Research On Phase Unwrapping Algorithm Based On Deep Learning

As a critical step of InSAR technology,phase unwrapping is essential for surface deformation processing.The traditional phase unwrapping methods usually have the problem of consuming too much runtime and it is difficult to achieve a good balance between the accuracy and efficiency in phase unwrapping.Deep Learning(DL)has attracted wide attention because of its powerful feature representation and feature learning ability,showing obvious advantages in many tasks in different fields such as image classification,object detection,attitude estimation,image segmentation,face recognition and so on.This paper studies phase unwrapping based on deep learning.The main work is as follows:(1)The traditional phase unwrapping methods such as the quality-guided method,branch-cut method and the iterative least squares method are analyzed and studied through the experiments of phase unwrapping for interferograms.(2)A phase unwrapping method based on position and channel attention high-resolution networks is proposed.The network fuses multi-scale information to repeatedly concatenate between different scale transformations,and the features extracted from backbone network are input to the attention model in parallel for feature refinement;At the same time,a cross-level gated decoder is added behind the attention model to selectively enhance the spatial information;The relevant data sets are constructed to train the network to effectively predict the wrap-counts of wrapped pixels,and the minimum discontinuity method is used as a post-processing step to correct the unwrapped phase to improve the accuracy of phase unwrapping.(3)A phase unwrapping method based on global attention up-sampling(GAUPU)and a phase unwrapping method based on spatial and channel attention networks(SCAPU)are proposed respectively.Here,the GAUPU network combines global attention up-sampling mechanism and PU-M-Net to construct a network suitable for unwrapping wrapped phase images.The SCAPU network utilizes deeplabv3+ as the backbone,adopts a serial-parallel atrous spatial pyramid pooling module,implements multi-scale skip connections between the encoder-decoder models,and fuses a convolutional block attention module.Secondly,the noise level evaluation system for interferograms,which has been demonstrated well,is used to estimate the noise level of the interferograms,and the data sets with different noise levels are constructed to train the network model to ensure the trained networks suitable for unwrapping interferograms with different noise levels.Finally,noisy wrapped phase images are unwrapped by the trained networks with the same noise levels as the noisy interferograms.The experimental results of phase unwrapping for interferograms fully verify the performance of GAUPU and SCAPU’s method.