Batch fringe extraction from the single FPP fringe pattern based on a triple serial and parallel convolution neural network

In the paper, we propose a batch fringe extraction method for the single FPP fringe pattern based on a triple serial and parallel convolution neural network. The proposed network is a combination of three deep convolution neural networks in a serial and parallel way. We train the network by pairs of the original FPP fringe patterns and the corresponding fringe components. When the network is trained successfully, the fringe component can be obtained directly by feeding the original FPP fringe pattern into the trained network. Based on the extracted fringe component, we get the desired phase. We test the proposed method on many FPP fringe patterns and compare them with four reference methods including the Fourier transform, Shearlet transform, bi-dimensional empirical mode decomposition, and variational image decomposition methods. We also evaluate the performances with three quantitative metrics and four visual presentations. The experiment results show that the proposed method can extract the fringe component more accurately than the four reference methods. Besides, the proposed method can adaptively process different FPP fringe patterns in batch without any parameter fine-tuning. Additionally, the proposed method has been applied in a real dynamic measurement of a leaf in continuous dehydration successfully.