Scholarlink

<p>Tea geometrid is a highly prevalent and damaging tea garden pest, poses significant detection challenges due to its yellow-brown color resembling soil, dead leaves, and branches. In addition, the labor and material costs involved in collecting and labeling large numbers of tea geometrid samples are high. This study introduces a convolutional neural network TGDNet for few-shot, multi-scale tea geometrid detection in complex scenarios. TGDNet incorporates a Multi-scale Rectangular Focus (MSRF) module, tailored to the pest shape using strip convolutions for efficient long-range dependency capture, and a Frequency Domain Edge Enhancement (FDEE) module, enhancing contrast by amplifying high and low frequency features. A multi-scale aggregation (MSA) detector head is designed to make full use of the limited tea geometrid features and achieves high detection accuracy on small sample datasets. Experiments demonstrate TGDNet’s superiority over state-of-the-art models, achieving mAP@0.5 of 91.6% on region focus data, 92% on single point focus data, and 91% on the combined dataset. These values exceed the baseline YOLOv11n model by 5.9, 4.5, and 2.9% respectively, while reducing model size to half at only 3&#xa0;MB.</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

TGDNet: an efficient detection model for multi-scale tea geometrids in complex scenes

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Yongcheng Jiang,
  • Zijing Wei,
  • Gensheng Hu

摘要

Tea geometrid is a highly prevalent and damaging tea garden pest, poses significant detection challenges due to its yellow-brown color resembling soil, dead leaves, and branches. In addition, the labor and material costs involved in collecting and labeling large numbers of tea geometrid samples are high. This study introduces a convolutional neural network TGDNet for few-shot, multi-scale tea geometrid detection in complex scenarios. TGDNet incorporates a Multi-scale Rectangular Focus (MSRF) module, tailored to the pest shape using strip convolutions for efficient long-range dependency capture, and a Frequency Domain Edge Enhancement (FDEE) module, enhancing contrast by amplifying high and low frequency features. A multi-scale aggregation (MSA) detector head is designed to make full use of the limited tea geometrid features and achieves high detection accuracy on small sample datasets. Experiments demonstrate TGDNet’s superiority over state-of-the-art models, achieving mAP@0.5 of 91.6% on region focus data, 92% on single point focus data, and 91% on the combined dataset. These values exceed the baseline YOLOv11n model by 5.9, 4.5, and 2.9% respectively, while reducing model size to half at only 3 MB.