Fig. 3: Overview of the training procedure.

In the first stage, the CNN is trained to reproduce the ground truth midpoint hologram with direct supervision and a dynamic focal stack loss. In the second stage, the CNN prediction is propagated to the target hologram plane, filtered by a second CNN, double phase encoded, Fourier filtered, and back propagated to the center of the 3D volume to obtain the post-encoding midpoint hologram. No ground truth phase-only hologram is provided, and the CNN is trained to discover an optimal solution with the dynamic focal stack loss between the post-encoding focal stack and the target focal stack plus a regularization loss (see Eq. (11)). The Fourier space amplitude is visualized in the logarithmic scale