A framework for multi-class surface damage segmentation, quantification, and severity assessment of reinforced concrete structures with GAN data augmentation

Concrete surface damage is a key indicator of deterioration in reinforced concrete (RC) structures, directly affecting durability, serviceability, and structural safety. Accurate detection, segmentation, and rapid severity assessment are therefore essential for efficient inspection and maintenance planning. This study proposes an end-to-end framework that integrates multi-class surface damage segmentation, geometric quantification, and automatic severity grading of RC structures, moving beyond defect identification toward engineering-oriented damage evaluation. To address class imbalance in damage datasets, a generative adversarial network (GAN) is employed to synthesize realistic samples of underrepresented damage types, thereby enhancing data diversity and model robustness. A lightweight attention-enhanced segmentation network is subsequently trained to perform pixel-level detection of multiple surface damage categories. Based on the extracted geometric information, a classification-driven module converts pixel-level measurements into engineering-relevant severity levels by estimating crack width and spalled area, enabling rapid damage grading of RC members. Experimental validation on benchmark datasets demonstrates that the proposed segmentation network achieves a mean Intersection over Union (mIoU) of approximately 4.07%, outperforming representative baseline models. Furthermore, automated crack-width estimation yields a mean absolute error of 0.061 mm and a mean relative error of 6.76% compared with manual measurements, with fully consistent severity grading results. Grad-CAM visualization confirms that the model focuses on structurally meaningful regions consistent with human inspection behavior. By integrating data augmentation, automated damage quantification, the proposed approach establishes a complete pathway from raw inspection images to member-level damage grades and code-compliant severity assessment, providing a practical and deployable solution for AI-assisted surface damage inspection and condition assessment of RC infrastructure in real-world scenarios. © 2026 The Author(s).