Scholarlink

<p>Recently, the dual-branch network has been favored by many researchers due to its remarkable effectiveness in real-time semantic segmentation tasks. However, the lightweight backbone networks currently used in dual-branch networks have some limitations. On the one hand, the receptive field of the lightweight backbone network is limited, which restricts further improvement of model performance. On the other hand, the feature loss at the object edges in images caused by stepwise downsampling operation seriously degrades the overall performance. Therefore, we propose DRFRNet, which consists of three main components: a deep dual-resolution backbone network, a parallel pyramid pooling module (PAASPP), and a feature rectification fusion module (FRF). The deep dual-resolution backbone network adopts a dual-branch structure to extract spatial and semantic features from images separately. Multiple bilateral fusions are performed between the two branches to promote feature interaction and complementarity. Before the bilateral fusion, the spatial branch is fed into a multi-scale dilated convolution block (MDCB) to expand the overall receptive field of the model. In addition, we designed the PAASPP module to expand the effective receptive field and fuse multi-scale contexts based on low-resolution feature maps. Finally, in the FRF module, before fusing the features of each branch, we first use feature resampling to rectify the features extracted by each branch and then integrate a gating mechanism to adaptively fuse them. Our method achieves a competitive balance between inference speed and accuracy on the Cityscapes, Camvid, and BDD datasets. Specifically, DRFRNet-S achieves 76.9% mIoU at 143.2 FPS on the Cityscapes dataset, 77.2% mIoU at 174.4 FPS on the Camvid dataset, and 58.5% mIoU at 145.7 FPS on the BDD dataset. The code is available at <a href="https://github.com/JAYCHUS/DRFRnet">https://github.com/JAYCHUS/DRFRnet</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

DRFRNet: a dual-resolution network with feature rectification for real-time semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Haifeng Sima,
  • Longfei Zhu,
  • Jianlong Wang,
  • Fukai Zhang,
  • Zhanqiang Huo

摘要

Recently, the dual-branch network has been favored by many researchers due to its remarkable effectiveness in real-time semantic segmentation tasks. However, the lightweight backbone networks currently used in dual-branch networks have some limitations. On the one hand, the receptive field of the lightweight backbone network is limited, which restricts further improvement of model performance. On the other hand, the feature loss at the object edges in images caused by stepwise downsampling operation seriously degrades the overall performance. Therefore, we propose DRFRNet, which consists of three main components: a deep dual-resolution backbone network, a parallel pyramid pooling module (PAASPP), and a feature rectification fusion module (FRF). The deep dual-resolution backbone network adopts a dual-branch structure to extract spatial and semantic features from images separately. Multiple bilateral fusions are performed between the two branches to promote feature interaction and complementarity. Before the bilateral fusion, the spatial branch is fed into a multi-scale dilated convolution block (MDCB) to expand the overall receptive field of the model. In addition, we designed the PAASPP module to expand the effective receptive field and fuse multi-scale contexts based on low-resolution feature maps. Finally, in the FRF module, before fusing the features of each branch, we first use feature resampling to rectify the features extracted by each branch and then integrate a gating mechanism to adaptively fuse them. Our method achieves a competitive balance between inference speed and accuracy on the Cityscapes, Camvid, and BDD datasets. Specifically, DRFRNet-S achieves 76.9% mIoU at 143.2 FPS on the Cityscapes dataset, 77.2% mIoU at 174.4 FPS on the Camvid dataset, and 58.5% mIoU at 145.7 FPS on the BDD dataset. The code is available at https://github.com/JAYCHUS/DRFRnet.