Linear Combinations of Patches are Unreasonably Effective for Single-Image Denoising

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Deep neural networks have advanced image denoising but require high-quality training data.
• Single-image denoising methods aim to denoise images without external datasets or prior knowledge.

Purpose of the Study:

• To investigate the effectiveness of linear combinations of patches for single-image denoising.
• To develop a parametric approach for estimating denoising weights using pilot images.

Main Methods:

• Utilized linear combinations of image patches for denoising.
• Employed quadratic risk approximation with multiple pilot images to estimate combination weights.
• Tested on artificially and real-world noisy images.

Main Results:

• Linear combinations of patches achieve state-of-the-art single-image denoising performance.
• The proposed method is competitive with leading single-image denoisers.
• Outperformed recent neural network-based techniques in speed and interpretability.

Conclusions:

• Simple linear combinations of patches are sufficient for effective single-image denoising.
• The developed parametric approach offers a fast, interpretable, and high-performing alternative to data-dependent methods.