Scholarlink

<p>Semantic segmentation, a cornerstone of computer vision, aims to classify each pixel in an image, offering pixel-level understanding crucial for applications like medical imaging, autonomous driving, and satellite remote sensing. Despite advances, real-time semantic segmentation faces challenges in balancing accuracy and speed. This paper introduces AEAFFNet, a dual-branch real-time semantic segmentation network featuring attention-enhanced adaptive feature fusion. AEAFFNet employs a densely connected detail branch for fine spatial detail extraction and a context branch with a channel feature filtering attention module (CFFAM) and a multi-scale information extraction module (MSIEM) for enhanced semantic feature representation. An attention-enhanced adaptive feature fusion module (AEAFFM) integrates features from both branches. Compared with recent dual-branch works, AEAFFNet preserves fine spatial details via a densely connected detail branch and enhances semantic discriminability through globally adaptive attention-based feature fusion. Experiments on Cityscapes and CamVid datasets demonstrate AEAFFNet’s superiority, achieving 78.9% mIoU at 65.7 FPS and 77.3% mIoU at 128.7 FPS, respectively, showcasing its effectiveness in maintaining high segmentation performance while ensuring fast inference. Our project repository is available at <a href="https://github.com/ZeroF-ai/AEAFFNet-main">https://github.com/ZeroF-ai/AEAFFNet-main</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

AEAFFNet: enhancing real-time semantic segmentation through attention-enhanced adaptive feature fusion

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Shan Zhao,
  • Wenjing Fu,
  • Fukai Zhang,
  • Zhanqiang Huo,
  • Yingxu Qiao

摘要

Semantic segmentation, a cornerstone of computer vision, aims to classify each pixel in an image, offering pixel-level understanding crucial for applications like medical imaging, autonomous driving, and satellite remote sensing. Despite advances, real-time semantic segmentation faces challenges in balancing accuracy and speed. This paper introduces AEAFFNet, a dual-branch real-time semantic segmentation network featuring attention-enhanced adaptive feature fusion. AEAFFNet employs a densely connected detail branch for fine spatial detail extraction and a context branch with a channel feature filtering attention module (CFFAM) and a multi-scale information extraction module (MSIEM) for enhanced semantic feature representation. An attention-enhanced adaptive feature fusion module (AEAFFM) integrates features from both branches. Compared with recent dual-branch works, AEAFFNet preserves fine spatial details via a densely connected detail branch and enhances semantic discriminability through globally adaptive attention-based feature fusion. Experiments on Cityscapes and CamVid datasets demonstrate AEAFFNet’s superiority, achieving 78.9% mIoU at 65.7 FPS and 77.3% mIoU at 128.7 FPS, respectively, showcasing its effectiveness in maintaining high segmentation performance while ensuring fast inference. Our project repository is available at https://github.com/ZeroF-ai/AEAFFNet-main.