.png){kind=icon}
.png){kind=icon}
Return
BladeSplat: high-fidelity surface reconstruction of wind turbine blades with structure-guided Gaussian splatting
.png){kind=icon}
PRE
AI
Z
Z
X
F
DOI：10.1016/j.aei.2026.104771.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Surface damage assessment of large-scale industrial assets often relies on high-fidelity 3D geometric models. Wind turbine blades represent a typical example due to their large size and susceptibility to surface defects during long-term operation. However, the large-scale and low-texture characteristics of blade surfaces pose significant challenges for conventional 3D reconstruction methods, frequently leading to blurred edges, loss of fine surface details, and geometric distortions that hinder reliable automated inspection. To address these limitations, a novel structure-guided Gaussian Splatting framework, termed BladeSplat, is proposed. In this framework, explicit geometric priors and structural constraints are systematically integrated into the optimization process of 2D Gaussian Splatting. First, a mask-guided structure-point initialization strategy is introduced to enrich spatial optimization primitives within target regions, thereby mitigating photometric ambiguity. Subsequently, geometric regularization and a masked perceptual loss are employed to enforce structural consistency and improve visual fidelity without overfitting to background noise. Finally, a two-stage decoupled optimization strategy separating geometry establishment from appearance refinement is adopted to ensure stable convergence. Experiments on a public wind turbine blade dataset demonstrate that the proposed framework significantly outperforms existing baseline methods, achieving up to 66.19% improvement in perceptual similarity and showing strong structural robustness under sparse viewpoints.
Keywords:
wind turbine blades
Gaussian Splatting
surface reconstruction
geometric priors
structure-guided optimization
Journal
IF:
9.9
Papers:
4.0K
Citations:
1.7W
Organization
No organization information available
Cited Papers