Scholarlink

<p>Real-time semantic segmentation holds broad application prospects in autonomous driving and robot navigation. Recently, real-time semantic segmentation networks primarily adopt encoder–decoder and multi-branch architectures. However, both approaches have inherent limitations. Specifically, encoder–decoder networks still fall short in utilizing spatial information. On the other hand, although multi-branch architectures better leverage spatial information, they face challenges in fully integrating information extracted from multiple branches. Additionally, both approaches suffer from a lack of information guidance, preventing the network from effectively learning spatial details and boundary information. To address these issues, we propose a dual-branch semantic segmentation network with boundary information guidance (DBBGNet). At the end of the semantic branch, we introduce an efficient aggregation pyramid pooling module (EAPPM) that extracts rich multi-scale contextual information. During the information fusion stage, we design a bilateral feature aggregation module (BFAM) to fully integrate the spatial and semantic information extracted by both branches. Furthermore, to obtain clear boundary information for constructing boundary loss, we introduce a boundary detail generation module (BDGM). Extensive experiments on Cityscapes, CamVid and iSAID datasets demonstrate competitive results. Specifically, DBBGNet achieves 78.6% mIoU at 69.2 FPS on the Cityscapes test dataset and 78.8% mIoU at 137.2 FPS on the CamVid test dataset. The source code can be accessed via: <a href="https://github.com/chenchaungya/DBBGNet">https://github.com/chenchaungya/DBBGNet</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

DBBGNet: a dual-branch network with boundary information guidance for real-time semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Chuang Chen,
  • Guohua Liu

摘要

Real-time semantic segmentation holds broad application prospects in autonomous driving and robot navigation. Recently, real-time semantic segmentation networks primarily adopt encoder–decoder and multi-branch architectures. However, both approaches have inherent limitations. Specifically, encoder–decoder networks still fall short in utilizing spatial information. On the other hand, although multi-branch architectures better leverage spatial information, they face challenges in fully integrating information extracted from multiple branches. Additionally, both approaches suffer from a lack of information guidance, preventing the network from effectively learning spatial details and boundary information. To address these issues, we propose a dual-branch semantic segmentation network with boundary information guidance (DBBGNet). At the end of the semantic branch, we introduce an efficient aggregation pyramid pooling module (EAPPM) that extracts rich multi-scale contextual information. During the information fusion stage, we design a bilateral feature aggregation module (BFAM) to fully integrate the spatial and semantic information extracted by both branches. Furthermore, to obtain clear boundary information for constructing boundary loss, we introduce a boundary detail generation module (BDGM). Extensive experiments on Cityscapes, CamVid and iSAID datasets demonstrate competitive results. Specifically, DBBGNet achieves 78.6% mIoU at 69.2 FPS on the Cityscapes test dataset and 78.8% mIoU at 137.2 FPS on the CamVid test dataset. The source code can be accessed via: https://github.com/chenchaungya/DBBGNet.