Dynamic quantitative phase imaging provides an effective solution for measuring the dynamic process of time-varying objects, such as biological samples, fluids, and flexible materials. However, there have been no effective approaches considering the spatial–temporal information of the dynamic process. Here, we report Ynet convolutional long short-term memory (Ynet-ConvLSTM) neural network; it learns the spatial features of the measured object that changes continuously along the time axis of a dynamic process by exploiting the known information of the interferogram sequence and phase images. According to our results, Ynet-ConvLSTM network improved the accuracy of phase image reconstruction in different dynamic circumstances.

动态定量相位成像为测量生物样品、流体、柔性材料等随时间变化的物体的动态过程提供了有效的解决方案。然而，目前还没有考虑到动态过程的时空信息的有效方法。在这里，我们报告了 Ynet 卷积长短期记忆 (Ynet-ConvLSTM) 神经网络；它通过利用干涉图序列和相位图像的已知信息来学习沿动态过程的时间轴连续变化的测量对象的空间特征。根据我们的结果，Ynet-ConvLSTM 网络提高了不同动态情况下相位图像重建的准确性。