Absolute phase recovery from reduced phase-shifted fringe patterns via composite-loss-driven deep learning

Authors: Guorui Zhao, Qiang Liu, Tianhao Wu, guofeng zhang, Hongwei Zhao, Rongliang Zhu, Fengshuo Yang

Journal: Optics Express

Publication year: 2026

DOI: 10.1364/oe.614672

Abstract

Conventional multi-frequency heterodyne phase-shifting profilometry enables high-accuracy three-dimensional reconstruction but typically requires the acquisition of 12 fringe patterns, which limits measurement efficiency. This paper proposes a few-step absolute phase recovery method based on a multi-frequency fringe completion network (MFC-Net) and a composite loss function. At each of the three frequencies, two fringe patterns with phase shifts of 0 and π are captured, resulting in a total of six measured images. The network predicts the missing π/2- and 3π/2-shifted fringe patterns, which are combined with the measured patterns to form a complete three-frequency four-step sequence. Absolute phase is subsequently recovered using standard four-step phase demodulation and multi-frequency heterodyne unwrapping. The composite loss function comprises an L1 loss, a hard example mining loss, and a spatial gradient consistency loss, which constrain global intensity fidelity, high-error pixels, and local structures, respectively. Experiments were conducted on multiple test objects using two frequency combinations, 100/90/81 and 64/56/49. For objects with smoothly varying surfaces, the absolute phase RMSE of the proposed method was only 3.2%–8.5% higher than that of the three-frequency four-step method. For an object containing complex edges and occlusions, the proposed method reduced the RMSE by 42.8% compared with the three-frequency three-step method under the 64/56/49 frequency combination. Compared with other deep-learning-based recovery strategies, the proposed method generally achieved better point-cloud completeness and local-profile preservation. These results demonstrate that the proposed method reduces the number of physically captured fringe patterns from 12 to 6 while largely preserving absolute phase accuracy and three-dimensional reconstruction quality.

View this paper on Academic Now