Deep learning approach for full-field displacement estimation in depth-resolved wavelength-scanning interferometry

Abstract

To minimize the effect of optical crosstalk-generated noise (crosstalk), we present a deep learning approach to precisely estimate the full-field displacements for depth-resolved wavelength-scanning interferometry (DRWSI). A deep convolution neural network, where the transformer block is introduced to effectively capture higher-order features of the wrapped phase difference map in a strong noise environment, is applied for phase unwrapping. Furthermore, a binary phase noise map is used to update the loss function in an improved training model, enhancing the network generalization. Finally, the full-field displacements are estimated from phase unwrapping maps with a high signal-to-noise ratio. The simulation and the loading experiment verified that a higher accuracy reconstruction of displacement is implemented using our approach. The contribution of this work can make the DRWSI more practical in quantifying the mechanical property inside the sample.