Universal linear intensity transformations using spatially incoherent diffractive processors.

Area of Science:

• Optics and Photonics
• Optical Engineering
• Computational Imaging

Background:

• Diffractive optical networks can perform complex linear transformations under spatially coherent light.
• Designing such networks for spatially incoherent light, which is more common in natural environments, presents unique challenges.

Purpose of the Study:

• To design and demonstrate a diffractive optical processor capable of arbitrary linear intensity transformations using spatially incoherent light.
• To explore the potential of these networks for all-optical processing tasks, including image classification.

Main Methods:

• Utilized numerical simulations and deep learning for supervised training of diffractive networks.
• Investigated the relationship between the number of diffractive features (N) and the input/output fields-of-view (N_i, N_o).
• Designed networks for simultaneous processing at multiple wavelengths and for handwritten digit classification.

Main Results:

• Demonstrated that spatially incoherent diffractive networks can approximate arbitrary linear intensity transformations.
• Achieved >95% test accuracy in classifying handwritten digits using an all-optical diffractive network under incoherent illumination.
• Showcased the capability of these networks to process intensity information across multiple wavelengths simultaneously.

Conclusions:

• Spatially incoherent diffractive networks offer a viable approach for all-optical linear intensity transformations.
• These networks hold significant promise for developing visual processors that operate effectively under natural, incoherent light.
• The demonstrated digit classification highlights the practical applicability of this technology in machine vision.