Ps-ViT: phase space vision transformer pre-training for the depth estimation in computer-generated holograms

Area of Science:

• Computer Vision
• Holographic Imaging
• Machine Learning

Background:

• Neural network pre-training advances computer vision, especially with limited data.
• Self-supervised learning extracts robust features from large, unlabeled datasets.
• Holography has lagged due to challenges in tailored pre-training strategies.

Purpose of the Study:

• To bridge the gap in applying advanced neural network pre-training to holography.
• To develop an effective pre-training method specifically for holographic data.
• To improve feature descriptor learning for holographic imaging applications.

Main Methods:

• Introduced a novel pre-training method utilizing the hologram phase space representation.
• Employed self-supervised learning techniques adapted for holographic data.
• Focused on learning efficient feature descriptors for dense depth map estimation.

Main Results:

• Successfully learned transferable and robust image feature descriptors from holographic data.
• Demonstrated optimized feature learning for dense depth map estimation in holography.
• Unlocked new potential for holographic imaging applications through improved feature representation.

Conclusions:

• The proposed pre-training method effectively addresses limitations in applying deep learning to holography.
• Leveraging hologram phase space representation is key to learning optimized holographic features.
• This work paves the way for enhanced performance in holographic computer vision tasks.