Pyramid diffractive optical networks for unidirectional image magnification and demagnification

Area of Science:

• Optics
• Machine Learning
• Optical Engineering

Background:

• Diffractive deep neural networks (D²NNs) are optical processors using diffractive layers and deep learning for computational tasks.
• Existing D²NNs can perform various optical functions but lack directional specificity for imaging.
• Unidirectional optical processing is desirable for applications requiring controlled image manipulation.

Purpose of the Study:

• To develop a novel diffractive optical network architecture for unidirectional image magnification and demagnification.
• To design a compact and efficient optical processor capable of directional imaging.
• To investigate the wavelength robustness and scalability of the proposed architecture.

Main Methods:

• Introduced a pyramid-structured diffractive optical network (P-D²NN) with pyramidally scaled diffractive layers.
• Optimized the P-D²NN using supervised deep learning for unidirectional magnification/demagnification.
• Designed a wavelength-multiplexed P-D²NN for simultaneous bidirectional operations at different wavelengths.
• Validated the P-D²NN experimentally using terahertz illumination.

Main Results:

• The P-D²NN successfully achieved high-fidelity unidirectional image magnification and demagnification.
• The P-D²NN demonstrated robust performance across a wide range of illumination wavelengths, despite single-wavelength training.
• Cascading P-D²NN modules enabled higher magnification factors.
• Experimental validation using terahertz waves confirmed the simulation results.

Conclusions:

• The P-D²NN architecture provides an effective strategy for task-specific visual processors with unidirectional imaging capabilities.
• This physics-inspired design offers a compact and efficient solution for directional optical magnification and demagnification.
• The P-D²NN shows potential for advanced optical computing and imaging applications.