基于深度学习的相位解缠方法

Phase Unwrapping Method Based on Deep Learning

针对干涉测量技术应用中的缠绕相位解缠问题,提出了一种基于U-Net3+的深度学习相位解缠方法。该方法以U-Net3+为架构,首先,利用全尺度的跳跃连接(skip connection)整合来自不同尺度特征图中的高级语义与低级语义信息;其次,利用深度监督加快网络收敛速度,同时增强干涉图特征信息;为防止梯度爆炸,在网络中添加残差网络,最终搭建适用于不同条纹类型干涉图解缠的网络架构,建立从缠绕相位到真实相位的直接映射关系。完成训练后的网络可直接对缠绕干涉图进行相位解缠以获得其真实相位估计,并在不同条纹类型的干涉图相位解缠实验中均表现出良好的稳定性。

Aiming at the problem of phase unwrapping for interferograms in interferometry,a phase unwrapping method based on U-Net3+ is proposed to retrieve the unwrapped phase from the wrapped phase.This method uses U-Net3+ as the skeleton,firstly integrates high-level and low-level semantic information from feature graphs at different scales by using full-scale skip connections.Secondly,deep supervision is used to accelerate the convergence speed of the network and enhance the characteristic information of the interferograms.Finally,the residual network is added to the network to prevent gradient explosion,and a network architecture suitable for unwrapping different types of fringe patterns is constructed to realize the direct mapping from the wrapped phase to the true unwrapped phase.The trained network can obtain the unwrapped phase from the wrapped phase,and shows good robustness in the experiments of phase unwrapping for different types of fringe patterns.