Scholarlink

<p>The non-local means (NLM) method has always been a topic of great interest in the field of image processing. There are two implementation methods for NLM. One is the pixel-wise implementation, and the other is the patch-wise implementation. They are called the pixel-wise NLM (Pixel-NLM) and the patch-wise NLM (Patch-NLM), respectively. Pixel-NLM excels at preserving edges but may introduce more artifacts. Conversely, Patch-NLM reduces artifacts but lacks sufficient preservation for image features such as edges. To address these issues, this study proposes a two-step iterative scheme which combines Pixel-NLM with Patch-NLM (CPP-NLM). The mean square error (MSE) and structural similarity (SSIM) are used to compute the weight in different steps of the proposed NLM. In the first step, MSE based Pixel-NLM produces the initial result. In the second step, Patch-NLM that considers both MSE and SSIM is used to eliminate residual noise and artifacts from the initial result. It is noted that the second step focuses on the residual image from the initial result, and the weights in this step are carefully analyzed. The whole algorithm can also be integrated into a single NLM framework based on a pre-filter with Pixel-NLM. The weight of center pixel (CW) is emphasized. The test shows that the proposed CPP-NLM achieves satisfactory subjective and objective results. For example, compared to the Pixel-NLM benchmark, the recent state-of-the-art MSE-SSIM-NLM method improves PSNR by 1.18&#xa0;dB and SSIM by 0.057. The proposed method further enhances these metrics by 36% and 32% respectively over the MSE-SSIM-NLM results.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Two-Step Image Denoising by Combining Pixel-Wise and Patch-Wise Non-local Means

  • 期刊论文

  • 发表日期：

    2025年11月16日

查看全文
  • Xiaobo Zhang

摘要

The non-local means (NLM) method has always been a topic of great interest in the field of image processing. There are two implementation methods for NLM. One is the pixel-wise implementation, and the other is the patch-wise implementation. They are called the pixel-wise NLM (Pixel-NLM) and the patch-wise NLM (Patch-NLM), respectively. Pixel-NLM excels at preserving edges but may introduce more artifacts. Conversely, Patch-NLM reduces artifacts but lacks sufficient preservation for image features such as edges. To address these issues, this study proposes a two-step iterative scheme which combines Pixel-NLM with Patch-NLM (CPP-NLM). The mean square error (MSE) and structural similarity (SSIM) are used to compute the weight in different steps of the proposed NLM. In the first step, MSE based Pixel-NLM produces the initial result. In the second step, Patch-NLM that considers both MSE and SSIM is used to eliminate residual noise and artifacts from the initial result. It is noted that the second step focuses on the residual image from the initial result, and the weights in this step are carefully analyzed. The whole algorithm can also be integrated into a single NLM framework based on a pre-filter with Pixel-NLM. The weight of center pixel (CW) is emphasized. The test shows that the proposed CPP-NLM achieves satisfactory subjective and objective results. For example, compared to the Pixel-NLM benchmark, the recent state-of-the-art MSE-SSIM-NLM method improves PSNR by 1.18 dB and SSIM by 0.057. The proposed method further enhances these metrics by 36% and 32% respectively over the MSE-SSIM-NLM results.