Weakly supervised phase unwrapping for fringe projection profilometry

Abstract

Phase unwrapping is a vital task applied in fringe projection profilometry (FPP). In FPP, fast-speed and high-accuracy three-dimensional (3D) imaging has been the goal. One prominent approach to achieving this objective is the dualfrequency temporal phase unwrapping method (DF-TPU). However, the highest period number for the DF-TPU approach is usually constrained to no more than 16 or 32 by inevitable phase errors, thereby limiting reconstruction precision. For single-camera FPP systems, existing deep learning-based methods capable of unwrapping high-frequency phase maps require accurate labels. This paper proposes a novel deep-learning-based phase unwrapping method for single-camera FPP systems. The inaccurate unit-period phase map is used as the weakly supervised label to guide the convergence of the unwrapping of the high-frequency phase map. The trained network can unwrap the phase map of 64 periods. The proposed approach has been validated in multiple real-world scenarios, including motion blur, isolated objects, non-uniform reflectivity, and phase discontinuity.