Phase dual-resolution networks for a computer-generated hologram.

Area of Science:

• Optics and Photonics
• Computer Vision
• Deep Learning

Background:

• Computer-generated holograms (CGHs) are crucial for optical field interference patterns.
• Existing iterative CGH algorithms face a speed-accuracy trade-off, while non-iterative methods lack hologram accuracy.
• Deep learning offers potential for faster and more accurate hologram generation.

Purpose of the Study:

• To propose a novel deep learning approach, the phase dual-resolution network (PDRNet), for generating phase-only holograms.
• To achieve high-fidelity hologram generation with fixed computational complexity, overcoming limitations of existing methods.
• To enable unsupervised training without ground-truth holograms.

Main Methods:

• Developed a phase dual-resolution network (PDRNet) utilizing a dual-resolution network architecture for hologram generation.
• Employed the differentiability of the angular spectrum method for unsupervised training of the convolutional neural network.
• Utilized a combined loss function of multi-scale structural similarity (MS-SSIM) and mean square error (MSE) for high-fidelity hologram generation.

Main Results:

• PDRNet successfully generated high-fidelity 1080P resolution holograms in 57 milliseconds.
• The proposed method demonstrated enhanced mapping capability through the optimized dual-resolution network.
• Experimental holographic displays showed a reduction in speckles in the reconstructed images.

Conclusions:

• PDRNet offers a significant advancement in computer-generated hologram technology, balancing speed and accuracy.
• The unsupervised deep learning approach provides a viable alternative for training hologram generation models.
• The method holds promise for improving the quality and performance of holographic displays.