Global Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• State-of-the-art image denoising often relies on patch-based methods, which are limited by patch matching accuracy and diminishing returns with more patches.
• Existing algorithms like BM3D show excellent performance but struggle with increasing image complexity and size.
• Patch-based methods' reliance on local similarity restricts their ability to capture global image structures.

Purpose of the Study:

• To develop a novel image denoising paradigm based on global filtering, estimating each pixel using all image data.
• To provide a statistical analysis of the proposed global filter using spectral decomposition.
• To derive an efficient implementation of the global filter using the Nyström extension for approximating spectral components.

Main Methods:

• Development of a global filtering framework for image denoising.
• Statistical analysis of the global filter's operator via spectral decomposition.
• Application of the Nyström extension to approximate principal spectral components.
• Efficient implementation through strategic pixel sampling.

Main Results:

• The proposed global filter effectively utilizes all image pixels for denoising, surpassing patch-based method limitations.
• Statistical analysis confirms the filter's properties and the impact of spectral decomposition truncation.
• The Nyström extension enables efficient global filtering with a small subset of image pixels.
• Experimental results demonstrate significant improvements over state-of-the-art patch-based denoising techniques.

Conclusions:

• A novel global filtering approach offers a powerful alternative to traditional patch-based image denoising.
• The method provides a statistically grounded and efficiently implementable solution for complex image denoising tasks.
• This strategy can enhance existing denoising filters by enabling global pixel estimation, leading to superior performance.