Dynamic polarization fusion network (DPFN) for imaging in different scattering systems

Area of Science:

• Optics and Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning excels in imaging through scattering media, with polarization offering stable light characteristics.
• Training specialized networks for diverse scattering conditions is often impractical due to data limitations.
• Generalist networks require larger models and datasets to handle varied polarization data.

Purpose of the Study:

• To develop a flexible dynamic learning framework for de-scattering imaging.
• To enable adaptation to diverse environmental scattering conditions.
• To optimize the selection of polarization characteristics for improved imaging.

Main Methods:

• Introduced a dynamic learning framework that adaptively adjusts weights of polarization components.
• Incorporated a Gating Network (GTN) to integrate multiple polarization features.
• Trained and evaluated the network across a range of continuous scattering conditions.

Main Results:

• The dynamic learning framework demonstrated robust generalization across continuous scattering environments.
• The Gating Network effectively integrated diverse polarization information for scenario-specific adaptation.
• The proposed method achieved enhanced de-scattering performance compared to traditional approaches.

Conclusions:

• The dynamic learning framework offers a flexible and effective solution for de-scattering imaging in varied conditions.
• Adaptive integration of polarization features is crucial for robust performance in complex scattering media.
• This approach facilitates the selection of optimal polarization characteristics for optimal imaging outcomes.