RSAGAN: Rapid self-attention generative adversarial nets for single-shot phase-shifting interferometry

Abstract

The phase-shift operation in phase-shifting interferometry (PSI) often introduces deviation and limits its application in the field of dynamic phase measurement. Based on this, a rapid self-attention generative adversarial nets (RSAGAN) is constructed and applied to realize single-shot PSI. By constructing the neural network of attention mechanism, we can quickly obtain the region dependency of the whole interferogram, so as to obtain a more realistic result after effective iterations. Accordingly, we can take one interferogram as the input of the network, get another interferogram with arbitrary phase shifts, and then combine the two-step phase-shifting algorithms to get the measured phase information. This method only needs one interferogram combined with the trained RSAGAN to achieve the acquisition of phase information with a simple experimental device, and has better generalization ability than the direct end-to-end phase acquisition method. Compared with the network without attention mechanism, the designed network has higher learning accuracy and has the potential to be applied in the related fields of dynamic phase measurement.