Deep Graph-Convolutional Image Denoising.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Non-local self-similarity is a key prior for effective image denoising.
• Convolutional neural networks (CNNs) excel at image denoising but typically use local information.
• Integrating non-local priors into CNNs for denoising remains an underexplored area.

Purpose of the Study:

• To propose a novel end-to-end trainable neural network architecture for image denoising.
• To incorporate non-local self-similarity priors into CNNs using graph convolution operations.
• To enhance the representation learning capabilities of neural networks for uncovering self-similar patterns.

Main Methods:

• Developed a neural network architecture with layers based on graph convolution operations.
• Generalized classic convolution to arbitrary graphs, creating neurons with non-local receptive fields.
• Dynamically computed graphs from hidden feature similarities and introduced lightweight Edge-Conditioned Convolution to address gradient and parameterization issues.

Main Results:

• Achieved state-of-the-art performance in image denoising tasks.
• Demonstrated improved qualitative and quantitative results compared to existing methods.
• Successfully applied the method to both synthetic Gaussian noise and real-world noise scenarios.

Conclusions:

• The proposed graph convolution-based neural network effectively leverages non-local self-similarity for superior image denoising.
• The novel architecture enhances CNNs' ability to learn from non-local image priors.
• The method offers a promising direction for advancing image restoration techniques.