Texture Variation Adaptive Image Denoising With Nonlocal PCA.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Image textures contain crucial high-frequency variations vital for human visual perception.
• Preserving these fine details, especially in stochastic or small-scale textures, remains a challenge for current denoising algorithms.
• State-of-the-art methods often use nonlocal approaches but struggle with texture fidelity.

Purpose of the Study:

• To develop an image denoising technique that effectively preserves texture variations.
• To improve upon existing nonlocal denoising methods by enhancing texture preservation capabilities.
• To address the limitations of current algorithms in handling high-frequency texture details.

Main Methods:

• An adaptive clustering method for grouping image patches with similar texture elements (texels).
• An over-clustering-and-iterative-merging approach to enhance noise robustness in patch grouping.
• A proposed PCA-transform-domain variation adaptive filtering method to preserve local texture variations within principal component analysis (PCA) coefficients.

Main Results:

• The proposed method demonstrates superior texture-preserving denoising compared to traditional PCA-based filtering techniques.
• Experiments show significant improvements in both quantitative and visual evaluations, particularly for irregular textures.
• The effectiveness is validated through denoising natural images and camera raw images.

Conclusions:

• The developed denoising method offers favorable texture-preserving performance.
• The novel PCA-transform-domain filtering effectively captures and preserves local texture variations.
• The approach shows promise for applications requiring high-fidelity texture retention in image denoising.