Deep learning denoising diffusion probabilistic model applied to holographic data synthesis.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Vision

Background:

• Holographic data synthesis is crucial for applications like holographic displays and data storage.
• Current methods for generating holographic data can be complex and computationally intensive.
• Deep learning offers potential for efficient and accurate data generation.

Purpose of the Study:

• To introduce a novel method for holographic data synthesis using a deep learning probabilistic diffusion model (DDPM).
• To demonstrate the capability of the DDPM to generate complex-valued holographic data from various image datasets.
• To showcase the synthesis of diverse holograms, including 2D characters, vehicles, and 3D scenes.

Main Methods:

• Conversion of color image datasets into complex-valued holographic data via backpropagation.
• Training a deep learning probabilistic diffusion model (DDPM) on the generated holographic datasets.
• Utilizing a U-Net convolutional neural network within the DDPM to learn the noise-reversal process.

Main Results:

• Successful demonstration of holographic data synthesis using a DDPM for the first time.
• Generation of holograms with features similar to the training datasets.
• Synthesis of holograms containing color images of 2D characters, vehicles, and 3D scenes at varying propagation distances.

Conclusions:

• The developed DDPM provides an effective approach for synthesizing complex-valued holographic data.
• This method enables the generation of a wide variety of holograms by inputting Gaussian random noise.
• The study opens new avenues for advanced holographic applications powered by deep learning.