Holographic image denoising for dense droplet field using conditional diffusion model.

Area of Science:

• Optics and Photonics
• Image Processing
• Machine Learning

Background:

• Holographic imaging is crucial for visualizing 3D structures, but suffers from noise like twin-images and speckles.
• Dense particle fields with large depth of field (DOF) present significant denoising challenges.
• Existing denoising methods often struggle with complex holographic noise patterns.

Purpose of the Study:

• To develop a novel generative approach for holographic image denoising.
• To suppress twin-image and speckle noises in dense particle fields with large DOF.
• To evaluate the proposed method's effectiveness and generalization capabilities.

Main Methods:

• A conditional diffusion model framework inspired by generative paradigms was developed.
• The model was trained and configured for specific holographic denoising tasks.
• Evaluation involved calibration dot board and droplet field data (gel atomization, aviation kerosene swirl spray) using inline and off-axis holography.
• Performance was assessed using three distinct quantitative metrics.

Main Results:

• The proposed method demonstrated superior noise reduction capabilities.
• It effectively preserved fine details within the holographic images.
• The model exhibited strong generalization across different holographic datasets and noise types.
• Performance surpassed two other comparative methods.

Conclusions:

• The generative conditional diffusion model represents a pioneering approach to holographic image denoising.
• The method offers significant advantages in noise suppression and detail preservation.
• Its reduced dependency on high-quality training labels suggests strong potential for industrial applications.