Deep learning-based single frame phase retrieval

摘要

In the field of precise 3D reconstruction, fringe pattern profilometry (FPP) is always regarded as the preferred method for it provides relatively higher accuracy. However, the phase acquisition process generally requires a sequence of images with different phase shift, which is rather time-consuming. Thus the application scenario of FPP is greatly limited and this has long been a bottleneck in practice. Although single-frame based phase retrieval algorithms like Fourier transform profilometry (FTP) has been proposed and extensively studied, they still suffer from relatively unbearable loss of accuracy. In response to this problem, we take advantage of the deep learning techniques and present a deep-learning based phase acquisition system in which the phase can be acquired by a single frame of fringe pattern image. The network is constructed according to the procedure of phase retrieval, which is trained by thousands of fringe pattern images with the phase data being known in advance. And it can predict more preciously the phase of a new fringe pattern map. Experiments illustrate the effect of our method which will be promising for practical use.