In this paper, we propose a deep nonlinear CNN model, named as PU-M-Net for phase unwrapping with speckle reduction and structure protection in ESPI. Our PU-M-Net consists of four pathways in “M” shape, and merges them by abundant skip connections. By this means, our PU-M-Net can improve the flow of deep features, promote the fusion of shallow information and deep features, and enhance the utilization of initial features in phase unwrapping. We propose a structure-consideration loss function by combining the MAE and MS-SSIM loss functions, and also construct an available pertinent dataset for ESPI phase unwrapping by means of the phase shifting method. With the proposed dataset and the proposed loss function, we train the PU-M-Net successfully under a limited training condition. With the trained network, we directly get the results from the original wrapped phases. We test our method on many simulated and experimental ESPI wrapped phases, and compare it with a least-squares method, a quality-guided method, and two deep learning methods. We also compared it with the PU-M-Net trained by MSE loss function. The performance is evaluated quantitatively and qualitatively in terms of phase unwrapping and structure protection. Results demonstrate that our method can reduce speckles and protect structures in phase unwrapping, and get the better results than the compared methods. Besides, it does not require any parameter fine-turning and any pre-process or post-process procedure, even for the case of high-level speckles or high-dense wrapped phases. Moreover, it has full advantages on excellent generalization and batch performance, and can be used to process a great number of images rapidly. Indeed, it has been applied to the dynamic measurement successfully in this paper.

在本文中，我们提出了一种深度非线性 CNN 模型，称为 PU-M-Net，用于在 ESPI 中使用散斑减少和结构保护进行相位解缠。我们的 PU-M-Net 由四个“M”形路径组成，并通过丰富的跳过连接将它们合并。通过这种方式，我们的 PU-M-Net 可以改善深层特征的流动，促进浅层信息和深层特征的融合，并提高相位展开中初始特征的利用率。我们通过结合 MAE 和 MS-SSIM 损失函数提出了一种结构考虑损失函数，并通过相移方法构建了一个可用的 ESPI 相位展开的相关数据集。使用建议的数据集和建议的损失函数，我们在有限的训练条件下成功训练了 PU-M-Net。通过训练好的网络，我们直接从原始包裹阶段得到结果。我们在许多模拟和实验 ESPI 包裹相位上测试我们的方法，并将其与最小二乘法、质量导向方法和两种深度学习方法进行比较。我们还将其与通过 MSE 损失函数训练的 PU-M-Net 进行了比较。在相位展开和结构保护方面对性能进行定量和定性评估。结果表明，我们的方法可以减少散斑并保护相位展开中的结构，并且比比较方法获得更好的结果。此外，它不需要任何参数微调和任何预处理或后处理程序，即使对于高水平散斑或高密度包裹相的情况也是如此。此外，它在出色的泛化和批处理性能方面具有充分的优势，并可用于快速处理大量图像。事实上，它已经成功地应用于本文的动态测量。