Deep neural network for multi-depth hologram generation and its training strategy.

Area of Science:

• Computational optics
• Digital holography
• Machine learning applications

Background:

• Generating multi-depth holograms is crucial for advanced optical displays and imaging.
• Conventional computer-generated holograms (CGHs) can be computationally intensive and complex to design for multiple depths.

Purpose of the Study:

• To develop a deep neural network capable of generating multi-depth holograms.
• To propose an effective training strategy and dataset composition method for deep learning holograms (DLHs).

Main Methods:

• A deep neural network architecture was designed to accept multiple depth images as input and output complex holograms.
• A novel dataset compositing method was developed, using adjusted random dot densities and basic shapes for training.
• The generated DLHs were verified through numerical and optical reconstruction.

Main Results:

• The proposed network successfully generated deep learning holograms (DLHs) that reconstruct input images at their corresponding depths.
• Numerical and optical reconstructions demonstrated that DLHs can produce clear images at multiple depths, comparable to conventional CGHs.
• Quantitative evaluation using peak signal-to-noise ratio (PSNR) confirmed the quality of reconstructed images.

Conclusions:

• The presented deep neural network and training strategy offer an effective method for generating multi-depth holograms.
• The DLH approach shows promise for applications requiring high-quality, multi-depth holographic reconstruction.
• The developed dataset composition method significantly improves the performance of DLHs.