基于MAPNet的深度学习相位解缠方法  

Phase Unwrapping Method Based on Multiple Attending Path Network

作　　者：侯国政 谢先明 贾金国 田冲宵 李蓉 HOU Guozheng;XIE Xianming;JIA Jinguo;TIAN Chongxiao;LI Rong(School of Automation,Guangxi University of Science and Technology,Liuzhou,Guangxi 545006,China;School of Electronic Engineering,Guangxi University of Science and Technology,Liuzhou,Guangxi 545006,China)

机构地区：[1]广西科技大学自动化学院,广西柳州545006 [2]广西科技大学电子工程学院,广西柳州545006

出　　处：《遥感信息》2024年第5期86-101,共16页Remote Sensing Information

基　　金：国家自然科学基金(62161003);广西自然科学基金(2023CXNSFAA026024)。

摘　　要：文章提出一种基于深度学习神经网络的相位解缠方法。首先,以多路径神经网络为骨架,采用并行多路径架构对干涉图进行不同尺度信息处理,以获取不同分辨率特征图,并利用通道注意力机制模块与空间池化模块融合不同分辨率的特征图,达到兼顾全局语义信息与局部相位细节信息的目的,有利于提升网络相位解缠性能。其次,利用噪声量级评估系统制作不同噪声量级数据集对网络模型进行训练,使训练完成的网络具有解缠不同噪声量级干涉图的能力。最后,利用构建的网络模型对不同噪声量级干涉图进行相位解缠,获得其解缠相位。实验结果表明,该网络模型可对不同条纹特征干涉图进行相位解缠且效果良好。A deep-learning-neural-network-based phase unwrapping approach is presented to recover unambiguous unwrapped phases from wrapped images.Firstly,this method considers a multipath neural network as the skeleton,and uses a parallel multipath architecture to extract the feature maps with different resolutions from interferograms.The channel attention mechanism module and the spatial pooling module are used to effectively integrate the feature maps with different resolutions to balance the global semantic information and local detailed phase information,which is conducive to improving the performance of the network in phase unwrapping for interferograms.Secondly,different noise level datasets are created using a noise level evaluation system to train the network model,enabling the trained network to have the ability to unwrap interferograms with different levels of noise.Finally,the constructed network can directly unwrap the interferograms with different noise levels.Experimental results show that this network can effectively unwrap the interferograms with different fringe features and achieve robust performance.

关 键 词：相位解缠 多路径神经网络 深度学习 干涉图 

分 类 号：P237[天文地球—摄影测量与遥感]