In quantitative phase imaging (QPI), it is greatly important to extract the phase from the phase-shifting interferograms. Despite extensive research efforts for decades, how to retrieve the actual phase using the minimum number of interferograms, continues to be an important problem. To cope with this problem, a deep-learning-based method of phase extraction is proposed in QPI. After the fringe pattern features of interferograms associated with phase retrieval are extracted, the proposed approach can establish the pixel-level mapping relation between the interferograms and ground-truth phases so that it can rapidly recover the true phase, without phase unwrapping, from one-frame interferogram. The feasibility and applicability of this method are demonstrated, respectively, by the datasets of the microsphere, neuronal cell with refractive index variation and red blood cell. The results show that this method has obvious advantages in terms of phase extraction, compared with the traditional phase retrieval algorithms.

在定量相位成像（QPI）中，从相移干涉图中提取相位非常重要。尽管数十年来进行了广泛的研究，但是如何使用最少数量的干涉图来检索实际相位仍然是一个重要问题。为了解决这个问题，在QPI中提出了一种基于深度学习的相位提取方法。提取与相位检索相关的干涉图的条纹图案特征后，该方法可以在干涉图和地真相之间建立像素级映射关系，从而可以从一个相位快速恢复真实相位，而无需展开相位。帧干涉图。通过微球的数据集分别证明了该方法的可行性和适用性，折射率变化的神经元细胞和红细胞。结果表明，与传统的相位检索算法相比，该方法在相位提取方面具有明显的优势。