A deep learning-based approach for motion-induced error reduction in fringe projection profilometry

Abstract

Fringe projection profilometry (FPP) is one of popular 3D measurement techniques, which can be divided into two categories: Fourier transform profilometry (FTP) and phase-shifting profilometry (PSP). Compared to FTP, PSP has been increasingly appealing to researchers due to its merits of higher accuracy and sensitivity. Although PSP works well with the assumption that the object stays quasi-static, it is sensitive to the object motion and causes the motion-induced error in dynamic measurement due to the multi-frame measurement mechanism. However, the multi-frequency phase unwrapping is always utilized to solve the problem of surface discontinuities, which limits the reduction of numbers of projected patterns. Besides, using the high-speed camera will increase the cost of the hardware. Recently, researchers have demonstrated that the phase unwrapping, fringe denoising and dynamic range can be improved with the assistance of deep learning technique. Therefore, in this paper, a deep learning-based method is proposed for motion-induced error reduction. With the aid of strong fitting capability of the neural network, the motion-induced errors can be significantly reduced even under low capture frame rate. The proposed method is experimentally verified on its applicability for dynamic 3D measurement.