A weighted dictionary learning model for denoising images corrupted by mixed noise

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Image noise significantly degrades visual quality and hinders subsequent analysis.
• Existing denoising methods struggle with complex mixed noise scenarios (e.g., Gaussian-impulse).
• Sparse representations and energy minimization are powerful tools for image restoration.

Purpose of the Study:

• To propose a novel energy minimization model for robust mixed noise removal from images.
• To develop an efficient algorithm for optimizing the proposed model.
• To demonstrate superior performance in preserving image details and textures.

Main Methods:

• A general weighted l(2)-l(0) norms energy minimization model is proposed.
• A new data fidelity function is derived from maximum likelihood estimation for easier optimization.
• Sparse representations over a trained dictionary and a modified K-SVD algorithm are utilized.
• An adaptive weighting function aids in noise detection and parameter estimation.

Main Results:

• The proposed model effectively removes various mixed noise types (Gaussian-Gaussian mixture, impulse, Gaussian-impulse).
• The method demonstrates excellent preservation of image textures and fine details.
• Experimental results show superior performance compared to existing state-of-the-art denoising techniques.

Conclusions:

• The weighted l(2)-l(0) norms model offers a robust and efficient solution for mixed noise image denoising.
• The adaptive weighting function enhances noise detection capabilities.
• The approach provides a significant advancement in preserving image fidelity during noise removal.