Temporal phase unwrapping with a lightweight deep neural network

Absolute phase unwrapping is important for optical phase-based three-dimensional (3D) measurements of isolated objects or discontinuous surfaces. For fringe projection profilometry, the fringe order of each stripe can be obtained uniquely by temporal phase unwrapping (TPU) that employs at least two different phase maps of the same measured scene. In this work, we present a novel TPU strategy using deep learning. Our idea is to treat the calculation of fringe order as a classification problem, which can be solved by a lightweight fully connected neural network. Consequently, the training process of our network can be finished within an hour, which saves a large amount of time and makes it possible to deploy the network on mobile devices. Moreover, rather than obtaining the training data from time-consuming real experiments, we simulate the data with software under different noises.