A Phase Fusion and Restoration Method Based on Global Neural Calibration Model

Abstract

Fringe projector profilometry provides incomplete and noisy measurement data for machined metal surfaces that are extensively utilized in various industrial applications. Previous methods have attempted to address this issue through high dynamic range imaging or by introducing reflectance information of extra facilities with multiple lights. However, these methods decrease measurement efficiency and complicate the measurement facility. Therefore, this article combines these two strategies by utilizing the reflectance information from nonfringe images to guide the fusion and restoration of incomplete phase information at different exposure times, where an advanced anisotropic reflectance model is employed to obtain the reflectance information. A global neural calibration model based on a neural radiance field (NeRF) is proposed to bridge the phase and reflectance information. This model comprises global neural phase-to-coordinate and phase-to-light models. The global neural phase-to-coordinate model uses the NeRF to establish a relationship from the pixel position and the relative phase to the relative point coordinate. Besides, the global neural phase-to-light model employs the NeRF to describe the light distribution of the projector. The global neural calibration model transforms and propagates the reflectance information to optimize the phase information, thereby enhancing the performance of fringe projection profilometry (FPP) for machined metal surfaces. Real experiments on standard ball bars demonstrate that the global neural calibration model achieves calibration accuracy of $31.89~\mu $ m. In addition, experiments on six machined metal workpieces demonstrate that the proposed method achieves the measurement accuracy of $58.5~\mu $ m compared with coordinate measuring machine (CMM) results.