Displacement-agnostic coherent imaging through scatter with an interpretable deep neural network

Area of Science:

• Optics
• Computational Imaging
• Machine Learning

Background:

• Coherent imaging through scattering media is a significant challenge in various scientific fields.
• Existing model-based and data-driven methods have limitations in handling complex scattering scenarios.
• Previous deep learning approaches showed promise but lacked generalizability to unseen perturbations.

Purpose of the Study:

• To develop a robust and interpretable deep learning model for imaging through scattering media.
• To create a model agnostic to a wider range of perturbations, including scatterer changes and defocus.
• To analyze and visualize the model's generalization mechanisms.

Main Methods:

• Proposed a novel deep neural network architecture for inverse scattering problems.
• Developed an analysis framework using unsupervised dimension reduction for model interpretability.
• Tested the model's performance against various perturbations like scatterer changes, displacements, and system defocus.

Main Results:

• The deep neural network achieved high-quality predictions even with unseen diffusers.
• The model demonstrated robustness to scatterer change, displacements, and up to 10x depth of field defocus.
• Analysis revealed the model effectively unmixes scattering information and extracts object-specific details.

Conclusions:

• The developed deep learning model offers a robust and interpretable solution for imaging through scattering media.
• The approach generalizes well under diverse scattering conditions, outperforming previous methods.
• This work advances the application of AI in overcoming physical imaging limitations.