Deep learning-assisted slightly off-axis digital holographic quantitative phase imaging

Digital holographic microscopy (DHM) is a typical quantitative phase imaging technique, in which the entire complex wavefront information is interferometrically encoded as a fringe pattern (so-called hologram) and then quantitatively demodulated by fringe analysis methods. Yet the off-axis digital holographic phase demodulation typically requires sufficiently high carrier spatial frequency for separating ±1-order and 0-order spectrum in the Fourier domain, limiting the space-bandwidth product (SBP) of the system. The in-line holographic configuration can realize full detector-bandwidth phase reconstruction at the cost of time resolution. In this work, we proposed a high-accuracy artifacts-free single-frame low-carrier frequency fringe demodulation scheme for the slightly off-axis digital holography, optimizing the system’s SBP effectively. This scheme acts as a method of deep-learning assisted physical model, incorporating a convolution neural network into a complete physical model by the idea of residual compensation, which enhances the imaging precision of the physical method while promises the interpretability of deep learning. The effectiveness of the proposed method is quantitatively analyzed through numerical simulation and experimentally verified by live-cells experiment. The phase recovered accuracy can be improved by one order of magnitude (the MAE up to 0.0065) compared with the traditional physical method. The live-cells experiment demonstrates the practicality of the method in biological research. Furthermore, it’s worth noting that the proposed method achieves a higher reconstruction accuracy utilizing only a small fraction of the datasets of the classical end-to-end deep learning model (without a physical model). The proposed deep learning-assisted physical model idea in this article is expected to bring more solutions for diverse computational imaging techniques.