Scholarlink

<p>Semantic segmentation-based Complex Waterway Scene Understanding exhibits significant promise in the environmental perception of Unmanned Surface Vehicles (USVs). Existing methods suffer from poor edge estimation of obstacles under blurred water surface conditions and high false positive rates in extreme water conditions. To address these issues, we propose a novel Lightweight Dual-branch Attention Network—LDANet. To enable the model to better analyze scenes from different perspectives, we have refined the dual-branch network structure by incorporating multiple atrous branches for local fusion. Furthermore, to better integrate highly diverse feature information, we introduce the Difference-feature Attention Fusion (DAF) method. This approach utilizes spatial dimension information reorganization to achieve mixed-domain attention calculations. Finally, we employ a clever connection scheme that combines DAF with Parallel Aggregation Pyramid Pooling Module (PAPPM) multi-scale processing, adaptively enhancing both local and global information. Our method achieves an mIoU of 96.2% at 81FPS on the MaSTr1325 dataset, 95.6% mIoU on the LaRS dataset, and 99.13% mIoU on Water Segmentation in the USV Inland. Notably, it excels in handling challenging waterway images with complex lighting conditions, fluctuations, and inhomogeneous reflections. Experimental results demonstrate that LDANet not only provides an effective tool for Complex Waterway Scene Understanding but also serves as a reference for semantic segmentation tasks in other complex environments.</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

LDANet: enhancing USV's capacity for better segmentation of complex waterway scenes

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Tongyang Dai,
  • Huiyu Xiang,
  • Chongjie Leng,
  • Song Huang,
  • Guanghui He,
  • Shishuo Han

摘要

Semantic segmentation-based Complex Waterway Scene Understanding exhibits significant promise in the environmental perception of Unmanned Surface Vehicles (USVs). Existing methods suffer from poor edge estimation of obstacles under blurred water surface conditions and high false positive rates in extreme water conditions. To address these issues, we propose a novel Lightweight Dual-branch Attention Network—LDANet. To enable the model to better analyze scenes from different perspectives, we have refined the dual-branch network structure by incorporating multiple atrous branches for local fusion. Furthermore, to better integrate highly diverse feature information, we introduce the Difference-feature Attention Fusion (DAF) method. This approach utilizes spatial dimension information reorganization to achieve mixed-domain attention calculations. Finally, we employ a clever connection scheme that combines DAF with Parallel Aggregation Pyramid Pooling Module (PAPPM) multi-scale processing, adaptively enhancing both local and global information. Our method achieves an mIoU of 96.2% at 81FPS on the MaSTr1325 dataset, 95.6% mIoU on the LaRS dataset, and 99.13% mIoU on Water Segmentation in the USV Inland. Notably, it excels in handling challenging waterway images with complex lighting conditions, fluctuations, and inhomogeneous reflections. Experimental results demonstrate that LDANet not only provides an effective tool for Complex Waterway Scene Understanding but also serves as a reference for semantic segmentation tasks in other complex environments.