Abstract In this paper, a phase unwrapping method based on Residual Auto Encoder Network is proposed. Phase unwrapping is regarded as a multiple classification problem and it will be solved by the trained network model. Through training and validation stages, optimal network model can be served as a predictor to predict wrap count distribution map of wrapped phase. Then merge wrapped phase and count together to complete unwrapping. Software simulation and hardware acquisition are the sources for the generation of training dataset. To further improve the accuracy of unwrapping, an image analysis based optimization method is designed that can remove misclassification and noise points in initial result. In addition, phase data stitching by Iterative Closest Point (ICP) is adopted to realize dynamic resolution and enhance the flexibility of method. Point diffraction interferometer (PDI) and multi-step phase extraction technique are introduced, which is the foundation of proposed method. It can be concluded from the experiments that the proposed method is superior to state-of-art ones in terms of unwrapping performance, time efficiency, anti-noise ability and flexibility.

中文翻译：

---

基于残差自编码神经网络的点衍射干涉仪相位展开方法

摘要 本文提出了一种基于残差自编码器网络的相位解缠方法。相位解缠被认为是一个多分类问题，它将由经过训练的网络模型解决。通过训练和验证阶段，最优网络模型可以作为预测器来预测包裹阶段的包裹计数分布图。然后合并包裹阶段并一起计数以完成展开。软件模拟和硬件获取是生成训练数据集的来源。为了进一步提高展开的准确性，设计了一种基于图像分析的优化方法，可以去除初始结果中的错误分类和噪声点。此外，采用迭代最近点（ICP）拼接相位数据，实现动态分辨率，增强方法的灵活性。介绍了点衍射干涉仪（PDI）和多步相位提取技术，这是该方法的基础。从实验中可以得出结论，所提出的方法在解包性能、时间效率、抗噪能力和灵活性方面均优于最先进的方法。