Scholarlink

<p>With the increasing demand for intelligent development in animal husbandry, traditional manual creation of cattle face datasets faces problems such as low efficiency and large errors. This paper proposes an automatic cattle face image capture and cropping algorithm based on improved YOLOv3 in small-sample scenarios, aiming to achieve automated construction of cattle face datasets. Due to the low-latency requirements for subsequent transplantation into embedded systems, the lightweight MobileNet is first used as the feature extraction network to improve detection speed, and the K-means +  + algorithm is adopted to re-cluster anchor boxes,&#xa0;thereby&#xa0;optimizing target scale adaptability. Meanwhile, data augmentation and transfer learning strategies are introduced to address overfitting issues caused by small-sample training. Ablation experiments based on a dataset of 7286 images from 48 cattle have shown that the improved model achieves an accuracy of 99.86% and a detection time of only 39.24&#xa0;min, with a 9% increase in accuracy and a 50% reduction in detection time compared to the original YOLOv3 model, and a detection efficiency 27 times higher than&#xa0;that of manual cropping. This algorithm provides an efficient data acquisition solution for rapid modeling of cattle face recognition systems and can be migrated to intelligent management scenarios of other livestock species.</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

Automatic capture and interception algorithm for cattle face images based on improved YOLOv3 in small-sample scenarios

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月31日

查看全文
  • Zheng Li,
  • Juan Yin

摘要

With the increasing demand for intelligent development in animal husbandry, traditional manual creation of cattle face datasets faces problems such as low efficiency and large errors. This paper proposes an automatic cattle face image capture and cropping algorithm based on improved YOLOv3 in small-sample scenarios, aiming to achieve automated construction of cattle face datasets. Due to the low-latency requirements for subsequent transplantation into embedded systems, the lightweight MobileNet is first used as the feature extraction network to improve detection speed, and the K-means +  + algorithm is adopted to re-cluster anchor boxes, thereby optimizing target scale adaptability. Meanwhile, data augmentation and transfer learning strategies are introduced to address overfitting issues caused by small-sample training. Ablation experiments based on a dataset of 7286 images from 48 cattle have shown that the improved model achieves an accuracy of 99.86% and a detection time of only 39.24 min, with a 9% increase in accuracy and a 50% reduction in detection time compared to the original YOLOv3 model, and a detection efficiency 27 times higher than that of manual cropping. This algorithm provides an efficient data acquisition solution for rapid modeling of cattle face recognition systems and can be migrated to intelligent management scenarios of other livestock species.