Fig. 1. Phase recovery network training with DD and PD strategies.

Fig. 2. Description of DD deep learning phase recovery methods.

Fig. 3. Description of PD deep learning phase recovery methods. (a) Network inference for the uPD. (b) Network training and inference for the tPD. (c) Network training and inference for the tPDr.

Fig. 4. Inference results of DD, uPD, tPD, and tPDr.

Fig. 5. Results of DD, tPD, and CD. The blue box represents low-frequency information and the green box represents high-frequency information.

Fig. 6. Cross-inference results of DD and tPD for the datasets of ImageNet, LFW, and MNIST. The metric below each result is the average SSIM for that testing dataset.

Fig. 7. Ill-posedness adaptability test of DD and tPD. Blue part represents a single hologram as the network input, red part represents a single hologram with aperture constraints as the network input, and yellow part represents multiple holograms as the network input.

Fig. 8. Dataset generation and network training for the case of imaging aberration.

Fig. 9. Prior capacity test of DD and tPD.

Fig. 10. Experimental tests of DD, tPD, CD, and uPD(tPDr). (a) Inference results of one field of view. (b) Inference results of another field of view.