Iterative Joint Image Demosaicking and Denoising using a Residual Denoising Network.

Area of Science:

• Digital Image Processing
• Computer Vision
• Machine Learning

Background:

• Digital cameras perform sequential denoising and demosaicking, which are often ill-posed problems.
• Existing machine learning solutions for joint denoising-demosaicking typically use generic architectures lacking physical image model considerations.

Purpose of the Study:

• To develop a novel algorithm for joint denoising-demosaicking that addresses limitations of current methods.
• To create a transparent and interpretable algorithm by combining classical image regularization, optimization, and deep learning.

Main Methods:

• Developed an iterative optimization algorithm incorporating a trainable denoising network.
• Inspired the algorithm by classical image regularization and large-scale optimization techniques.
• Designed a principled denoising network architecture.

Main Results:

• The proposed method outperforms existing approaches in reconstruction quality for both noisy and noise-free data across diverse datasets.
• The algorithm demonstrates superior performance due to its rigorous derivation and principled network design.
• Achieved state-of-the-art results with fewer trainable parameters and significantly less training data.

Conclusions:

• The novel joint denoising-demosaicking algorithm offers a transparent and effective alternative to black-box deep learning methods.
• The principled approach leads to superior image reconstruction and improved efficiency in terms of parameters and training data.
• This work advances the field of image processing by providing a more robust and data-efficient solution for denoising and demosaicking.