Scholarlink

<p>Semantic segmentation plays a crucial role in various computer vision applications. However, improving the accuracy of existing models without significantly increasing computational cost remains a challenging task. Furthermore, current models often fall short in effectively utilizing spatial and contextual information. In this paper, we propose SCFI-ESeg, a competitive and efficient segmentation framework based on Transformer, which strengthens spatial information and aggregates multi-stage contextual information. We introduce a compact Multi-Stage Attention module and a Continuous Atrous Spatial Pyramid Pooling to enhance the integration of features at different levels and improve the model’s understanding of multi-scale information. Additionally, we present a Spatial Feature Enhancement Module, which leverages query features from the encoder to specifically enhance spatial information, thus improving the model’s ability to perceive fine image details. Experimental results demonstrate that SCFI-ESeg performs exceptionally well on public datasets, ADE20K and Pascal Context and Cityscapes, especially in complex scenarios. Compared to the baseline Segmenter, our framework achieves an average performance improvement of 1.6% on the ADE20K dataset, with a 2.5% improvement under the ViT-tiny configuration. Notably, this gain is obtained with a small absolute increase of 1.7M parameters and 0.6G GFLOPs. While the relative parameter growth is more noticeable for lightweight models such as ViT-tiny, the overall computational overhead remains modest, and the trade-off between accuracy and efficiency is favorable. Source code and models are available at Github: <a href="https://github.com/Eadeath/SCFI-ESeg">https://github.com/Eadeath/SCFI-ESeg</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

SCFI-ESeg: spatial and content feature integration for efficient semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Ning Li,
  • Xudong Zhang,
  • Gaochao Yang,
  • Bo Li,
  • Baohua Yuan

摘要

Semantic segmentation plays a crucial role in various computer vision applications. However, improving the accuracy of existing models without significantly increasing computational cost remains a challenging task. Furthermore, current models often fall short in effectively utilizing spatial and contextual information. In this paper, we propose SCFI-ESeg, a competitive and efficient segmentation framework based on Transformer, which strengthens spatial information and aggregates multi-stage contextual information. We introduce a compact Multi-Stage Attention module and a Continuous Atrous Spatial Pyramid Pooling to enhance the integration of features at different levels and improve the model’s understanding of multi-scale information. Additionally, we present a Spatial Feature Enhancement Module, which leverages query features from the encoder to specifically enhance spatial information, thus improving the model’s ability to perceive fine image details. Experimental results demonstrate that SCFI-ESeg performs exceptionally well on public datasets, ADE20K and Pascal Context and Cityscapes, especially in complex scenarios. Compared to the baseline Segmenter, our framework achieves an average performance improvement of 1.6% on the ADE20K dataset, with a 2.5% improvement under the ViT-tiny configuration. Notably, this gain is obtained with a small absolute increase of 1.7M parameters and 0.6G GFLOPs. While the relative parameter growth is more noticeable for lightweight models such as ViT-tiny, the overall computational overhead remains modest, and the trade-off between accuracy and efficiency is favorable. Source code and models are available at Github: https://github.com/Eadeath/SCFI-ESeg