DamViT: A Multi-scale Vision Transformer for Damage Detection in Structural Health Analysis of Concrete Dams

Abstract

Structural health analysis (SHA) of dams and water reservoirs is essential for ensuring their long-term safety, durability, and operational reliability. Conducting timely assessments, however, remains challenging due to practicing visual inspection methods are prone to persistent moisture, efflorescence, complex illumination conditions, and natural occlusions such as vegetation. These factors significantly reduce the robustness and consistency of automatic damage detection. To overcome these limitations, we propose a novel multi-scale Dam Vision Transformer (DamViT) architecture specifically designed for robust and automated damage detection and segmentation. DamViT utilizes a dual-branch framework: a Coarse branch that efficiently identifies potential damage regions by capturing global context and a Fine-grained branch that performs high-precision segmentation and delineation of specific defects, namely cracks and spalling, by extracting localized, high-resolution features. To facilitate rigorous development and evaluation, we meticulously annotated and compiled a high-quality dataset comprising 5,000 high-resolution image patches (512 x 512 pixels) of dam surfaces. Thorough Experimentation on the test set demonstrates the superior performance of the proposed approach. For damage detection, DamViT achieves mAP of 0.92 for cracks and 0.95 for spalling, and for segmentation, it attains mAP scores of 0.72 and 0.80, respectively. These results outperform well-established approaches (e.g., U-Net, DeepLabV3+, DeepCrack, CrackFormer) and recent YOLO variants. The proposed DamViT advances the state-of-the-art in SHA by enabling timely and reliable damage detection for cost-effective long-term infrastructure maintenance.

Keywords: Structural Health Analysis, Vision Transformer, Automatic Damage Detection, Cracks and Spalling