Automated multi-type damage detection framework in reinforced concrete structures via data augmentation and deep segmentation networks

Abstract

Accurate and automated detection of multiple damage types in reinforced concrete (RC) structures is vital for ensuring long-term structural safety and operational performance. Despite recent advances in deep learning, the effectiveness of semantic segmentation algorithms is often constrained by limited and imbalanced training datasets, especially for complex and less common damage types such as rebar corrosion and concrete crushing. This study proposes a comprehensive damage detection framework that synergistically integrates classical data augmentation techniques (e.g., flipping, cropping, brightness/contrast adjustments) and generative augmentation via StyleGAN2 to enrich training data and mitigate class imbalance. A curated and annotated dataset containing 2,119 images across five RC damage categories—including cracks, spalling, rebar exposure, rebar corrosion, and concrete crushing—was used to train the DeepLabv3 + segmentation model. Through comparative analysis of 22 augmentation strategies, results show that a triple-combination strategy with Generative adversarial network (GAN) integration achieved the best performance, with average mIoU improved by 21.8% and mean F1-score exceeding 82.7%, compared to the baseline model. Additionally, the proposed framework enables pixel-level damage size quantification, with validated performance on both isolated and multi-damage test images. Experimental findings reveal that excessive augmentation complexity may introduce feature bias, whereas carefully balanced combinations enhance both generalization and category-level sensitivity. This study not only provides actionable insights into optimizing data augmentation for multi-damage detection, but also lays the groundwork for practical deployment of AI-based structural health monitoring systems capable of delivering quantitative, real-time damage assessment in the field.