Low-Light Sparse Polarization Demosaicing Network (LLSPD-Net): Polarization Image Demosaicing Based on Stokes Vector...

Area of Science:

• Optics and Photonics
• Computer Vision
• Image Processing

Background:

• Polarization imaging is vital for applications like autonomous driving but suffers in low light due to low photon counts and conventional sensor limitations.
• Traditional demosaicing methods struggle with noise and distortion in low-light polarization imaging, necessitating improved techniques.

Purpose of the Study:

• To propose a novel network, the Low-Light Sparse Polarization Demosaicing Network (LLSPD-Net), for simulating and processing polarization images under low-light conditions.
• To develop a polarization demosaicing method that avoids traditional interpolation by utilizing Stokes vector complementation guided by enhanced intensity images.

Main Methods:

• The LLSPD-Net comprises an intensity image enhancement network and a Stokes vector complementation network.
• The intensity enhancement network generates high-quality RGB images from low-light inputs.
• A novel polarization demosaicing approach uses the enhanced image to guide Stokes vector completion, replacing traditional interpolation.

Main Results:

• The proposed LLSPD-Net effectively enhances low-light polarization images and performs demosaicing.
• Experiments on self-constructed and public datasets demonstrate superior performance compared to state-of-the-art methods.
• The method shows significant improvements in both qualitative and quantitative evaluations for low-light polarization image processing.

Conclusions:

• LLSPD-Net offers a robust solution for polarization imaging in low-light environments.
• The novel Stokes vector complementation method overcomes limitations of traditional polarization demosaicing.
• This work advances the field of polarization imaging, enabling better performance in challenging lighting conditions.