Scholarlink

<p>Image denoising is a fundamental task in image processing and computer vision. Traditional supervised methods heavily rely on large sets of noisy-clean image pairs, which are impractical to obtain for real-world noisy images. Self-supervised denoising methods offer a viable alternative but often struggle with spatial correlation in noise. In this paper, we introduce the asymmetric mask blind-spot network (AM-BSN), designed to disrupt spatial correlations of large-scale noise in real-world images. Our network features a dual-branch architecture: a local branch employing a 3<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>3 central mask convolution for fine detail recovery, and a global branch utilizing a <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(5\times 5\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>5</mn> <mo>×</mo> <mn>5</mn> </mrow> </math></EquationSource> </InlineEquation> ‘X’-shaped mask convolution and dilated convolutions for global structure reconstruction. Experimental results on real-world datasets demonstrate that AM-BSN outperforms recent state-of-the-art self-supervised denoising methods (e.g., AP-BSN, C-BSN, MMBSN, LG-BPN and SASL), achieving a PSNR of 37.90&#xa0;dB and an SSIM of 0.885 on the SIDD benchmark dataset, and a PSNR of 38.49&#xa0;dB and an SSIM of 0.940 on the DND benchmark dataset. This research advances self-supervised denoising techniques, providing a practical solution for real-world applications. The code is available at <a href="https://github.com/wam730/AM-BSN.">https://github.com/wam730/AM-BSN.</a></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

Enhancing self-supervised image denoising with asymmetric mask blind-spot network

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Yujie Wang,
  • Weiwei Wang,
  • Yi Wu,
  • Kaige Cui

摘要

Image denoising is a fundamental task in image processing and computer vision. Traditional supervised methods heavily rely on large sets of noisy-clean image pairs, which are impractical to obtain for real-world noisy images. Self-supervised denoising methods offer a viable alternative but often struggle with spatial correlation in noise. In this paper, we introduce the asymmetric mask blind-spot network (AM-BSN), designed to disrupt spatial correlations of large-scale noise in real-world images. Our network features a dual-branch architecture: a local branch employing a 3 \(\times \) × 3 central mask convolution for fine detail recovery, and a global branch utilizing a \(5\times 5\) 5 × 5 ‘X’-shaped mask convolution and dilated convolutions for global structure reconstruction. Experimental results on real-world datasets demonstrate that AM-BSN outperforms recent state-of-the-art self-supervised denoising methods (e.g., AP-BSN, C-BSN, MMBSN, LG-BPN and SASL), achieving a PSNR of 37.90 dB and an SSIM of 0.885 on the SIDD benchmark dataset, and a PSNR of 38.49 dB and an SSIM of 0.940 on the DND benchmark dataset. This research advances self-supervised denoising techniques, providing a practical solution for real-world applications. The code is available at https://github.com/wam730/AM-BSN.