Scholarlink

To address the challenges of complex background interference and insufficient multi-scale morphological adaptability in tomato seedling detection under greenhouse conditions, an improved OMB-YOLO model is proposed based on the YOLOv11 framework. The model reconstructs the backbone network with full-dimensional dynamic convolution (ODConv) to enhance the feature representation of stem and leaf textures. A feature enhancement module based on Multi-Scale Dilated Attention (MSDA) is introduced to suppress glare noise interference, while a Bi-directional Weighted Feature Pyramid Network (Bi-FPN)-based neck structure optimizes dense object localization, achieving synergistic improvements in detection performance. Experimental results show that the improved model achieves an mAP of 98.0% on a custom dataset, a 1.6% improvement over the original model. With a computational cost of 6.8 GFLOPs, the model improves mAP by 1.9–4.8% compared to the YOLOv5–10 series, with Precision and Recall reaching 94.8% and 93.4%, respectively. Through dynamic architecture optimization, this model provides an efficient solution for greenhouse tomato seedling detection.

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

Greenhouse Tomato Seedling Detection Algorithm Based on OMB-YOLO

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Boyang Li,
  • Daming Liu

摘要

To address the challenges of complex background interference and insufficient multi-scale morphological adaptability in tomato seedling detection under greenhouse conditions, an improved OMB-YOLO model is proposed based on the YOLOv11 framework. The model reconstructs the backbone network with full-dimensional dynamic convolution (ODConv) to enhance the feature representation of stem and leaf textures. A feature enhancement module based on Multi-Scale Dilated Attention (MSDA) is introduced to suppress glare noise interference, while a Bi-directional Weighted Feature Pyramid Network (Bi-FPN)-based neck structure optimizes dense object localization, achieving synergistic improvements in detection performance. Experimental results show that the improved model achieves an mAP of 98.0% on a custom dataset, a 1.6% improvement over the original model. With a computational cost of 6.8 GFLOPs, the model improves mAP by 1.9–4.8% compared to the YOLOv5–10 series, with Precision and Recall reaching 94.8% and 93.4%, respectively. Through dynamic architecture optimization, this model provides an efficient solution for greenhouse tomato seedling detection.