SIFT flow-based pattern registration for accurate deep learning-based single-shot fringe projection profilometry

Xin Liu, Haotian Yu, Chuang Liang, Lianfa Bai, Jing Han, Dongliang Zheng,
SIFT flow-based pattern registration for accurate deep learning-based single-shot fringe projection profilometry,
Optics and Lasers in Engineering,
Volume 195,
2025,
109307,
ISSN 0143-8166,
https://doi.org/10.1016/j.optlaseng.2025.109307.
(https://www.sciencedirect.com/science/article/pii/S0143816625004920)
Abstract: Single-shot fringe projection profilometry (FPP) shows great potential in dynamic 3-D sensing applications. It is necessary to extract the fundamental spectrum from the frequency domain and calculate the desired phase. But the spectrum overlapping problem constrains the phase accuracy. For single-shot FPP, deep learning techniques have recently gained success, which improve the phase accuracy obviously but the performance is still not the best. In this paper, we verify that the retrieved phase by the traditional deep learning-based single-shot FPP is limited due to the spectrum overlapping problem. For solving the problem, a SIFT flow-based pattern registration method is proposed. In a dynamic scene containing moving objects, SIFT flow establishes pixel correspondence among a sequence of patterns. The proposed method first extracts the fringe background, and then calculates flow vectors for establishing correspondence. From which, deep learning-based single-shot phase retrieval technique can obviously reduce the phase error caused by the spectrum overlapping problem. Experimental results verify that the proposed method can achieve accurate single-shot FPP especially for objects with complex textures. Compared with traditional deep learning-based single-shot FPP, the measurement accuracy is improved and the 3-D details can be recovered more completely without sacrificing measurement speed.
Keywords: 3-D measurement; Fringe projection profilometry; Single-shot; Pattern registration; SIFT flow