Pol2Pol: self-supervised polarimetric image denoising.

Area of Science:

• Computer Vision
• Image Processing
• Optics

Background:

• Polarimetric imaging captures rich information but is susceptible to noise.
• Traditional denoising methods often require paired noisy and clean images.
• Self-supervised learning offers a promising alternative for image denoising.

Purpose of the Study:

• To develop a self-supervised method for polarimetric image denoising.
• To enable effective denoising using only single, noisy polarimetric images.
• To create an extensible and compatible denoising framework.

Main Methods:

• Proposed a novel polarization generator to create synthetic training pairs from single noisy images.
• Exploited inherent polarization relationships for data synthesis.
• Adapted existing supervised denoising networks for the self-supervised Pol2Pol framework.

Main Results:

• Pol2Pol effectively denoises polarimetric images with a single noisy input.
• The method outperforms existing self-supervised denoising techniques.
• Achieved performance comparable to traditional supervised denoising methods.

Conclusions:

• Self-supervised polarimetric image denoising is feasible and effective.
• The Pol2Pol method offers a flexible and powerful approach to image denoising.
• This technique reduces the need for extensive paired training data in polarimetric imaging.