Quantized neural network for complex hologram generation.

Area of Science:

• Computer vision
• Holography
• Machine learning

Background:

• Computer-generated holography (CGH) is crucial for augmented reality (AR) displays but faces computational challenges.
• Neural networks accelerate CGH, yet efficient models are needed for embedded systems.

Purpose of the Study:

• To develop a lightweight neural network model for complex hologram generation.
• To reduce computational cost, memory footprint, and power consumption for embedded CGH.

Main Methods:

• Introduced neural network quantization, specifically converting a tensor holography model from 32-bit floating-point (FP32) to 8-bit integer (INT8) precision.
• Evaluated hologram quality, model size, and processing speed.
• Implemented the INT8 model on a system-on-module for embedded deployment.

Main Results:

• The INT8 model achieved hologram quality comparable to the FP32 model.
• Model size was reduced by approximately 70%.
• Processing speed increased fourfold, demonstrating high power efficiency and deployability on embedded platforms.

Conclusions:

• Neural network quantization offers an effective solution for efficient CGH.
• The developed INT8 model is suitable for real-time AR applications on resource-constrained devices.
• This approach overcomes the limitations of traditional CGH for practical AR display implementation.