Deep Demosaicing for Polarimetric Filter Array Cameras.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Polarisation Filter Array (PFA) cameras offer cost-effective light polarization analysis.
• Demosaicing PFA images requires accounting for local filter variations and scene characteristics.
• Non-linear effects like cross-talk challenge traditional demosaicing methods.

Purpose of the Study:

• To develop a data-driven approach for demosaicing raw PFA images directly to per-pixel Stokes vectors.
• To overcome challenges in acquiring ground-truth polarization data for training.
• To improve the accuracy of polarization state analysis from PFA camera imagery.

Main Methods:

• A novel CNN architecture utilizing Mosaiced Convolutions tailored to PFA filter arrangements.
• A new data acquisition technique using a consumer LCD screen, invariant to monitor gamma and lighting.
• Direct demosaicing of raw PFA images to per-pixel Stokes vectors.

Main Results:

• The proposed CNN model consistently outperformed existing algorithmic and learning-based demosaicing techniques.
• Significant reduction in error, especially for polarization angle estimation.
• Successful acquisition of real-world training data using the novel LCD screen method.

Conclusions:

• The CNN-based approach provides a robust and accurate solution for PFA image demosaicing.
• The novel data acquisition method facilitates training for real-world polarization imaging applications.
• This work advances the capabilities of polarization imaging systems through improved data processing.