Related Experiment Video
Updated: Jul 25, 2025
10:16
Digital Inline Holographic Microscopy DIHM of Weakly-scattering Subjects
Published on: February 8, 2014
12.3K
Diffraction model-driven neural network trained using hybrid domain loss for real-time and high-quality
Optics Express
|June 29, 2023
Summary
This study introduces Res-Holo, a new diffraction model-driven neural network for generating high-quality phase-only holograms (POHs). It significantly improves image quality and reduces artifacts in holographic displays.
Area of Science:
- Optics
- Computer Science
- Artificial Intelligence
Background:
- Learning-based computer-generated holography (CGH) shows promise for real-time holographic displays.
- Existing methods struggle with high-quality hologram generation due to CNN limitations in cross-domain tasks.
Purpose of the Study:
- To develop a novel neural network for generating high-fidelity phase-only holograms (POHs).
- To enhance hologram quality by addressing limitations in current learning-based CGH algorithms.
Main Methods:
- Introduced Res-Holo, a diffraction model-driven neural network.
- Utilized pretrained ResNet34 weights for feature extraction and overfitting prevention.
- Incorporated hybrid domain loss (spatial and frequency) for improved hologram generation.
Main Results:
- Achieved a 6.05 dB improvement in peak signal-to-noise ratio (PSNR) using hybrid domain loss.
- Generated high-fidelity 2K resolution POHs with an average PSNR of 32.88 dB.
- Demonstrated effective image quality improvement and artifact suppression in optical experiments.
Conclusions:
- Res-Holo effectively generates high-quality POHs, outperforming existing methods.
- The hybrid domain loss strategy is crucial for enhancing holographic image fidelity.
- The method shows practical applicability in both monochrome and full-color holographic displays.
