Fringe projection profilometry based on deep learning phase demodulation combined with temporal phase unwrapping

Abstract

In fringe projection profilometry (FPP), phase shifting profilometry (PSP) combined with temporal phase unwrapping (TPU) algorithms can be used to reliably obtain 3D information from complex measured scenes. However, collecting too many fringe patterns for phase demodulation reduces measurement efficiency. Some studies have shown that deep learning techniques can achieve phase demodulation on single-frame fringe pattern, suggesting that combining deep learning-based phase demodulation with TPU could potentially enable high-speed, high-precision 3D measurements. In this paper, we propose the FPP based on deep learning phase demodulation combined with TPU to achieve 3D measurements using only three fringe patterns. Furthermore, based on different network input strategies and TPU algorithms, the proposed method has four different implementation processes. Comparative experiments analyze the impact of different network input strategies, TPU algorithms, and network structures on the accuracy of phase demodulation and unwrapping. The results demonstrate that using multiple fringe patterns with different frequencies as a joint input significantly improves the phase demodulation accuracy for various frequencies, particularly for lower frequencies, compared to using a single pattern with a neural network. In contrast, enhancing the network structure alone yields relatively modest improvements in phase demodulation accuracy compared to adjusting the input strategy. By analyzing phase demodulation and unwrapping errors, this paper provides guidance on selecting the appropriate implementation process for the proposed method under varying levels of noise interference.