High-speed computer-generated holography using an autoencoder-based deep neural network.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Traditional computer-generated holography (CGH) requires extensive datasets for supervised learning.
• Existing methods often struggle with generating high-resolution, speckle-free holograms efficiently.

Purpose of the Study:

• To develop an unsupervised learning method for rapid, high-fidelity phase-only hologram generation.
• To improve hologram quality by reducing speckles compared to conventional algorithms.

Main Methods:

• An autoencoder-based neural network, termed holoencoder, was designed for phase-only hologram generation.
• Physical diffraction propagation was integrated into the autoencoder's decoding process.
• The model was trained in an unsupervised manner to learn latent hologram encodings.

Main Results:

• The holoencoder successfully generated high-fidelity 4K resolution holograms in just 0.15 seconds.
• Reconstruction results demonstrated the holoencoder's strong generalizability.
• Experimental comparisons showed a significant reduction in speckles compared to existing CGH algorithms.

Conclusions:

• The proposed unsupervised holoencoder offers a fast and effective approach for generating high-quality phase-only holograms.
• This method advances holographic display technology by improving hologram generation speed and image fidelity.
• The holoencoder shows promise for practical applications requiring efficient and high-quality holographic content.