Scholarlink

<p>Semantic segmentation of unmanned aerial vehicle (UAV) nighttime oblique images is critical for UAV navigation and nighttime urban interior research. However, complex lighting conditions at night often obscure the details of the building facade in single-exposure images, causing previous segmentation methods to suffer from blurred edges and detail loss. To address boundary blurring and information loss in the segmentation of nighttime building facade, we propose a three-branch improved DeeplabV3 network (TE-DeeplabV3) that incorporates high/low exposure data. In the high- and low-exposure branches, an image luminance-guided exposure attention module (LGEA) is added to prioritize feature-rich regions in the corresponding exposure images. A cross-exposure attention module (CEA) is integrated into the medium-exposure branch to enable comprehensive feature extraction. Experiments on a custom UAV nighttime dataset show that TE-DeeplabV3 achieves a mIoU of 93.6% (1.8% improvement over DeeplabV3) and a building facade category IoU of 91.8% (2.5% improvement over DeeplabV3), providing a more accurate boundary segmentation for building facades.</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

TE-DeeplabV3: Semantic Segmentation of Nighttime Building Facades Taking into Account High and Low Exposure

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Chengzhi Zong,
  • Liang Zhou,
  • Deping Li,
  • Daoquan Zhang,
  • Jiejie Wu,
  • Tingting Jiang,
  • Chen Huang,
  • Penggui Xie

摘要

Semantic segmentation of unmanned aerial vehicle (UAV) nighttime oblique images is critical for UAV navigation and nighttime urban interior research. However, complex lighting conditions at night often obscure the details of the building facade in single-exposure images, causing previous segmentation methods to suffer from blurred edges and detail loss. To address boundary blurring and information loss in the segmentation of nighttime building facade, we propose a three-branch improved DeeplabV3 network (TE-DeeplabV3) that incorporates high/low exposure data. In the high- and low-exposure branches, an image luminance-guided exposure attention module (LGEA) is added to prioritize feature-rich regions in the corresponding exposure images. A cross-exposure attention module (CEA) is integrated into the medium-exposure branch to enable comprehensive feature extraction. Experiments on a custom UAV nighttime dataset show that TE-DeeplabV3 achieves a mIoU of 93.6% (1.8% improvement over DeeplabV3) and a building facade category IoU of 91.8% (2.5% improvement over DeeplabV3), providing a more accurate boundary segmentation for building facades.