Scholarlink

<p>In cotton planting scenarios, the accurate identification of terminal buds faces three main challenges: mutual occlusion due to high-density plants, complex backgrounds, and limited computing power of edge devices. To address these challenges, this study proposes a lightweight object detection model, YOLOv8-CTB, which improves both detection accuracy and performance through three key optimizations: (1) The ExtraDW module is introduced to optimize the C2f module, reducing parameters and improving feature extraction for better accuracy and efficiency; (2) A lightweight cross-scale feature fusion module (BF-CCFM) replaces the Neck of YOLOv8n, enhancing multi-scale feature fusion and detection robustness; (3) The SEAM attention mechanism adaptively weights key features in occluded regions to suppress background interference and reduce false detections. Experimental results on a self-collected dataset show that YOLOv8-CTB achieves 98.3% Precision, a 2.3% improvement over YOLOv8n. Its FLOPs, parameter count, and model size are reduced by 74.1%, 75.9%, and 47.1%, significantly improving hardware adaptability. Deployment on the Jetson Xavier NX with TensorRT acceleration maintains 91.2% accuracy and 31 FPS, satisfying real-time detection requirements in field applications. This work provides strong support for the engineering application of intelligent cotton topping.</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

YOLOv8-CTB: A lightweight and real-time cotton bud detection under field conditions with edge deployment

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Maochang Song,
  • Yubin Lan,
  • Xin Han,
  • Xinyue Gou,
  • Fengchao Wang,
  • Wenshuai Yang,
  • Lihua Cui

摘要

In cotton planting scenarios, the accurate identification of terminal buds faces three main challenges: mutual occlusion due to high-density plants, complex backgrounds, and limited computing power of edge devices. To address these challenges, this study proposes a lightweight object detection model, YOLOv8-CTB, which improves both detection accuracy and performance through three key optimizations: (1) The ExtraDW module is introduced to optimize the C2f module, reducing parameters and improving feature extraction for better accuracy and efficiency; (2) A lightweight cross-scale feature fusion module (BF-CCFM) replaces the Neck of YOLOv8n, enhancing multi-scale feature fusion and detection robustness; (3) The SEAM attention mechanism adaptively weights key features in occluded regions to suppress background interference and reduce false detections. Experimental results on a self-collected dataset show that YOLOv8-CTB achieves 98.3% Precision, a 2.3% improvement over YOLOv8n. Its FLOPs, parameter count, and model size are reduced by 74.1%, 75.9%, and 47.1%, significantly improving hardware adaptability. Deployment on the Jetson Xavier NX with TensorRT acceleration maintains 91.2% accuracy and 31 FPS, satisfying real-time detection requirements in field applications. This work provides strong support for the engineering application of intelligent cotton topping.