Diffraction model-informed neural network for unsupervised layer-based computer-generated holography.

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Learning-based computer-generated holography (CGH) shows potential for real-time holographic displays but typically requires large labeled datasets.
• Existing supervised methods face challenges in dataset creation and complexity scaling with object depth.

Purpose of the Study:

• To propose a diffraction model-informed neural network framework (self-holo) for unsupervised 3D phase-only hologram generation.
• To enable real-time holographic displays without the need for paired training data.

Main Methods:

• Developed a neural network framework (self-holo) incorporating angular spectrum propagation for unsupervised training.
• Inputting amplitude and depth map images, the network synthesizes 3D or 2D holograms.
• Complexity is managed by reconstructing holograms to a single depth layer, independent of total depth layers.

Main Results:

• Successfully demonstrated 3D reconstructions with significant 3D effect.
• Validated the generalizability of the self-holo framework through numerical and optical experiments.
• Achieved unsupervised training, eliminating the requirement for labeled datasets.

Conclusions:

• The self-holo framework offers an efficient and data-independent approach to 3D hologram generation.
• This method advances the development of real-time holographic displays.
• The unsupervised learning strategy simplifies the CGH pipeline and broadens its applicability.