Weakly supervised phase unwrapping for single-camera fringe projection profilometry

Abstract

Phase unwrapping is essential in fringe projection profilometry (FPP). FPP's primary focus is to simultaneously realize high-speed, high-precision three-dimensional (3D) reconstruction. The dual-frequency hierarchical temporal phase unwrapping method (DF-TPU) technique is one prominent technique for accomplishing this goal. However, phase errors are inevitable and typically limit the DF-TPU approach's maximal period number, reducing reconstruction accuracy. The fully supervised approaches based on deep learning can unwrap dual-frequency phase maps of single-camera FPP systems but require accurate labels of high-frequency absolute phase. This study proposes a novel deep learning phase unwrapping scheme suited to single-camera FPP setups. Without the need for accurate labels corresponding to the high-frequency phase maps, supervisory signals are extracted from the high-frequency phase map, the one-period phase map, and the geometry constraint to guide the training of the phase unwrapping network. This training is a weakly supervised learning, which makes the time-consuming and expertise-demanding precise labeling unnecessary. Data collected in multiple real-world circumstances, such as motion blur, isolated objects, non-uniform reflectivity, and phase discontinuities, are used to validate the effectiveness of this weakly supervised method. Experimental results demonstrate that the proposed method outperforms the fully supervised and conventional DF-TPU methods in terms of depth accuracy and measurement efficiency in terms of fringe pattern number.