Hybrid deep reconstruction for vignetting-free upconversion imaging through scattering in epsilon-near-zero materials

Area of Science:

• Optics
• Image Reconstruction
• Machine Learning

Background:

• Optical imaging is hindered by scattering in complex media, scrambling spatial and phase information.
• Conventional methods struggle to recover clear images from scattered light.
• Time-gating techniques offer a way to isolate ballistic photons but require advanced reconstruction.

Purpose of the Study:

• To develop a novel deep learning framework for high-fidelity optical image reconstruction from nonlinear scattering measurements.
• To overcome the limitations of scattering in turbid and heterogeneous environments.
• To enhance image quality metrics such as PSNR, SSIM, and IoU.

Main Methods:

• A hybrid-supervised deep learning framework combining a U-Net model (DeepTimeGate) and Deep Image Prior (DIP) refinement.
• Utilizing a time-gated epsilon-near-zero (ENZ) imaging system with four-wave mixing (FWM) in indium tin oxide (ITO) films.
• Acquiring nonlinear scattering measurements and applying the deep learning framework for image reconstruction.

Main Results:

• Achieved significant improvements in peak signal-to-noise ratio (PSNR) by 124% and structural similarity index (SSIM) by 231% compared to raw scattering inputs.
• Demonstrated a 10x improvement in intersection-over-union (IoU) for complex imaging scenarios.
• Successfully removed vignetting and expanded the effective field-of-view beyond the ENZ optical time gate output.

Conclusions:

• The proposed hybrid deep learning framework enables high-fidelity optical imaging through complex media.
• This method offers substantial improvements in image reconstruction quality and expands imaging scope.
• Potential applications include biomedical imaging and in-solution diagnostics where conventional imaging fails.