Variational based Mixed Noise Removal with CNN Deep Learning Regularization.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Mixed noise removal, particularly Gaussian mixture and Gaussian-impulse noise, presents a significant challenge in image processing.
• Existing methods often struggle to accurately discriminate between different noise types and their levels at a pixel-by-pixel basis.

Purpose of the Study:

• To develop an integrated approach combining variational methods and deep learning for robust mixed noise removal.
• To accurately classify noise types and levels for improved image reconstruction.

Main Methods:

• A novel variational method is proposed for iterative noise parameter estimation and automatic noise classification.
• Operator splitting is employed to decompose the variational problem into four subproblems: regularization, synthesis, parameter estimation, and noise classification.
• A deep learning approach, specifically a Convolutional Neural Network (CNN), is utilized to learn natural image priors for regularization.

Main Results:

• The integrated method demonstrates superior performance in removing Gaussian mixture and Gaussian-impulse noise compared to existing techniques.
• The CNN regularizer significantly enhances the quality of restored images.
• The synthesis step, informed by noise analysis, yields better reconstructions.

Conclusions:

• The proposed method effectively addresses the challenge of mixed noise removal by integrating model-based and learning-based techniques.
• The framework's ability to classify noise types and levels allows for more accurate image restoration.
• The approach is extensible to various image reconstruction and inverse problems, with the CNN acting as a variational functional operator.