Research On Phase Unwrapping Algorithm Based On Deep Learning

In recent years,the application of deep learning is more and more widely,which has become the basis of many modern artificial intelligence applications.Convolutional neural network has become one of the most potential deep neural network models for its powerful image feature extraction and learning ability.It has been widely used in image recognition and classification,SAR image classification,optical interferogram reconstruction and other fields.Phase unwrapping is the key technology in the application of InSAR,and it is also the most important step in many processes of InSAR Data processing.At present,the traditional phase unwrapping algorithm has the problem that it is difficult to balance the accuracy and efficiency of unwrapping.Aiming at the development trend of InSAR intelligence,this paper will carry out the research of phase unwrapping technology under the framework of deep learning theory.The main research contents are as follows:1.Research on traditional phase unwrapping algorithms,including classical path tracking,minimum norm,and nonlinear filter phase unwrapping algorithms,and analyze each algorithm in the phase unwrapping experiments of simulated and measured topographic interferograms Compare with.2.Aiming at the complex phase unwrapping problem,a phase unwrapping algorithm based on full convolution Dense Net is proposed.The dense blocks of convolution layer series structure are included in the model architecture,which makes the network achieve a better balance between interferogram feature extraction and detail keeping,and is conducive to improving the phase unwrapping accuracy and training efficiency of the network.The experimental results show that the algorithm has good unwrapping effect for high SNR interferogram.3.Combining the U-net architecture,the spatial pyramid pool(ASPP)network,and the bottleneck residual network,an improved U-net phase unwrapping algorithm is proposed.This method builds a robust phase unwrapping network based on the U-net architecture to establish the mapping relationship from the wrapped phase to the unwrapped phase.The ASPP combines the advantages of multi-scale information and dilated convolution,and gathers feature maps with different dilation rates to capture rich contextual information,and expands the feature receiving field without sacrificing feature spatial resolution,which is conducive to obtain the characteristic information of the wrapping interferogram accurately,and enhance the robustness of the phase unwrapping network.The bottleneck residual network can make the network model reduce the amount of parameter calculation while preventing network degradation,and improve the accuracy and efficiency of network training.The results obtained with simulated and measured data show that the proposed method can obtain more robust results,compared with other similar methods.In addition,in order to cooperate with the above algorithm research,the construction method of InSAR Data Set is studied,and the interferogram data set generated by simulated terrain,quasi measured terrain and digital elevation model data is created.