Imaging through opaque scattering layers via transmission matrix assisted learning

Area of Science:

• Optical Imaging
• Computational Imaging
• Deep Learning Applications

Background:

• Deep learning (DL) methods for imaging through scattering media often lack physical principles.
• Existing DL approaches typically require extensive, complete training datasets for transmission matrix (TM) measurement and image reconstruction.

Purpose of the Study:

• To propose a novel two-stage deep learning network for efficient TM measurement and image reconstruction.
• To overcome the data completeness limitations of current DL-based imaging techniques.
• To enable high-fidelity image reconstruction with significantly reduced training data.

Main Methods:

• A two-stage deep learning network combining TM measurement and image reconstruction.
• A Measurement Stage designed to drastically reduce data requirements.
• An Imaging Stage incorporating a self-closed-loop constraint to eliminate dependence on complete training sets.

Main Results:

• Achieved high-quality image reconstruction with a Structural Similarity Index Measure (SSIM) of 0.84 using only 10 training data pairs.
• Significantly reduced the amount of data needed for TM measurement.
• Demonstrated the capability of both stages to function as stand-alone methods with conventional algorithms.

Conclusions:

• The proposed two-stage network effectively addresses data limitations in DL-based scattering media imaging.
• This method advances transmission matrix-based imaging and offers a practical reference for optical imaging applications.
• The approach facilitates robust image reconstruction even with sparse training data.