Fig. 2: Deep-learning-based hologram reconstruction. | Light: Science & Applications

Fig. 2: Deep-learning-based hologram reconstruction.

From: Deep learning in holography and coherent imaging

Fig. 2

a An end-to-end CNN was trained to transform a hologram directly to a phase image4. Adapted with permission from ref. 4. b The raw hologram (i.e., without phase information) was numerically focused onto the sample plane and was used as an input for the network to match the phase-recovered image3. The sample is a Pap smear specimen. Adapted with permission from ref. 3. c The raw hologram was propagated to an approximate distance within a sample volume, and the deep network generated an extended depth-of-field reconstruction of the whole sample volume, also performing autofocusing5. The specimen is a 3D distributed aerosol sample. Adapted with permission from ref. 5. d Similar to (b) but implemented on holograms under low-photon and poor-SNR conditions6. Adapted with permission from ref. 6. e A CNN was trained to transform a low-resolution holographic reconstruction (created using iterative multiheight phase recovery) to an equivalent high-resolution image of the same sample FOV7. The sample is a Pap smear specimen. Adapted with permission from ref. 7

Back to article page