Advanced Crack Detection in Reinforced Autoclaved Aerated Concrete Using Generative Data Augmentation and Enhanced Segmentation

Abstract

Cracks in reinforced autoclaved aerated concrete (RAAC) pose significant structural risks, including water ingress, corrosion of reinforcement, and, in extreme cases, potential collapse. The challenge is worsened by the lack of accessible RAAC crack data, making it difficult to develop accurate and consistent detection frameworks. This limitation restricts the ability to perform timely interventions and implement effective maintenance strategies for RAAC cracks. Therefore, the study aims to assess the impact of data augmentation by using StyleGAN3, one of the generative adversarial networks, to address limited RAAC crack data. Furthermore, advanced convolutional neural network architectures were explored for improved semantic segmentation of RAAC cracks. Results revealed that StyleGAN3-generated data augmentation boosted model performance, and the newly developed RAAC-UNet++ model markedly enhanced segmentation accuracy. These findings offer valuable insights for improving RAAC crack detection, ultimately aiding in the effective maintenance and management of RAAC structures.