Noise-Resistant Demosaicing with Deep Image Prior Network and Random RGBW Color Filter Array

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Conventional RGB color filter arrays (CFAs) limit light transmission.
• Existing demosaicing methods struggle with noise and low-light conditions.
• Random RGBW CFAs offer improved light capture but pose reconstruction challenges.

Purpose of the Study:

• To develop a deep-image-prior-based demosaicing method for random RGBW CFAs.
• To enable color image reconstruction using only the RGBW CFA image for training.
• To improve image quality and noise resilience in demosaicing.

Main Methods:

• Utilized a deep image prior (DIP) network trained solely on the RGBW CFA image.
• Developed a specialized loss function to incorporate information from white pixels.
• Employed a single-shot reconstruction approach without complex auxiliary algorithms.

Main Results:

• Achieved superior color image reconstruction quality compared to existing methods.
• Demonstrated enhanced performance in the presence of noise and low-light conditions.
• Validated the method's effectiveness for joint demosaicing and denoising.

Conclusions:

• The proposed DIP-based demosaicing method effectively reconstructs color images from random RGBW CFAs.
• The method offers significant advantages in noisy and low-light imaging scenarios.
• This work represents a pioneering neural network approach for RGBW CFA demosaicing with single-image training.