Scholarlink

<p>The complex and variable nature of shadows in natural environments-including diverse shapes, intensities, and intricate scene compositions-presents major challenges for accurate shadow removal. To address these challenges, this study introduces MGME-ShadowNet, a novel shadow removal algorithm designed to enhance image quality by restoring details within shadowed regions, thereby improving the performance of subsequent image processing tasks. MGME-ShadowNet integrates a MaskGuideAttention mechanism and a Mask-Aware Enhancement Module (MAEM) within a diffusion model framework, enabling it to handle both soft and hard shadows, especially in complex scenes. The MaskGuideAttention mechanism adaptively adjusts the contributions of self-attention and spatial attention through learnable parameters and allocates attention based on shadow masks. Meanwhile, the MAEM module refines initial shadow masks generated by the shadow detection network, improving their accuracy through dilated convolutions and pyramid pooling. Experimental evaluations on the SRD dataset demonstrate that MGME-ShadowNet achieves improved performance compared to existing methods, with a 2.09 dB increase in peak signal-to-noise ratio (PSNR) in shadowed regions and a 0.56-point reduction in root mean square error (RMSE) for shadow quality, surpassing the performance of ShadowDiffusion. These results confirm MGME-ShadowNet’s effectiveness in handling complex shadow scenarios. The code is available at https://github.com/Donghui-Wang/MGMEShadowNet.</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

Image shadow removal algorithm based on MaskGuideAttention and mask-aware enhancement module

  • 期刊论文

  • 发表日期：

    2025年5月2日

查看全文
  • Donghui Wang,
  • Jinhua Wang,
  • Ning He,
  • Jingzun Zhang,
  • Sen Zhang,
  • Shuai Liu

摘要

The complex and variable nature of shadows in natural environments-including diverse shapes, intensities, and intricate scene compositions-presents major challenges for accurate shadow removal. To address these challenges, this study introduces MGME-ShadowNet, a novel shadow removal algorithm designed to enhance image quality by restoring details within shadowed regions, thereby improving the performance of subsequent image processing tasks. MGME-ShadowNet integrates a MaskGuideAttention mechanism and a Mask-Aware Enhancement Module (MAEM) within a diffusion model framework, enabling it to handle both soft and hard shadows, especially in complex scenes. The MaskGuideAttention mechanism adaptively adjusts the contributions of self-attention and spatial attention through learnable parameters and allocates attention based on shadow masks. Meanwhile, the MAEM module refines initial shadow masks generated by the shadow detection network, improving their accuracy through dilated convolutions and pyramid pooling. Experimental evaluations on the SRD dataset demonstrate that MGME-ShadowNet achieves improved performance compared to existing methods, with a 2.09 dB increase in peak signal-to-noise ratio (PSNR) in shadowed regions and a 0.56-point reduction in root mean square error (RMSE) for shadow quality, surpassing the performance of ShadowDiffusion. These results confirm MGME-ShadowNet’s effectiveness in handling complex shadow scenarios. The code is available at https://github.com/Donghui-Wang/MGMEShadowNet.