UARNet-3D: deep learning-based multiview 3D reconstruction for fringe projection profilometry

Deep learning-based direct 3D reconstruction for Fringe Projection Profilometry (FPP) has attracted significant attention. However, the focus remains on single-view reconstruction. Such single-view methods inherently suffer from occlusions in imaging, leading to compromised accuracy. To overcome such limitations, we introduce an end-to-end deep learning framework for high-fidelity 3D reconstruction utilizing fringe patterns acquired from four distinct perspectives. UARNet3D, a novel 3D-CNN-based multi-scale fusion network, integrates multi-view fringes to achieve enhanced depth precision and detail retention. It simultaneously processes fringe images from all four perspectives. Experiments using a DMD-based FPP system demonstrates its effectiveness and robustness. It achieves significantly lower reconstruction errors compared to state-of-the-art single-view networks, with a Mean Absolute Error (MAE) of 0.0541 mm and a Root Mean Square Error (RMSE) of 0.2007 mm. These results represent more than 50% reduction in both metrics. UAR-Net3D is a promising approach to address the increasing demands for FPP-based precision measurement.