Considering Image Information and Self-Similarity: A Compositional Denoising Network.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) are prevalent in image denoising, often enhanced by residual learning.
• Existing research primarily optimizes CNN architectures, overlooking limitations in residual learning, such as neglecting image information and self-similarity.

Purpose of the Study:

• To address the limitations of conventional residual learning in image denoising.
• To propose a novel compositional denoising network (CDN) that integrates image information and self-similarity for enhanced performance.

Main Methods:

• Developed a compositional denoising network (CDN) with two sub-paths: an image information path (IIP) and a noise estimation path (NEP).
• Trained IIP using an image-to-image approach for image information extraction.
• Employed a similarity-based training strategy for NEP to leverage image self-similarity for noise distribution estimation.

Main Results:

• The proposed CDN integrates image information and noise distribution estimation for comprehensive denoising.
• CDN demonstrated superior performance compared to existing CNN-based methods on both synthetic and real-world noisy images.
• Achieved state-of-the-art results in image denoising tasks.

Conclusions:

• The compositional denoising network (CDN) effectively overcomes limitations of standard residual learning in image denoising.
• CDN's dual-path approach, utilizing image information and self-similarity, offers a significant advancement in noise reduction.
• The findings suggest CDN as a powerful new method for state-of-the-art image denoising.