Weighted Tensor Rank-1 Decomposition for Nonlocal Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Numerical Analysis

Background:

• Natural images exhibit significant local and nonlocal self-similarity.
• Low-rank models are effective for image denoising, but tensor extensions are challenging due to NP-hard decomposition.
• Existing methods struggle to fully leverage tensor properties for image denoising.

Purpose of the Study:

• To propose a novel weighted tensor rank-1 decomposition (WTR1) method for nonlocal image denoising.
• To effectively utilize both local and nonlocal self-similarity in low-rank tensor models.
• To develop an efficient tensor decomposition algorithm for low-rank approximation.

Main Methods:

• Grouping similar image patches into 3-D stacks.
• Converting patch stacks into a finite sum of rank-1 products.
• Employing intrinsic low-rank tensor approximation via a novel CANDECOMP/PARAFAC (CP) decomposition algorithm.

Main Results:

• The WTR1 method successfully exploits local and nonlocal self-similarity.
• Achieved improved nonlocal image denoising quality compared to existing methods.
• Experimental results demonstrate superior performance over state-of-the-art denoising techniques.

Conclusions:

• The proposed WTR1 method offers an effective approach for nonlocal image denoising.
• WTR1 provides a robust strategy for low-rank tensor approximation in image processing.
• The method demonstrates significant potential for enhancing image quality in practical applications.