Scholarlink

<p>Accurately detecting fish feeding behavior is crucial for optimizing feeding strategies and improving aquaculture efficiency. This study presents a novel deep learning-based model, SA-RepVGG-PLK, which combines the strengths of three advanced techniques: SA-Net for spatial attention, RepVGG for efficient feature extraction, and Pyramidal Lucas–Kanade for capturing dynamic motion information. SA-RepVGG-PLK was trained and evaluated on a dataset of sea bass feeding behavior collected from a controlled aquaculture environment. Experimental results show that SA-RepVGG-PLK achieves state-of-the-art performance, with an accuracy of 95.8%, precision of 96.9%, recall of 97.0%, and an F1 score of 97.0%, surpassing state-of-the-art convolutional models. The integration of motion information between feeding frames significantly reduces misjudgments and enhances the robustness of the predictions. While SA-RepVGG-PLK sacrifices some inference speed, its precision and reliability make it ideal for applications requiring high accuracy, though further optimization is needed for real-time deployment on edge devices. Future research will focus on optimizing computational efficiency and extending SA-RepVGG-PLK's application to diverse aquaculture environments. The proposed SA-RepVGG-PLK uniquely integrates spatial attention, efficient re-parameterized convolutions, and multi-scale motion tracking, achieving robust feeding detection in dynamic underwater environments, which is critical for reducing feed waste in large-scale aquaculture.</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-model fusion method for detecting the feeding behavior of sea bass

  • 期刊论文

  • 发表日期：

    2025年7月17日

查看全文
  • Wan Yi Li,
  • Teng Ke Li,
  • Zi Heng Liao,
  • Qiu Xian Wu,
  • Jun Quan Lin,
  • Ting Wu,
  • Li Lin,
  • Hai Xia Xu,
  • Ling Yang,
  • Juan Zou

摘要

Accurately detecting fish feeding behavior is crucial for optimizing feeding strategies and improving aquaculture efficiency. This study presents a novel deep learning-based model, SA-RepVGG-PLK, which combines the strengths of three advanced techniques: SA-Net for spatial attention, RepVGG for efficient feature extraction, and Pyramidal Lucas–Kanade for capturing dynamic motion information. SA-RepVGG-PLK was trained and evaluated on a dataset of sea bass feeding behavior collected from a controlled aquaculture environment. Experimental results show that SA-RepVGG-PLK achieves state-of-the-art performance, with an accuracy of 95.8%, precision of 96.9%, recall of 97.0%, and an F1 score of 97.0%, surpassing state-of-the-art convolutional models. The integration of motion information between feeding frames significantly reduces misjudgments and enhances the robustness of the predictions. While SA-RepVGG-PLK sacrifices some inference speed, its precision and reliability make it ideal for applications requiring high accuracy, though further optimization is needed for real-time deployment on edge devices. Future research will focus on optimizing computational efficiency and extending SA-RepVGG-PLK's application to diverse aquaculture environments. The proposed SA-RepVGG-PLK uniquely integrates spatial attention, efficient re-parameterized convolutions, and multi-scale motion tracking, achieving robust feeding detection in dynamic underwater environments, which is critical for reducing feed waste in large-scale aquaculture.