Scholarlink

<p>Evaluating fish school&#xa0;feeding intensity is essential for optimizing feed efficiency, lowering aquaculture production costs, and ensuring healthy fish development in pond culture. Given that fish schools exhibit diverse feeding morphologies with dynamically shifting aggregation patterns in the pond environment, current models struggle to capture the critical spatial relationships between localized feeding hotspots and global distribution characteristics, leading to compromised accuracy. To tackle this issue, a fish school feeding intensity assessment method based on EfficientNetB0-FPSANet is proposed in this paper. At the data preparation phase, input images first undergo a specialized water surface glare removal processing to eliminate overexposed artifacts caused by water surface reflections, followed by random crops and statistical normalization to optimize visual quality and enhance sample diversity. Secondly, the MBConv layer of EfficientNetB0 is redesigned by removing the Squeeze-Excitation (SE), achieving a lighter architecture without compromising efficacy. Meanwhile, the Fusion Pyramid Squeeze-Excitation Attention (FPSA) is appended at the end of EfficientNetB0, which extracts features at varying scales and effectively preserves the local details and global information. Finally, through rigorous testing on actual fish school image datasets, the developed EfficientNetB0-FPSANet model delivers 98.73% assessment accuracy. Our model achieves superior performance with an accuracy improvement of 2.69%, 2.50%, 2.34%, 1.71%, and 1.56% in comparison with other fish school feeding intensity assessment models based on Swin_transformer, ConvNeXt_base, ResNet50, ShuffleNetV2, and DenseNet121, respectively. Experimental results confirm the effectiveness of our approach in accurately assessing fish school&#xa0;feeding intensity in practical aquaculture environments.</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

A fish school feeding intensity assessment method based on EfficientNetB0-FPSANet

  • 期刊论文

  • 发表日期：

    2025年11月28日

查看全文
  • Lu Zhang,
  • Le Wang,
  • Suchang Li

摘要

Evaluating fish school feeding intensity is essential for optimizing feed efficiency, lowering aquaculture production costs, and ensuring healthy fish development in pond culture. Given that fish schools exhibit diverse feeding morphologies with dynamically shifting aggregation patterns in the pond environment, current models struggle to capture the critical spatial relationships between localized feeding hotspots and global distribution characteristics, leading to compromised accuracy. To tackle this issue, a fish school feeding intensity assessment method based on EfficientNetB0-FPSANet is proposed in this paper. At the data preparation phase, input images first undergo a specialized water surface glare removal processing to eliminate overexposed artifacts caused by water surface reflections, followed by random crops and statistical normalization to optimize visual quality and enhance sample diversity. Secondly, the MBConv layer of EfficientNetB0 is redesigned by removing the Squeeze-Excitation (SE), achieving a lighter architecture without compromising efficacy. Meanwhile, the Fusion Pyramid Squeeze-Excitation Attention (FPSA) is appended at the end of EfficientNetB0, which extracts features at varying scales and effectively preserves the local details and global information. Finally, through rigorous testing on actual fish school image datasets, the developed EfficientNetB0-FPSANet model delivers 98.73% assessment accuracy. Our model achieves superior performance with an accuracy improvement of 2.69%, 2.50%, 2.34%, 1.71%, and 1.56% in comparison with other fish school feeding intensity assessment models based on Swin_transformer, ConvNeXt_base, ResNet50, ShuffleNetV2, and DenseNet121, respectively. Experimental results confirm the effectiveness of our approach in accurately assessing fish school feeding intensity in practical aquaculture environments.