental results demonstrate that Dugdale-GAN outperforms existing methods in generating physically plausible crack patterns.