Dual Mixture Model Based CNN for Image Denoising.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Non-Gaussian residual errors and noise are prevalent in real-world image applications.
• Traditional methods struggle to integrate these noise types into Convolutional Neural Network (CNN) architectures for image denoising.
• Existing CNN-based denoising methods often lack robust handling of non-Gaussian noise characteristics.

Purpose of the Study:

• To propose a novel deep learning approach for effectively handling non-Gaussian residual errors in image denoising.
• To develop a CNN architecture capable of addressing non-Gaussian noise with non-uniform spatial distributions.
• To improve image denoising performance on challenging noise types compared to existing methods.

Main Methods:

• Leveraging the universal approximation property of probability density functions for non-Gaussian errors.
• Employing a maximum likelihood estimation duality to derive an adaptive weighting strategy for image fidelity.
• Integrating a learnable regularizer for enhanced image priors.
• Unrolling the iterative solution into a weighted residual CNN architecture.

Main Results:

• The proposed weighted residual block effectively handles non-Gaussian residuals, particularly noise with non-uniform spatial distribution.
• Numerical results demonstrate superior performance in removing non-Gaussian noise (e.g., Gaussian mixture, random-valued impulse noise) compared to existing methods.
• The adaptive weighting strategy enhances the fidelity of denoised images.

Conclusions:

• The developed deep learning approach offers a robust solution for image denoising under non-Gaussian noise conditions.
• The weighted residual CNN architecture provides a significant advancement in handling complex noise patterns.
• This method shows promise for real-world applications where non-Gaussian noise is a significant challenge.