Deep-learning-enabled single-shot high-frequency color fringe projection profilometry based on dual inner shifting-phase method

It is still a challenge to correctly retrieve the absolute phase from a high-frequency fringe pattern in a single-shot. To address this issue, a dual inner shifting-phase method is proposed by decomposing the fringe step order number into the product of two smaller step order numbers to overcome the fault-prone problem of step order solution by the existing single inner shifting-phase method. A U-Net model is utilized to train on mixed-frequency dual inner shifting-phase fringe datasets to extract six sinusoidal feature fringes rather than absolute phases to enhance the generalization ability of the DL-CFPP. The proposed method achieves a high phase unwrapping success rate of over 99% for high frequency fringe up to 100 periods, and accurate 3D shape reconstruction across various fringe frequencies using only one trained model even with fringe frequencies not included in the training dataset. Experiments in measuring dynamic objects verified the proposed technique's advantages of robust 3D shape reconstruction with high-frequency fringe.