Ultra-Wide-Field Noninvasive Imaging Through Scattering Media Via Physics-Guided Deep Learning.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Noninvasive imaging through scattering media is vital but limited by narrow fields of view (FOV).
• Existing learning-based methods require extensive real datasets and struggle with FOV beyond the optical memory effect (OME).

Purpose of the Study:

• To develop an ultra-wide-field noninvasive imaging method for scattering media.
• To overcome the limitations of current FOV constraints and data dependency in scattering imaging.

Main Methods:

• Proposed UNI-Net, a physics-guided adaptive dual-domain diffusion model.
• Synthesized large-scale pre-training data using a physical scattering imaging model.
• Employed multi-channel patch partitioning of speckle patterns and a spatial-channel parallel attention block.

Main Results:

• Reduced reliance on real experimental data by an order of magnitude.
• Achieved a PSNR of 31.23 dB at a 41 × OME range, a 49.5% improvement over existing methods.
• Successfully reconstructed complex scenes at an extreme 164 × OME range with a PSNR of 27.21 dB.

Conclusions:

• UNI-Net offers a robust solution for ultra-wide-field noninvasive imaging through scattering media.
• The physics-guided approach and novel attention mechanism enable high performance with reduced data requirements.
• This method significantly advances the capabilities of imaging through scattering environments.