Fig. 1. (a) Schematic of Holomer’s (our method) two-stage architecture and its training pipeline. (b) Network architecture schematic for the target phase generator and the phase-only hologram encoder. The target phase generator receives a single-channel target amplitude as input. The input of the phase-only hologram encoder is dual-channel amplitude and phase images. The U-shaped network architecture ensures the consistency between output and input dimensions, enhancing the network’s capability to learn multi-level features. (c) Holomer block schematic; each Holomer block contains two modules, window multihead self-attention (W-MSA) and sliding window multihead self-attention (SW-MSA). The two diagrams correspond to the receptive field of the two modules. By employing the sliding window self-attention mechanism, the Holomer block significantly enhances its receptive field, thereby improving its ability to learn long-range features of the diffraction process.

Fig. 2. (a), (b) Illustration of the amplitude and phase distribution of diffraction impulse response (DIR), respectively. In the computation of the DIR, the wavelength employed is 520 nm, the pixel size is 6.4 μm, and the propagation distance is 20 cm. (c) The schematic illustration for image reconstruction through the convolution of DIR with a hologram. To fulfill the conditions of linear convolution, it is necessary to zero-pad the hologram to twice its original size before proceeding with the convolution. (d) A comparative schematic illustration contrasting the receptive fields of Holomer (ours) and a single-layer CNN, along with their respective coverage of the DIR region size. (e) The curve illustrating the variation in the proportion of the sum of intensities over the region covered by the receptive field to the total DIR intensity with respect to changes in the receptive field size, annotated with markers for a single-layer CNN and Holomer. The DIR coverage of Holomer exhibits a notable improvement compared to CNN.

Fig. 3. A comparative demonstration of numerical simulation results encompasses Wirtinger, HoloNet, CCNN, and our proposed Holomer, utilizing images sourced from the DIV2K validation set. In specific experiments, Wirtinger underwent 200 iterations, while HoloNet, CCNN, and Holomer underwent training for an equivalent duration on the DIV2K training set.

Fig. 4. Captured reconstructed images of the green channel (520 nm) for Wirtinger, HoloNet, CCNN, and Holomer methods, along with the corresponding ground truth image.

Fig. 5. (a) Schematic diagram of the experimental setup used to verify the optical reconstruction of Holomer-generated holograms using an aperture to eliminate diffracted images from other levels. (b), (c) Optical reconstruction images of our method in color channels directly captured by a camera. (d), (e) The corresponding color holograms loaded on the SLM.