Robust 3D reconstruction of highly reflective objects via structured light

Abstract Achieving accurate three-dimensional (3D) measurement on metallic and similar components in industrial environments remains a major challenge in fringe projection profilometry, primarily due to severe phase information loss caused by overexposure in highly reflective regions and underexposure in low-reflectance areas commonly present on such surfaces. These issues often lead to incomplete or distorted depth reconstruction. To address this, we propose a novel fringe restoration framework, called multi-attention fusion network, designed to enhance the robustness and precision of structured light-based 3D reconstruction under complex reflectance conditions. A key component of our approach is the reflective-aware multi-attention module, which emphasizes critical phase features while suppressing noise from saturated pixels. This enables effective compensation for fringe distortion in reflective regions. To overcome data scarcity, we further introduce a mask-guided adaptive fringe fusion strategy, which augments training data by replacing degraded regions with high-quality patches guided by semantic segmentation masks. Experiments on challenging objects, such as aero-engine blades, demonstrate that our method reduces the mean absolute error of depth maps from 0.191 to 0.045 and improves valid point cloud coverage from 70.8 % to 98.4 %. These results demonstrate the effectiveness of the proposed method in improving the robustness, completeness, and relative accuracy of structured-light-based 3D reconstruction under high-reflectivity conditions.