Scholarlink

<p>Automatic cattle breed recognition is important for optimizing livestock management. Traditional visual assessment suffers from subjectivity. In this study, the EfficientNet-B3 architecture is applied to classify cattle breeds, and the hierarchical representation of features across the depth of the network is investigated. Feature extraction was performed at three levels. In which low-level features from early convolution layers capture textures and boundaries. Mid-level features from intermediate blocks encode morphological structures related to body shape and proportions. High-level features from deep layers represent semantic characteristics specific to the breed. The last layer of the Grad-CAM image was also visualized, showing that the model focuses on body contours and coat pigmentation, ignoring background artefacts. Hierarchical analysis shows how the network transitions from simple visual patterns to complex morphological markers. Based on the results of phenotype classification, the Holstein Friesian breed achieved 86% accuracy due to high intra-breed variability in markings. The main confusion arose between the Ayrshire and Red Dane breeds due to similar color patterns.</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

Multi-Level Visual Feature Analysis for Automatic Cattle Breed Classification in Smart Livestock Systems

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Assel Mukasheva,
  • Dina Koishiyeva,
  • Saule Amanzholova,
  • Jeong Won Kang

摘要

Automatic cattle breed recognition is important for optimizing livestock management. Traditional visual assessment suffers from subjectivity. In this study, the EfficientNet-B3 architecture is applied to classify cattle breeds, and the hierarchical representation of features across the depth of the network is investigated. Feature extraction was performed at three levels. In which low-level features from early convolution layers capture textures and boundaries. Mid-level features from intermediate blocks encode morphological structures related to body shape and proportions. High-level features from deep layers represent semantic characteristics specific to the breed. The last layer of the Grad-CAM image was also visualized, showing that the model focuses on body contours and coat pigmentation, ignoring background artefacts. Hierarchical analysis shows how the network transitions from simple visual patterns to complex morphological markers. Based on the results of phenotype classification, the Holstein Friesian breed achieved 86% accuracy due to high intra-breed variability in markings. The main confusion arose between the Ayrshire and Red Dane breeds due to similar color patterns.