Scholarlink

<p>Visual navigation, grounded in RGB imagery, offers a cost-effective and efficient solution for autonomous navigation in combine harvesters. Nevertheless, the precision of existing visual methods is compromised by factors such as stubble interference, variations in soil texture, and lighting fluctuations. Addressing these challenges, this study introduces GloCrop-UNet, a network designed for wheat harvesting scenes that enhances global semantic and local detailed features collaboratively through a dual-path architecture. The primary path employs the bidirectional feature shift-deep mechanism (BFSM) to establish longitudinal and transverse long-distance dependencies alongside cross-channel feature interactions, while the auxiliary path uses the hierarchical dynamic regional attention mechanism (HDRAM) to emphasize critical edge textures and reduce computational complexity. Furthermore, the dual-channel shared encoder facilitates the efficient reuse of low-level features. Experimental results demonstrate that GloCrop-UNet achieves a segmentation accuracy of 82.47% and a Dice coefficient of 90.25% in wheat edge detection tasks, surpassing the standard Rolling-UNet by 3.45% in accuracy and 2.02% in Dice coefficient, with a reduced parameter count of 2.0073 million. This model is particularly suited for deployment in embedded hardware within agricultural machinery, offering a promising advancement in autonomous agricultural navigation. The code of GloCrop-UNet has been opened on GitHub (<a href="https://github.com/shandian9876/GloCrop-UNet">https://github.com/shandian9876/GloCrop-UNet</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

Enhanced wheat edge detection via dual-channel global–local feature collaboration in GloCrop-UNet

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Chunyu Xu,
  • Mengyang Xu,
  • Zhihao Yan,
  • Jinhai Li,
  • Bingxing Mao,
  • Yufeng Wang

摘要

Visual navigation, grounded in RGB imagery, offers a cost-effective and efficient solution for autonomous navigation in combine harvesters. Nevertheless, the precision of existing visual methods is compromised by factors such as stubble interference, variations in soil texture, and lighting fluctuations. Addressing these challenges, this study introduces GloCrop-UNet, a network designed for wheat harvesting scenes that enhances global semantic and local detailed features collaboratively through a dual-path architecture. The primary path employs the bidirectional feature shift-deep mechanism (BFSM) to establish longitudinal and transverse long-distance dependencies alongside cross-channel feature interactions, while the auxiliary path uses the hierarchical dynamic regional attention mechanism (HDRAM) to emphasize critical edge textures and reduce computational complexity. Furthermore, the dual-channel shared encoder facilitates the efficient reuse of low-level features. Experimental results demonstrate that GloCrop-UNet achieves a segmentation accuracy of 82.47% and a Dice coefficient of 90.25% in wheat edge detection tasks, surpassing the standard Rolling-UNet by 3.45% in accuracy and 2.02% in Dice coefficient, with a reduced parameter count of 2.0073 million. This model is particularly suited for deployment in embedded hardware within agricultural machinery, offering a promising advancement in autonomous agricultural navigation. The code of GloCrop-UNet has been opened on GitHub (https://github.com/shandian9876/GloCrop-UNet).