Prior-free imaging unknown target through unknown scattering medium.

Area of Science:

• Optics
• Computational Imaging
• Machine Learning

Background:

• Deep learning methods for imaging through scattering media often rely on paired data and lack physical process integration.
• Existing approaches struggle with reconstructing hidden targets without pre-trained networks and exhibit limited generalization.

Purpose of the Study:

• To propose an unsupervised neural network that integrates a universal physical process for imaging through scattering media.
• To develop a method that reconstructs hidden targets using only a single speckle pattern and unpaired target data, independent of specific optical systems.

Main Methods:

• An unsupervised neural network architecture is proposed, incorporating a universal physical process for image reconstruction.
• The network optimizes online using the physical process, bypassing the need for extensive paired training data.
• Reconstruction is achieved with a single speckle pattern and unpaired target data, demonstrating system and prior information independence.

Main Results:

• The proposed network enables reconstruction without large-scale paired data or prior information.
• Online optimization is achieved through physical process integration, enhancing generalization capabilities.
• The method demonstrates universal applicability across different optical systems.

Conclusions:

• The developed unsupervised neural network effectively images through scattering media by integrating physical processes.
• This approach overcomes limitations of data-driven methods, offering improved generalization and practical applicability without extensive prior data.
• The method's independence from specific system parameters and its universal applicability make it a promising solution for real-world scattering imaging challenges.