Memory-Reduced Convolutional Neural Network for Fast Phase Hologram Generation

Computer-generated holography (CGH) has advanced the development of human-centric holographic near-eye displays. Recent work has proposed an end-to-end convolutional neural network that converts 2D images into 3D holograms, enabling real-time 3D holographic displays from widespread available images. However, the high computational cost of deep learning-based methods limits their practical application. Deploying such algorithms on resource-constrained mobile platforms requires more efficient models with reduced computational memory and power demands, which plays a crucial role in promoting human-centric virtual reality/augmented reality displays. In this article, a lightweight 3D hologram generation model is proposed using neural network quantization from the input of single 2D image. Specifically, a 2D-to-3D CGH model is quantized from 32-bit floating-point to 8-bit integer precision. The results show that the INT8 model reduces size by 60%, improves processing speed by a factor of three, and achieves comparable hologram quality to the FP32 model. This work enables the practical deployment of 2D-to-3D CGH model on low-power platforms, bridging the gap between high-performance holographic computation and real-world wearable display systems.