Scholarlink

<p>Accurate fruit detection under natural field conditions remains challenging due to small object size, occlusion, and complex background variability. While most studies focus on architectural modifications, the role of learning dynamics and training stability remains underexplored. This study proposes a learning-dynamics-driven, stability-aware optimisation framework for fig detection based on the YOLOv11l architecture. With only a modest increase in computational cost, the proposed YOLOv11l-AGRI-EMA model integrates Adam-based optimisation with an Efficient Multi-Scale Attention (EMA) mechanism to enhance feature selectivity and convergence behaviour. Experimental results demonstrate strong performance across multiple metrics, including precision (84.66%), recall (82.14%), F1-score (83.38%), and mAP@0.5–0.95 (65.11%). Multi-seed experiments confirm high reproducibility, while statistical analysis (Wilcoxon test, <i>p</i> &lt; 0.05) validates the significance of the improvements. The model also achieves reliable counting performance (R² = 0.971, MAE = 0.19) and robustness under diverse field conditions. Importantly, the findings provide empirical evidence that performance improvements are not solely dependent on architectural design. Instead, optimisation of learning dynamics emerges as a viable strategy that improves performance while introducing only limited computational overhead. These results suggest that optimisation-oriented training strategies, combined with attention mechanisms, can contribute to improved detection performance in agricultural object detection tasks.</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

A learning-dynamics-driven framework for robust small-object detection in complex field environments: YOLOv11l-AGRI-EMA framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Erhan Kahya

摘要

Accurate fruit detection under natural field conditions remains challenging due to small object size, occlusion, and complex background variability. While most studies focus on architectural modifications, the role of learning dynamics and training stability remains underexplored. This study proposes a learning-dynamics-driven, stability-aware optimisation framework for fig detection based on the YOLOv11l architecture. With only a modest increase in computational cost, the proposed YOLOv11l-AGRI-EMA model integrates Adam-based optimisation with an Efficient Multi-Scale Attention (EMA) mechanism to enhance feature selectivity and convergence behaviour. Experimental results demonstrate strong performance across multiple metrics, including precision (84.66%), recall (82.14%), F1-score (83.38%), and mAP@0.5–0.95 (65.11%). Multi-seed experiments confirm high reproducibility, while statistical analysis (Wilcoxon test, p < 0.05) validates the significance of the improvements. The model also achieves reliable counting performance (R² = 0.971, MAE = 0.19) and robustness under diverse field conditions. Importantly, the findings provide empirical evidence that performance improvements are not solely dependent on architectural design. Instead, optimisation of learning dynamics emerges as a viable strategy that improves performance while introducing only limited computational overhead. These results suggest that optimisation-oriented training strategies, combined with attention mechanisms, can contribute to improved detection performance in agricultural object detection tasks.