Enhancing light efficiency in phase-only holograms via neural network

Artificial neural networks have emerged as powerful tools for hologram synthesis and reconstruction, offering improvements in both image quality and computational efficiency. In this work, we present ShuffleResnet, a neural phase encoding approach designed to address the limitations of the conventional double phase encoding method (DPM) used for phase-only spatial light modulators (SLMs). The proposed model, ShuffleResnet, significantly enhances light efficiency, achieving a 59% increase compared to the conventional DPM. Numerical simulation results further demonstrate that the proposed model improves reconstruction quality and effectively suppresses artifacts in the complex field. Additionally, the model encodes complex field holograms at 1920 × 1080 resolution with an average inference speed of 4.74 milliseconds per hologram. The enhanced reconstruction fidelity increased light efficiency and fast inference, suggesting strong potential for real-time holographic applications.