Vision transformer empowered physics-driven deep learning for omnidirectional three-dimensional holography.

Area of Science:

• Photonics
• Computer Vision
• Holography

Background:

• 3D holograms face challenges like inter-plane crosstalk and limited axial resolution.
• Current methods to reduce crosstalk often lead to unidirectional holograms or speckle noise.
• Existing deep learning methods, primarily using CNNs, struggle with hologram display quality and axial resolution.

Purpose of the Study:

• To develop an advanced method for generating high-quality, omnidirectional 3D holograms.
• To overcome the limitations of existing techniques in holographic display technology.
• To improve both inter-plane crosstalk and axial resolution in reconstructed 3D holograms.

Main Methods:

• A physics-driven deep neural network empowered by a vision transformer (ViT) was developed.
• The global attention mechanism inherent in ViT was leveraged to address holographic display challenges.
• The model focuses on generating omnidirectional 3D computer-generated holograms (CGH).

Main Results:

• The proposed method successfully generates omnidirectional 3D holograms.
• The use of ViT resulted in significantly reduced inter-plane crosstalk.
• High axial resolution was achieved in the reconstructed 3D holograms.

Conclusions:

• The vision transformer-empowered network offers a novel solution for high-quality 3D holographic displays.
• This approach enhances holographic display performance by minimizing crosstalk and improving axial resolution.
• The work paves the way for complex inverse design applications in photonics.