Figure 6.

Representative reconstructed images versus iteration number at the different stages of iterative phase retrieval by using different initial starting methods. (a) Random initialization. (b) Using the NNS results as initialization. (c) Using the support from the machine-learning model and random phase as initialization. (d) Using the support and phase from the machine-learning model as initialization. The first three images in each row are the amplitudes of the particle and the rest are the corresponding phases. The corresponding iteration numbers of each of the images are given at the top of each column. When plotting the phase images of the particle, a 2D phase-unwrapping algorithm (see Appendix C for details) is applied to make the obtained phase continually change.