Scholarlink

<p>Suboptimal feed bunk management in feedlot cattle, currently reliant on subjective manual and visual methods, significantly compromises animal health, production, and economic efficiency. The present paper proposes an objective method for feed bunk evaluation through the application of deep learning-based Convolutional Neural Networks. We trained the models using the images from the publicly available Feed Bunk Score Images dataset, composed of 1,511 images annotated according to the South Dakota State University score system. We proposed a specific fine-tuning configuration for this problem, freezing certain layers for each network evaluated. To improve the results, we combined the best configuration from each network into a weighted average ensemble. To find the weights that best combine the networks for the score classification problem we applied the Simulated Annealing metaheuristic. Furthermore, another contribution is the evaluation of image classification according to diet adjustments (increase, maintenance or decrease). The ensemble demonstrated high efficacy, achieving an overall accuracy of 95.37% for feed bunk score classification and an impressive 97.88% for diet adjustment classification. These results show the potential of using neural networks to support the feed bunk management practices.</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

Feed bunk score images classification using deep transfer learning

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Gabriel Rezende da Silva,
  • Brenda Marques de Paula,
  • Lucas Silva Santana,
  • Brian Luís Coimbra Maia,
  • Mathews Edwirds Gomes Almeida,
  • Sabrina Evelin Ferreira,
  • Luiz Maurílio da Silva Maciel,
  • Saulo Moraes Villela,
  • Amália Saturnino Chaves

摘要

Suboptimal feed bunk management in feedlot cattle, currently reliant on subjective manual and visual methods, significantly compromises animal health, production, and economic efficiency. The present paper proposes an objective method for feed bunk evaluation through the application of deep learning-based Convolutional Neural Networks. We trained the models using the images from the publicly available Feed Bunk Score Images dataset, composed of 1,511 images annotated according to the South Dakota State University score system. We proposed a specific fine-tuning configuration for this problem, freezing certain layers for each network evaluated. To improve the results, we combined the best configuration from each network into a weighted average ensemble. To find the weights that best combine the networks for the score classification problem we applied the Simulated Annealing metaheuristic. Furthermore, another contribution is the evaluation of image classification according to diet adjustments (increase, maintenance or decrease). The ensemble demonstrated high efficacy, achieving an overall accuracy of 95.37% for feed bunk score classification and an impressive 97.88% for diet adjustment classification. These results show the potential of using neural networks to support the feed bunk management practices.