Deep-learning-based computer-generated hologram from a stereo image pair.

Area of Science:

• Computer vision
• Computational optics
• Deep learning

Background:

• Traditional hologram generation requires complex depth estimation and diffraction calculations.
• Real-time holographic display is limited by computationally intensive processes.

Purpose of the Study:

• To propose an efficient deep-learning-based framework for generating computer-generated holograms (CGHs) from stereo images.
• To enable rapid and straightforward CGH synthesis from real-world scene data.

Main Methods:

• Developed an end-to-end convolutional neural network, the Stereo-to-Hologram Network (SHNet).
• SHNet takes a stereo image pair as input and outputs a monochromatic 3D complex hologram.
• The network bypasses intermediate depth recovery and diffraction computations.

Main Results:

• Successfully synthesized monochromatic 3D complex holograms from stereo image pairs.
• Demonstrated rapid and straightforward CGH calculation from directly recorded real-world scenes.
• Achieved clear 3D reconstructions with depth cues via numerical simulations and optical experiments.

Conclusions:

• The SHNet framework offers an efficient alternative for generating CGHs.
• This deep-learning approach significantly reduces the computational complexity of hologram production.
• Validated the effectiveness of SHNet-generated CGHs for holographic 3D reconstruction and virtual reality displays.