Multipatch Unbiased Distance Non-Local Adaptive Means With Wavelet Shrinkage

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Existing non-local means (NLM) methods often exhibit limited denoising performance due to fixed weights or reliance on Euclidean distance.
• The fixed nature of weights and the limitations of Euclidean distance hinder the robustness and accuracy of traditional NLM techniques.

Purpose of the Study:

• To propose novel non-local adaptive means (NLAM) algorithms for enhanced image denoising.
• To introduce unbiased distance metrics for more robust similarity measurements between image patches.
• To develop adaptive multi-patch strategies and integrate wavelet shrinkage for superior denoising results.

Main Methods:

• Iterative updating of weights (ωij) as an optimization variable within the NLAM framework.
• Introduction of three unbiased distances: pixel-pixel, patch-patch, and coupled unbiased distances.
• Development of unbiased distance non-local adaptive means (UD-NLAM) and its multi-patch variant (MUD-NLAM), incorporating wavelet shrinkage (MUD-NLAM-WS).

Main Results:

• NLAM, UD-NLAM, and MUD-NLAM demonstrate superior performance compared to existing NLM methods.
• MUD-NLAM-WS achieves state-of-the-art denoising performance, outperforming current advanced methods.
• The proposed unbiased distance metrics offer greater robustness in measuring image pixel/patch similarity.

Conclusions:

• The proposed NLAM framework, particularly UD-NLAM and MUD-NLAM, significantly improves image denoising efficacy.
• MUD-NLAM-WS represents a new benchmark in image denoising by effectively combining adaptive multi-patch processing and wavelet shrinkage.
• Adaptive weight updating and unbiased distance metrics are crucial for robust and high-performance image denoising.