Scholarlink

The vertical orientation of garlic scale buds significantly impacts both yield and quality, with precise sowing direction control remaining a persistent challenge in garlic seeding machinery. To address this issue, we developed a comprehensive data acquisition protocol and a computer vision pipeline combining object detection and progressive learning to resolve this agrotechnical bottleneck. Our approach utilizes field-collected garlic scale bud images from agricultural planting conditions. The detection phase employs YOLOv8 architecture, achieving a remarkable mAP@0.5 of 0.994 on the test set. Subsequent orientation recognition utilizes a cascaded ResNet-18 framework, delivering stage-specific accuracies of 96.99% and 97.38% respectively. Through deployment on Raspberry Pi 5 using the NCNN inference framework, the system accomplishes single-image processing in 0.09 s. Field implementation on commercial garlic planting machines demonstrates 96.36% practical recognition accuracy, successfully enabling automated assessment of scale bud alignment for optimal planting. This integration marks a significant advancement in precision garlic cultivation technology.

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

Garlic Scale Bud Orientation Recognition Algorithm with YOLO and Progressive Learning Strategy

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Litan Sun,
  • Xiancheng Shen,
  • Yongshuai Shen,
  • Lexin Jiang,
  • Jun Chong,
  • Sijie Niu,
  • Guanghui Zhang,
  • Bo Zhang

摘要

The vertical orientation of garlic scale buds significantly impacts both yield and quality, with precise sowing direction control remaining a persistent challenge in garlic seeding machinery. To address this issue, we developed a comprehensive data acquisition protocol and a computer vision pipeline combining object detection and progressive learning to resolve this agrotechnical bottleneck. Our approach utilizes field-collected garlic scale bud images from agricultural planting conditions. The detection phase employs YOLOv8 architecture, achieving a remarkable mAP@0.5 of 0.994 on the test set. Subsequent orientation recognition utilizes a cascaded ResNet-18 framework, delivering stage-specific accuracies of 96.99% and 97.38% respectively. Through deployment on Raspberry Pi 5 using the NCNN inference framework, the system accomplishes single-image processing in 0.09 s. Field implementation on commercial garlic planting machines demonstrates 96.36% practical recognition accuracy, successfully enabling automated assessment of scale bud alignment for optimal planting. This integration marks a significant advancement in precision garlic cultivation technology.