[Paper Review] NAMF: A Non-local Adaptive Mean Filter for Salt-and-Pepper Noise Removal

Houwang Zhang, Zhu, Yuan+1 more|arXiv (Cornell University)|Oct 17, 2019

Image and Signal Denoising Methods21 references4 citations

TL;DR

This paper proposes NAMF, a non-local adaptive mean filter for salt-and-pepper (SAP) noise removal, which uses an adaptive window to detect noisy pixels based on intensity proportion and replaces them with neighboring mean values, followed by a SAP-specific non-local mean filter for enhanced restoration. NAMF achieves superior PSNR (up to 41.31 dB) and SSIM (0.9932) at 90% noise levels, outperforming state-of-the-art methods in detail preservation and image quality.

ABSTRACT

In this paper, a novel algorithm called a non-local adaptive mean filter (NAMF) for removing salt-and-pepper (SAP) noise from corrupted images is presented. We employ an efficient window detector with adaptive size to detect the noise, the noisy pixel will be replaced by the combination of its neighboring pixels, and finally we use a SAP noise based non-local mean filter to reconstruct the intensity values of noisy pixels. Extensive experimental results demonstrate that NAMF can obtain better performance in terms of quality for restoring images at all levels of SAP noise.

Motivation & Objective

• To address the degradation of image details and jagged boundaries in existing SAP noise removal methods under high noise intensity.
• To improve restoration quality by distinguishing noise pixels from texture pixels more accurately during detection.
• To enhance image reconstruction using a non-local mean filter tailored for SAP noise characteristics.
• To maintain high processing efficiency while achieving superior denoising performance across all noise levels.

Proposed method

• An adaptive window detector identifies potential noisy pixels by checking if their gray values match the global min/max of the image.
• A proportion-based thresholding method confirms noise by evaluating the ratio of identical-value pixels within the adaptive window.
• Noise-affected pixels are replaced with the mean intensity of their spatially neighboring pixels.
• A novel SAP noise-aware non-local mean filter further refines the denoised image by leveraging similarity across distant image patches.
• The method uses a 41×41 similarity window and a 5×5 search window, with parameters tuned via empirical fitting.
• The threshold for noise confirmation is set to T=0.8 after experimental validation.

Experimental results

Research questions

• RQ1 Can a proportion-based adaptive window method improve noise detection accuracy by reducing false positives on texture regions?
• RQ2 Does integrating a non-local mean filter specifically designed for SAP noise enhance image restoration quality beyond local filtering?
• RQ3 How does NAMF perform in preserving fine details and structural information at high SAP noise levels (e.g., 90%) compared to existing methods?
• RQ4 What is the trade-off between denoising performance and computational efficiency in the proposed NAMF framework?

Key findings

• NAMF achieves a PSNR of 41.3133 dB and SSIM of 0.9932 on the 'Lena' image with 90% SAP noise, significantly outperforming all compared methods.
• At 10% noise, NAMF’s PSNR is 15.4199 dB, and at 90% noise, it maintains a PSNR of 27.5748 dB, demonstrating robustness across noise levels.
• The average PSNR and SSIM curves show that NAMF consistently ranks highest, with a growing performance gap over other methods as noise intensity increases beyond 30%.
• NAMF’s running time is comparable to state-of-the-art methods, ranking in the middle but outperforming NAFSMF and matching DAMF in speed at high noise levels.
• Visual inspection confirms that NAMF preserves edges and textures better than AMF, AWMF, BPDF, and other methods, especially at high noise densities.
• The method effectively avoids blurring and jagged artifacts in restored images, even when over 90% of pixels are corrupted.