3D Holography Based on D2NN with Random Phase Image Target

In this paper, we propose a novel approach for 3D holography realization based on the diffractive deep neural network (D2NN), which is realized by multiple cascaded SLMs. Before the training process of the D2NN, random phases are added to target images located at different propagation distances to suppress the crosstalk between them. The L2 norm of the forward propagation fields and the target fields is used as the loss function and the Adam method is used to update the phases of the SLMs. After convergence, the simulation results of the 1-layer, 4-layer, and 6-layer D2NNs are compared to the Fresnel holography with random phases. The image clarity of the Fresnel holography and the 1-layer D2NN are similar with the mean squared error (MSE) as 1.211 and 1.212 and the peak signal-to-noise ratio (PSNR) as 59.375dB and 59.369dB respectively. While the results of the 4-layer and 6-layer D2NNs are better than the Fresnel holography: the MSEs are 1.077 and 1.056 and the PSNRs are 59.884dB and 59.968dB.