Scholarlink

<p>Automatic road extraction updates road networks efficiently. However, segmentation-based methods often suffer from fragmentation due to road occlusions caused by trees, buildings, and shadows. Inspired by the idea that occluded road segments can be inferred through connectivity, we propose a connectivity-aware road extraction network (CAFormer). The network employs a transformer-based architecture to capture long-range context and a connectivity-aware module (CA) using strip convolutions to extract directional connectivity information. The CA module is further enhanced by auxiliary supervision to improve connectivity learning. Additionally, we design a multi-scale feature fusion module that utilizes channel-wise cross-attention to aggregate multi-scale features, which is helpful for the inference of occluded roads. Extensive experiments on the DeepGlobe and RoadTracer datasets demonstrate that CAFormer achieves competitive performance in both segmentation accuracy and connectivity preservation. Specifically, on the DeepGlobe dataset, CAFormer improves the IoU by 2.43% and the APLS by 3.32% compared to the previous best method. On the RoadTracer dataset, it achieves an IoU improvement of 1.1% and an APLS improvement of 0.98%. The source code is available at <a href="https://github.com/1299361191/CAFormer.">https://github.com/1299361191/CAFormer.</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

CAFormer: a connectivity-aware vision transformer for road extraction from remote sensing images

  • 期刊论文

  • 发表日期：

    2025年3月6日

查看全文
  • Xite Wang,
  • Changsheng Qin,
  • Mei Bai,
  • Qian Ma,
  • Guanyu Li

摘要

Automatic road extraction updates road networks efficiently. However, segmentation-based methods often suffer from fragmentation due to road occlusions caused by trees, buildings, and shadows. Inspired by the idea that occluded road segments can be inferred through connectivity, we propose a connectivity-aware road extraction network (CAFormer). The network employs a transformer-based architecture to capture long-range context and a connectivity-aware module (CA) using strip convolutions to extract directional connectivity information. The CA module is further enhanced by auxiliary supervision to improve connectivity learning. Additionally, we design a multi-scale feature fusion module that utilizes channel-wise cross-attention to aggregate multi-scale features, which is helpful for the inference of occluded roads. Extensive experiments on the DeepGlobe and RoadTracer datasets demonstrate that CAFormer achieves competitive performance in both segmentation accuracy and connectivity preservation. Specifically, on the DeepGlobe dataset, CAFormer improves the IoU by 2.43% and the APLS by 3.32% compared to the previous best method. On the RoadTracer dataset, it achieves an IoU improvement of 1.1% and an APLS improvement of 0.98%. The source code is available at https://github.com/1299361191/CAFormer.