Scholarlink

<p>Semantic segmentation is pivotal in computer vision, with applications in medical imaging, geospatial analysis, and intelligent transportation. This paper introduces SGTNet, a real-time semantic segmentation network that integrates sparse Transformer features with multi-level local features. The proposed network comprises three key modules: the Sparse Transformer Module (STM) for efficient global context modeling, the Semantic Information Enhancement Module (SIEM) for fusing global and multi-scale local features, and the Attention-Guided Fusion Module (AGFM) for enhancing global context representation through attention mechanisms. Evaluated on four benchmark datasets, SGTNet achieves 77.9% mIoU on Cityscapes, 77.4% on CamVid, 53.4% on PASCAL VOC2012, and 36.6% on ADE20K, demonstrating superior performance in capturing global context and integrating multi-scale feature information. The code of this work is publicly available at: <a href="https://github.com/zhoukaiyu-zky/SGTNet">https://github.com/zhoukaiyu-zky/SGTNet</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

SGTNet: real-time semantic segmentation via sparse transformer integration and multi-scale feature fusion

  • 期刊论文

  • 发表日期：

    2026年6月14日

查看全文
  • Shan Zhao,
  • Kaiyu Zhou,
  • Fukai Zhang,
  • Zhanqiang Huo,
  • Yingxu Qiao

摘要

Semantic segmentation is pivotal in computer vision, with applications in medical imaging, geospatial analysis, and intelligent transportation. This paper introduces SGTNet, a real-time semantic segmentation network that integrates sparse Transformer features with multi-level local features. The proposed network comprises three key modules: the Sparse Transformer Module (STM) for efficient global context modeling, the Semantic Information Enhancement Module (SIEM) for fusing global and multi-scale local features, and the Attention-Guided Fusion Module (AGFM) for enhancing global context representation through attention mechanisms. Evaluated on four benchmark datasets, SGTNet achieves 77.9% mIoU on Cityscapes, 77.4% on CamVid, 53.4% on PASCAL VOC2012, and 36.6% on ADE20K, demonstrating superior performance in capturing global context and integrating multi-scale feature information. The code of this work is publicly available at: https://github.com/zhoukaiyu-zky/SGTNet.