Fig. 1. Proposed spectral convolutional neural networks architecture

Fig. 2. Comparison spectrogram before and after downscale. (a) Before downscale; (b) after downscale

Fig. 3. Noise level mapping M

Fig. 4. Simulated phase sample with three-peak beams. (a) Data sample 1; (b) data sample 2

Fig. 5. Digital holograms of two data samples from simulated phase. (a) Data sample 1; (b) data sample 2

Fig. 6. Simulated speckle noise holograms. (a) Data sample 1; (b) data sample 2

Fig. 7. Comparison of noise reduction based on different algorithms for analog noise hologram and its spectrum. (a) Original speckle noise hologram; (b) noise reduction hologram of BM3D algorithm; (c) noise reduction hologram of FFDNET network; (d) noise reduction hologram of SCNN network; (e) original speckle noise spectrogram; (f) noise reduction spectrum of BM3D algorithm; (g) noise reduction spectrum of FFDNET network; (h) noise reduction spectrum of SCNN network

Fig. 8. Reconstructed phase diagram and central cross section of their main peaks after noise reduction with different algorithms. (a) Reconstructed phase of original speckle noise; (b) noise reduction reconstruction phase of BM3D algorithm; (c) noise reduction reconstruction phase of FFDNET network; (d) noise reduction reconstruction phase of SCNN network; (e)--(h) are central sections along y axis of phase diagrams shown in Figs (a)--(d)

Fig. 9. Results of peak signal-to-noise ratio of simulated holograms with different noise reduction algorithms

Fig. 10. Noise and intensity reconstruction results of different algorithms for experimental hologram acquisition. (a) Reconstruction phase of original speckle noise; (b) noise reduction result of BM3D algorithm; (c) FFDNET network noise reduction; (d) SCNN network noise reduction; (e)--(h) are results of intensity reconstruction of holograms shown in Figs (a)--(d)

Fig. 11. Results of the peak signal-to-noise ratio of holograms collected by different noise reduction algorithms