Research On Phase Demodulation Technology Based On Deep Learning

Obtaining the phase map of the tested sample from an interferogram is a fundamental problem in optical interferometry.Since the phase map reflects the topographic information of the surface for the measured object,the reconstructed phase map determines the measurement accuracy directly in interferometry.Interferograms containing closed fringes are usually obtained in measuring optical spherical/aspherical surface.Currently,achieving phase demodulation from a single interferogram which containing closed fringes still suffers from the disadvantages of low phase reconstruction accuracy,large computational burden,and high sensitivity to coherent noise.In order to solve the above-mentioned problems,in this paper two distinct demodulation methods based on deep learning are proposed,respectively.With either of the two methods,high-accuracy phase reconstruction from a single interferogram containing closed fringes can be achieved.Firstly,a one-step phase demodulation method is proposed based on the HRUnet regression convolutional neural network.The HRUnet network combines both features of the Unet network and the HRnet network.It improves the expression ability under multiple resolutions(high and low resolution)so that the trained HRUnet network can directly output the high-accuracy unwrapped phase map.A large number of interferograms simulated by Zernike polynomial and experimental interferograms obtained from a ZYGO interferometer were used to construct the dataset jointly.The network can directly output the unwrapped phase map by inputting a single interferogram.By testing both the simulated interferograms and the experimental interferograms,results show that the accuracy of the phase map reconstructed from HRUnet is higher than the reconstruction results of the traditional HCNN neural network and the Unet neural network,respectively.In addition,the HRUnet network also demonstrates better generalization ability and higher immunity to coherent noise.Secondly,a two-step phase demodulation method is proposed based on the PhaseUnet++regression network.PhaseUnet++ network is constructed based on the Unet network and it also adopted the structure of the Res Net network.The PhaseUnet++ retains more dimensional feature information through dense network connections.Meantime,by using jump connections in the convolutional layer the gradient disappearance and network degradation problem is overcame effectively,which makes the network converge more easily.With the trained PhaseUnet++,an outputted sinusoidal interferogram is firstly obtained by simultaneously input a normalized cosine interferogram and an auxiliary map.Afterwards,wrapped phase map is calculated by the arctangent function.Finally,the unwrapped phase distribution is obtained using the unwrapping procedure.Simulation and experimental test results show that the proposed method using two maps as input has higher phase reconstruction accuracy than the method using a single interferogram as input;in addition,the PhaseUnet++ network shows better adaptability when dealing with interferograms of different sizes.