Surface crack edge-aware segmentation using hierarchical transformer encoder and Sobel edges for aerial inspection

Abstract

Structural crack detection remains a key component of structural health monitoring for large-scale civil infrastructure, where manual inspection is costly and hazardous. While aerial imaging enables rapid data acquisition, reliable crack analysis is limited by the difficulty of obtaining high-resolution images at close range without risking UAV damage. To address this, we introduce a UAV equipped with a protective spherical shell that allows safe surface contact. Although CNN-based architectures such as FPN and U-Net are established for crack segmentation, their encoders struggle to model long-range dependencies and fine edge structures, which result in poor delineation of thin, low-contrast, and discontinuous cracks. This limitation remains a persistent research gap in aerial crack analysis. To address this gap, this study investigates two complementary strategies: (i) replacing conventional CNN encoders with hierarchical Transformer backbones to capture global contextual relationships, and (ii) introducing edge-aware input augmentation using luminosity-based grayscale and Sobel gradient maps to optimize boundary sensitivity. We conducted experiments on our CrackUAS dataset, which consists of aerial images collected from infrastructure inspection scenarios. Integrating a hierarchical Transformer encoder with Sobel edge augmentation yields consistent performance gains across architectures. The ablation study concludes that F1-scores increase by approximately 33% for U-Net and 13% for FPN relative to their baseline configurations. These results show that integrating Transformer-based feature encoding with edge-aware input augmentation improves aerial crack segmentation.