Multi-view fringe projection profilometry based on phase texture and U-Net

Abstract

The separability of patterns in a light-intersected area is the fundamental property of multi-view fringe projection profilometry (FPP). The traditional method based on temporal discrete Fourier transform separation and periodic wrapped phase requires dozens of patterns for each reconstruction. To enhance projection efficiency in multi-view FPP, a phase texture technique is proposed to reduce the pattern number by encoding the wrapped phase as an aperiodic texture. The U-Net neural network is trained on virtual datasets and employed as the decoder to map the phase texture to projector coordinates. To improve the decoder's adaptability for real measurements, the virtual dataset is configured with noise and defocus, while a monotonic loss function is designed. Simulations and experiments demonstrate that the proposed patterns are separable and the encoding method achieved reconstructions with only one-fifth the number of patterns required by traditional separation methods. The experimental results prove the improved decoding performance of U-Net trained with the monotonic loss function and the enhanced dataset.