The resurgence in analog optical information processing is intertwined with advancements in artificial intelligence (AI), particularly deep learning. Diffractive deep neural networks (D²NNs), a product of this synergy, use light-matter interaction to process visual information. These networks are characterized by spatially engineered surfaces, whose transmission and reflection profiles are optimized using machine learning techniques. Once optimized, these networks are capable of performing a variety of visual computing tasks at light-speed, such as image classification, information encryption, and quantitative phase imaging. Earlier work demonstrated the capacity of spatially coherent D²NNs to perform arbitrary complex-valued linear transformations, contingent on the optimization of a sufficient number of diffractive features. For phase-only networks, this condition intensifies due to reduced degrees of freedom. Extending this to networks operating under spatially incoherent illumination, it was shown that a diffractive network could perform nonnegative linear transformations of optical intensity. However, these transformations were previously limited to one-dimensional inputs and did not cover arbitrary input and output apertures.
A new study published in the Advanced Photonics Nexus by Xilin Yang, Md Sadman Sakib Rahman, Bijie Bai, Jingxi Li, and led by Professor Aydogan Ozcan from the University of California Los Angeles, researchers demonstrated the processing of complex-valued data using compact diffractive optical networks under spatially incoherent illumination. This is achieved by preprocessing the input information through ‘mosaicking’ and ‘demosaicking’ operations, allowing these networks to perform arbitrary complex-valued linear transformations, a significant advancement for applications like all-optical image encryption under natural light.
The authors architecture allows the synthesis of arbitrary complex-valued linear transformations with spatially incoherent D²NNs. The mosaicking process involves finding nonnegative optical intensity-based representations of complex-valued elements. They showed that a spatially incoherent diffractive network, spanning less than 100 wavelengths, can perform any complex-valued linear transformation with negligible error when the number of optimizable diffractive features exceeds a certain threshold. They demonstrated that our spatially incoherent D²NN models can accurately approximate the desired complex-valued linear transformation. This success is further illustrated through the application of these networks in complex number-based image encryption-decryption schemes, which have been validated through rigorous numerical analysis.
The new study showcased the potential of spatially incoherent D²NNs in processing complex-valued data. The flexibility in the system’s design allows for a variety of implementations, potentially enhancing the security of optical information transmission. This is particularly useful in scenarios where the knowledge of mosaicking and demosaicking schemes is crucial for decrypting the transmitted data. Additionally, they explored the use of D²NNs in image encryption applications, demonstrating the feasibility of using these networks for both encryption and decryption processes. This opens up possibilities for creating more secure communication channels using optical systems. However, they focus remained on the numerical analysis of these concepts. While various D²NNs have been experimentally validated across the electromagnetic spectrum, there are challenges associated with fabrication errors and mechanical misalignments. Addressing these challenges is crucial for the practical application of these networks.
In conclusion, Professor Aydogan Ozcan and his group highlighted the capabilities of spatially incoherent diffractive networks in performing complex-valued linear transformations, marking a significant advancement in optical information processing. This technology holds promise for a wide range of applications, including image encryption and computational imaging, especially under natural light conditions.
Xilin Yang, Md Sadman Sakib Rahman, Bijie Bai, Jingxi Li, Aydogan Ozcan. Complex-valued universal linear transformations and image encryption using spatially incoherent diffractive networks. Advanced Photonics Nexus, Vol. 3, Issue 1, 016010 (January 2024). https://doi.org/10.1117/1.APN.3.1.016010