Phase-only hologram generated by a convolutional neural network trained using low-frequency mixed noise.

Area of Science:

• Optics
• Computer Vision
• Machine Learning

Background:

• Computer-generated holography (CGH) traditionally relies on real images for training.
• Convolutional Neural Networks (CNNs) offer a promising approach for hologram generation.
• Existing CNN-based methods can be limited by training data requirements and artifact generation.

Purpose of the Study:

• To propose a novel phase-only hologram generation method using a CNN trained with a unique dataset.
• To improve the quality and reduce artifacts in reconstructed holographic images.
• To demonstrate a flexible and efficient approach for training hologram-generating CNNs.

Main Methods:

• A phase-only hologram was generated using a CNN trained with a custom dataset named low-frequency mixed noise (LFMN).
• The LFMN dataset comprises various noise images processed at low frequencies, replacing conventional real images for CNN training.
• The proposed CNN model was trained and evaluated on the DIV2K valid dataset.

Main Results:

• The proposed method achieved a hologram generation speed of 0.094 s/frame for 2160 × 3840 pixels.
• The average peak signal-to-noise ratio (PSNR) of the reconstructed images was approximately 29.2 dB.
• Optical experiments confirmed the theoretical predictions, showing superior reconstructed image quality and reduced artifacts compared to conventional methods.

Conclusions:

• The LFMN dataset provides a simple and flexible approach for training CNNs for hologram generation.
• The proposed CNN-based method significantly enhances the quality of reconstructed holographic images.
• This technique effectively mitigates artifacts, offering a substantial improvement over traditional CGH methods.