Scholarlink

As a fundamental task in computer vision, image denoising significantly improves image quality and supports downstream applications. However, existing self-supervised methods inadequately utilize structural information when processing complex images. This paper proposes a novel structure-aware adaptive masking denoising framework (SAMGW) that enhances self-supervised denoising through structured mask design and gradient-weighted learning. We introduce three complementary mask modes that adaptively select optimal combinations based on local structural features, design a gradient-consistency weighted loss function that preserves edges and textures while reducing noise, and develop an integrated strategy that dynamically fuses predictions based on regional importance. Experiments on real noisy datasets demonstrate our method outperforms existing self-supervised approaches, achieving 35.87 dB PSNR and 0.884 SSIM on SIDD validation, and 34.56 dB PSNR and 0.894 SSIM on FMDD, with significant advantages in preserving complex textures and fine structures. The proposed method offers a new research direction particularly suitable for specialized images with rich structures, such as fluorescence microscopy images.

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

SAMGW: Structure-Preserving Mask and Gradient Learning for Complex Image Denoising

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jingrui Xu,
  • Cheng-Le Qu,
  • Zheng-Yue Song

摘要

As a fundamental task in computer vision, image denoising significantly improves image quality and supports downstream applications. However, existing self-supervised methods inadequately utilize structural information when processing complex images. This paper proposes a novel structure-aware adaptive masking denoising framework (SAMGW) that enhances self-supervised denoising through structured mask design and gradient-weighted learning. We introduce three complementary mask modes that adaptively select optimal combinations based on local structural features, design a gradient-consistency weighted loss function that preserves edges and textures while reducing noise, and develop an integrated strategy that dynamically fuses predictions based on regional importance. Experiments on real noisy datasets demonstrate our method outperforms existing self-supervised approaches, achieving 35.87 dB PSNR and 0.884 SSIM on SIDD validation, and 34.56 dB PSNR and 0.894 SSIM on FMDD, with significant advantages in preserving complex textures and fine structures. The proposed method offers a new research direction particularly suitable for specialized images with rich structures, such as fluorescence microscopy images.