Adaptive 3D descattering with a dynamic synthesis network.

Area of Science:

• Computational imaging
• Deep learning
• Optical physics

Background:

• Deep learning is used for image recovery in scattering media, often requiring separate expert networks for each condition.
• Expert networks perform poorly when testing conditions differ from training, limiting their generalizability.
• Generalist networks require large datasets and complex architectures to handle diverse scattering conditions.

Purpose of the Study:

• To develop an adaptive deep learning framework for robust image recovery across a wide range of scattering conditions.
• To introduce a novel "mixture of experts" architecture for dynamic network synthesis.
• To demonstrate the framework's effectiveness in holographic 3D particle imaging.

Main Methods:

• Proposed a dynamic synthesis network (DSN) with a gating network to blend multiple "expert" networks.
• Trained the DSN using simulated data covering a continuum of scattering conditions.
• Evaluated the DSN's performance in holographic 3D particle imaging simulations and experiments.

Main Results:

• The DSN demonstrated generalization across a continuum of scattering conditions in simulations.
• The DSN trained solely on simulated data generalized effectively to experimental holographic 3D particle imaging.
• Achieved robust 3D descattering and image recovery under various scattering scenarios.

Conclusions:

• The dynamic synthesis network (DSN) offers a highly adaptive deep learning solution for imaging in scattering media.
• The "mixture of experts" approach enables dynamic network synthesis, improving generalization.
• This framework opens new possibilities for adaptive computational imaging techniques beyond descattering.