Scholarlink

<p>In recent years, deep learning models based on non-local attention (NLA) have achieved remarkable success in image denoising. The success of NLA stems from its ability to capture long-range interactions guided by both the data itself and learnable parameters within the attention mechanism. Building on the classical insight from traditional non-local filtering methods, it has been observed that more accurate calculation of interactions or similarities between tokens can significantly improve denoising performance. However, existing non-local attention networks designed for denoising have failed to consider the trustworthiness of these interaction computations. In this paper, motivated by the statistic principles underlying similarity calculation, we propose a exacter set-level non-local attention based on the learning weights (ESNLA). This approach leverages similar tokens within the local neighborhood to mitigate the impact of noise and enhance the robustness of similarity calculations under noisy conditions. Designed as an efficient modular component, ESNLA can be seamlessly integrated into any deep convolutional architecture for image denoising. Ablation studies validate the superiority of ESNLA over existing NLA modules. By embedding multiple ESNLA blocks into a ResNet backbone, we further construct a deep ESNLA Network (ESNLANet). Extensive experiments demonstrate that ESNLANet achieves highly competitive denoising performance, outperforming a wide range of state-of-the-art methods.</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

Exacter set-level nonlocal attention mechanism based on learning weights for image denoising

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Dai-Qiang Chen,
  • Ying-Chun Chen

摘要

In recent years, deep learning models based on non-local attention (NLA) have achieved remarkable success in image denoising. The success of NLA stems from its ability to capture long-range interactions guided by both the data itself and learnable parameters within the attention mechanism. Building on the classical insight from traditional non-local filtering methods, it has been observed that more accurate calculation of interactions or similarities between tokens can significantly improve denoising performance. However, existing non-local attention networks designed for denoising have failed to consider the trustworthiness of these interaction computations. In this paper, motivated by the statistic principles underlying similarity calculation, we propose a exacter set-level non-local attention based on the learning weights (ESNLA). This approach leverages similar tokens within the local neighborhood to mitigate the impact of noise and enhance the robustness of similarity calculations under noisy conditions. Designed as an efficient modular component, ESNLA can be seamlessly integrated into any deep convolutional architecture for image denoising. Ablation studies validate the superiority of ESNLA over existing NLA modules. By embedding multiple ESNLA blocks into a ResNet backbone, we further construct a deep ESNLA Network (ESNLANet). Extensive experiments demonstrate that ESNLANet achieves highly competitive denoising performance, outperforming a wide range of state-of-the-art methods.