GS-RNN: a phase-only hologram generation model based on the fusion of the Gerchberg-Saxton algorithm and a neural network

XianFei Hu , Jinbin Gui , Lei Hu

Optics Express +

|

June 14, 2025

View abstract on PubMed

Summary

We introduce GS-RNN, a novel recurrent neural network model that generates high-quality phase-only holograms (POH) faster than traditional methods. This computer-generated hologram (CGH) approach enhances iterative hologram generation efficiency.

Area Of Science

Optics and Photonics
Computer Vision
Artificial Intelligence

Background

Traditional Gerchberg-Saxton (GS) algorithm for hologram generation can be computationally intensive.
Phase-only holograms (POH) are crucial for various optical applications.
Improving the efficiency and quality of computer-generated holograms (CGH) is an ongoing research area.

Purpose Of The Study

To develop a novel iterative generation model for computer-generated holograms (CGH).
To enhance the efficiency and quality of phase-only holograms (POH) generation.
To reduce the number of iterations required in hologram generation algorithms.

Main Methods

Proposed a recurrent neural network (RNN) architecture combined with an improved residual complementary neural network.
Integrated trainable neural network parameters into the traditional Gerchberg-Saxton (GS) algorithm, creating the GS-RNN model.
Trained the model using sample data to learn optimal parameters for hologram generation.

Main Results

The GS-RNN model achieved a three-fold increase in iterative computation efficiency compared to the traditional GS algorithm.
The model successfully generated high-quality phase-only holograms (POH).
GS-RNN demonstrated robustness, maintaining performance even when physical parameters for POH generation were altered, without retraining.

Conclusions

The proposed GS-RNN model significantly improves the efficiency of iterative hologram generation.
The integration of RNNs with the GS algorithm offers a powerful approach for generating high-quality CGH.
Experimental validation confirmed the effectiveness and practical applicability of the GS-RNN model.