Asymmetrical neural network for real-time and high-quality computer-generated holography

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Neural network-based computer-generated holography (CGH) offers real-time hologram generation.
• Current lightweight networks limit fitting capacity, impacting display quality.
• A need exists for methods balancing real-time performance with high fidelity.

Purpose of the Study:

• To propose a novel asymmetrical neural network with a non-end-to-end structure for enhanced CGH.
• To improve the fitting capacity and real-time display quality of neural network-based CGH.
• To maintain a lightweight architecture while boosting performance.

Main Methods:

• Developed an asymmetrical neural network with a non-end-to-end architecture.
• Decomposed the CGH task into two sub-tasks: phase prediction and hologram encoding.
• Employed distinct network layers tailored to each sub-task for optimized feature extraction and encoding.

Main Results:

• The proposed asymmetrical network demonstrated enhanced fitting capacity.
• Superior real-time display quality was achieved compared to existing methods.
• Numerical reconstructions and optical experiments confirmed the method's effectiveness.

Conclusions:

• The asymmetrical, non-end-to-end neural network effectively addresses limitations in current CGH methods.
• This approach successfully balances lightweight architecture with high-fidelity hologram generation.
• The validated method shows significant promise for advanced real-time holographic applications.