Scholarlink

<p>The precise quantification of corn seedlings is of considerable importance for evaluating planting density, yield estimation, and implementing precision agricultural management. Nevertheless, objects frequently exhibit high spatial density in field environments, are subject to occlusions, and are affected by complex background interference. Furthermore, edge-computing devices are constrained by limited computational resources and energy capacity, making it challenging for traditional models to achieve accuracy and real-time performance simultaneously. To address this, this paper proposes a lightweight corn seedling detection model, YOLO11-SPE, and constructs a drone image dataset for evaluation, based on diverse scenes and a specialized occlusion subset. The model is optimized based on YOLO11n with three key improvements: (1) StarNet is employed as the lightweight backbone, utilizing element-wise multiplication to enhance nonlinear channel interactions, thereby improving the separation of occlusions and edge adhesion; (2) C3k2_PC is introduced at the network neck, incorporating pooled token mixing and gated channel modeling to strengthen multi-scale and contextual representations, while suppressing false targets; (3) an EfficientHead module is designed, combining decoupled prediction, distributed focal loss (DFL), and anchor-free decoding strategies to enhance the localization stability of dense small targets and occluded areas. Experimental results show that YOLO11-SPE maintains a relatively stable mAP@0.5 while reducing the model’s parameters, FLOPs, and size by approximately 43%, 46%, and 42%, respectively, with an inference speed of 98.9 FPS. In the counting regression task, the MAE and RMSE are 4.51 and 6.13, respectively, with an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textrm{R}^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>R</mtext> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> value of 0.9807. This study also conducted edge-computing device deployment testing on the Jetson Nano, achieving 24.8 FPS, which is sufficient to meet real-time detection requirements.</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

YOLO11-SPE: a lightweight object detection model for corn seedling counting

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Yize Wu,
  • Yubing Luo,
  • Hongpu Chen,
  • Feng Chen,
  • Hao Ye,
  • Xin Chen,
  • Xiang Li

摘要

The precise quantification of corn seedlings is of considerable importance for evaluating planting density, yield estimation, and implementing precision agricultural management. Nevertheless, objects frequently exhibit high spatial density in field environments, are subject to occlusions, and are affected by complex background interference. Furthermore, edge-computing devices are constrained by limited computational resources and energy capacity, making it challenging for traditional models to achieve accuracy and real-time performance simultaneously. To address this, this paper proposes a lightweight corn seedling detection model, YOLO11-SPE, and constructs a drone image dataset for evaluation, based on diverse scenes and a specialized occlusion subset. The model is optimized based on YOLO11n with three key improvements: (1) StarNet is employed as the lightweight backbone, utilizing element-wise multiplication to enhance nonlinear channel interactions, thereby improving the separation of occlusions and edge adhesion; (2) C3k2_PC is introduced at the network neck, incorporating pooled token mixing and gated channel modeling to strengthen multi-scale and contextual representations, while suppressing false targets; (3) an EfficientHead module is designed, combining decoupled prediction, distributed focal loss (DFL), and anchor-free decoding strategies to enhance the localization stability of dense small targets and occluded areas. Experimental results show that YOLO11-SPE maintains a relatively stable mAP@0.5 while reducing the model’s parameters, FLOPs, and size by approximately 43%, 46%, and 42%, respectively, with an inference speed of 98.9 FPS. In the counting regression task, the MAE and RMSE are 4.51 and 6.13, respectively, with an \(\textrm{R}^{2}\) R 2 value of 0.9807. This study also conducted edge-computing device deployment testing on the Jetson Nano, achieving 24.8 FPS, which is sufficient to meet real-time detection requirements.