Scholarlink

<p>In real-time semantic segmentation tasks, traditional dual-branch networks have problems such as pixel misalignment and information loss when fusing different resolution feature maps from different branches. To alleviate these problems, in this paper we propose an Attention-Nested Dual-Branch Network (ANDNet). This network comprises two parts: the detail branch and the semantic branch, which interact through a Semantic Correspondence Module. This module establishes the positional relationships between feature mappings from the detail branch and the semantic branch, thereby improving the pixel misalignment issue. Additionally, we construct a Contextual Nested Attention Fusion Module to integrate the image feature information extracted from the semantic and detail branches, further enhancing the model’s segmentation accuracy. We conduct experiments on the Cityscapes, CamVid and ADE20K dataset. The results demonstrate that our ANDNet achieves a satisfactory balance between segmentation accuracy and inference speed. The code and model is available at <a href="https://github.com/XingJL-1214/ANDNet/tree/master">https://github.com/XingJL-1214/ANDNet/tree/master</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

Attention-nested dual-branch network in real-time semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Jiali Xing,
  • Siming Jia,
  • Lintao Zheng,
  • Yongsheng Dong

摘要

In real-time semantic segmentation tasks, traditional dual-branch networks have problems such as pixel misalignment and information loss when fusing different resolution feature maps from different branches. To alleviate these problems, in this paper we propose an Attention-Nested Dual-Branch Network (ANDNet). This network comprises two parts: the detail branch and the semantic branch, which interact through a Semantic Correspondence Module. This module establishes the positional relationships between feature mappings from the detail branch and the semantic branch, thereby improving the pixel misalignment issue. Additionally, we construct a Contextual Nested Attention Fusion Module to integrate the image feature information extracted from the semantic and detail branches, further enhancing the model’s segmentation accuracy. We conduct experiments on the Cityscapes, CamVid and ADE20K dataset. The results demonstrate that our ANDNet achieves a satisfactory balance between segmentation accuracy and inference speed. The code and model is available at https://github.com/XingJL-1214/ANDNet/tree/master.