Deep Learning-based Single-Shot Composite Fringe Projection Profilometry with Pixel-Wise Uncertainty Quantification

TL;DR

This paper introduces HSURE-CFPP, a deep learning method for single-shot 3D shape measurement that provides both high accuracy and pixel-wise uncertainty estimates, enabling reliable and fast optical metrology.

Contribution

It proposes a heteroscedastic snapshot-ensemble network for composite fringe profilometry, offering uncertainty quantification and improved reliability in single-shot 3D imaging.

Findings

Achieves high-accuracy 3D reconstruction from a single composite fringe.

Produces pixel-wise uncertainty maps that correlate with reconstruction errors.

Demonstrates robustness in static and dynamic scenes.

Abstract

Driven by the growing demand for high-speed 3D measurement in advanced manufacturing, optical metrology algorithms must deliver high accuracy and robustness under dynamic conditions. Fringe projection profilometry (FPP) offers high precision, yet the 2pi ambiguity of the wrapped phase means that conventional absolute phase recovery typically relies on multiple coded patterns, sacrificing temporal resolution. Deep learning-based composite FPP (CFPP) shows promise for single-shot phase recovery from a composite fringe, but limited interpretability makes it difficult to assess reconstruction reliability or trace error sources in the absence of ground truth. To address this, we propose HSURE-CFPP (Heteroscedastic Snapshot-ensemble Uncertainty-aware Ratio Estimation for CFPP). HSURE-CFPP predicts the numerator-denominator ratio used for wrapped-phase computation with a heteroscedastic snapshot-ensemble network, enabling ultra-fast 3D imaging from a single composite fringe and producing pixel-wise uncertainty maps for confidence assessment and unreliable-region identification. Specifically, a heteroscedastic likelihood jointly estimates pixel-wise noise variance to capture data uncertainty, while a snapshot ensemble quantifies model uncertainty via dispersion across snapshots, yielding total predictive uncertainty as an interpretable reliability measure. Experiments on static and dynamic scenes demonstrate that HSURE-CFPP achieves high-accuracy reconstruction at high speed and that the predicted uncertainty correlates well with reconstruction errors, providing a deployable quality-assessment mechanism for deep-learning-based FPP.