Scholarlink

<p>As the strawberry industry expands, the inefficiency and cost of manual harvesting have spurred the development of vision-based automated picking systems. Among them, ripeness detection models are crucial for enabling precise and efficient harvesting. However, existing models often struggle in complex greenhouse environments, particularly under occlusion and small-object conditions, leading to reduced accuracy and excessive model size. To address these challenges, this study proposes YOLO11-LES, a novel lightweight strawberry ripeness detection model based on YOLO11n. It integrates three newly designed modules: (1) the Lightweight Adaptive Weighting Downsampling Structure (LAWDS) module replaces standard downsampling convolutions with adaptive weighting to reduce parameters and computation; (2) the Edge-Intensified Feature Extraction Stem (EIEStem) module enhances edge features and mitigates small-object detail loss during early feature extraction; and (3) the Spatial-Enhanced Attention Module Head (SEAMHead) module improves occlusion awareness through integrated spatial and channel attention. Experiments show that YOLO11-LES achieves a compact model size of 4.6 MB, with a precision of 81.5%, recall of 84.1%, and mAP50 of 86.5%. Compared to the baseline YOLO11n, it improves these metrics by 2.9%, 9.0%, and 3.2%, respectively, while reducing the model size by 0.9 MB. YOLO11-LES effectively balances detection accuracy with lightweight design, making it well suited for deployment on real-time edge devices.</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

Beyond obstacles: feather-light YOLO11-LES for real-time ripeness detection of occluded strawberries in greenhouses

  • 期刊论文

  • 发表日期：

    2025年8月23日

查看全文
  • Zheng Li,
  • Xiaonan Hu,
  • Xiaobei Zhao,
  • Hao Ye,
  • Feng Chen,
  • Xin Chen,
  • Xiang Li

摘要

As the strawberry industry expands, the inefficiency and cost of manual harvesting have spurred the development of vision-based automated picking systems. Among them, ripeness detection models are crucial for enabling precise and efficient harvesting. However, existing models often struggle in complex greenhouse environments, particularly under occlusion and small-object conditions, leading to reduced accuracy and excessive model size. To address these challenges, this study proposes YOLO11-LES, a novel lightweight strawberry ripeness detection model based on YOLO11n. It integrates three newly designed modules: (1) the Lightweight Adaptive Weighting Downsampling Structure (LAWDS) module replaces standard downsampling convolutions with adaptive weighting to reduce parameters and computation; (2) the Edge-Intensified Feature Extraction Stem (EIEStem) module enhances edge features and mitigates small-object detail loss during early feature extraction; and (3) the Spatial-Enhanced Attention Module Head (SEAMHead) module improves occlusion awareness through integrated spatial and channel attention. Experiments show that YOLO11-LES achieves a compact model size of 4.6 MB, with a precision of 81.5%, recall of 84.1%, and mAP50 of 86.5%. Compared to the baseline YOLO11n, it improves these metrics by 2.9%, 9.0%, and 3.2%, respectively, while reducing the model size by 0.9 MB. YOLO11-LES effectively balances detection accuracy with lightweight design, making it well suited for deployment on real-time edge devices.