Non-invasive imaging through scattering medium and around corners beyond 3D memory effect.

Area of Science:

• Computational imaging
• Optical physics
• Machine learning

Background:

• The three-dimensional (3D) memory effect (ME) is limited in scope, restricting speckle correlation technology to small imaging fields of view (FOV) and depths of field (DOF).
• Existing methods struggle to image targets beyond the 3D ME range or in complex scattering environments.

Purpose of the Study:

• To develop a novel method for restoring targets beyond the 3D ME range using an untrained neural network.
• To enable reconstruction of hidden targets and targets around corners in scattering media.

Main Methods:

• An untrained neural network was constructed as an optimization tool.
• The autocorrelation consistency relationship and generative adversarial strategy were combined.
• Online optimization was performed using single-frame speckle and unaligned real targets, eliminating the need for prior training or additional system modulation.

Main Results:

• The proposed method successfully restored targets beyond the 3D ME range.
• Reconstruction of hidden targets and targets around corners was achieved.
• The neural network did not require pre-training, simplifying the experimental setup.

Conclusions:

• The combination of a generative adversarial framework with physical priors effectively decouples aliasing information and reconstructs targets.
• This approach overcomes the limitations of traditional speckle correlation technology in terms of FOV and DOF.
• The method offers a promising new direction for computational imaging, particularly for non-line-of-sight imaging applications.