Research On Phase Unwrapping Algorithm Of Optical Image Based On Deep Learning

With the rapid development of deep learning,the nonlinear fitting ability of deep convolutional neural networks is becoming stronger.In addition,many problems that cannot be solved in other fields are solved by deep learning.As an indispensable key technology in the application of optical phase measurement,phase unwrapping plays a key role in the application of holography and structured light illumination system.At present,the traditional phase unwrapping technology has entered bottleneck period in the speed and quality of phase unwrapping.Therefore,this paper proposes the method of deep learning for phase unwrapping.This paper mainly discusses the research of phase unwrapping based on residual regression network and semantic segmentation network.This paper first proposes the phase unwrapping using the residual regression network,and deeply explores the mapping relationship between the wrapped phase map and the unwrapped phase map.A large number of simulated phase map data is generated using the Zernike polynomial.The wrapped phase map is used as the input of the residual regression network while the unwrapped phase map is used as output.Due to the existence of residual structure,the network only needs to learn the mapping relationship between wrapped phase map and the difference of unwrapped phase map and wrapped phase map.The experimental results show that the algorithm can only unwrap about 87% of the wrapped phase maps in the test set.At the same time,there is a pixel error in the result of phase unwrapping.Secondly,the phase unwrapping algorithm based on semantic segmentation network proposed in this paper consists of three steps: segmentation,summation and refinement.In the step of segmentation,we use the Deep Lab V3+ network to semantically segment the wrapped phase map according to the multiple of 2?.Then,the multiplication value of 2? in the segmentation result is multiplied by 2? and then add to the wrapped phase map.Hence a temporary phase unwrapping result is obtained.Lastly,a refinement step is deployed on the intermediate result,yielding the final result.Compared with the latest phase unwrapping algorithms based on quality-guided of path dependent algorithm(SRFNP)and the path independent algorithm based on the transport of intensity equation(TIE,ITIE and RTIE),the results show that our algorithm is more robust and has higher quality when there is noise.