Scholarlink

<p>Recognition of fish feeding intensity enables precision feeding, which is beneficial for enhancing economic returns and protecting the aquatic environment. Vision-based methods are often hindered by water turbidity and surface disturbances, while existing acoustic approaches face high cost, model complexity, and noise sensitivity. To address low-cost, lightweight, and robust requirements, we propose FTPG-CNN6, an improved Convolutional Neural Network (CNN) with Frequency-Time Attention (FT-Attention), Frequency-Prioritized convolution (FP-Conv), and GhostConv. Raw audio signals collected by low-cost hydrophones are processed to generate Mel-frequency cepstral coefficient (MFCC) feature maps, reducing data dimensionality while preserving key acoustic information. These feature maps are then input into the FTPG-CNN6 network, which incorporates the three modifications to enhance recognition: GhostConv to reduce computational cost, FP-Conv to capture critical frequency features and adapt to the uneven frequency distribution of signals collected by low-cost hydrophones, the FT-Attention mechanism to emphasize important regions and mitigate the impact of environmental noise. The resulting lightweight model is deployed on embedded devices, enabling real-time recognition of fish feeding intensity. On a self-built dataset, the method achieved 97.30% accuracy, outperforming EfficientNet and ResNet18 by 5.37% and 3.35%, respectively, while reducing memory use by 32.16% and 71.17%. These results demonstrate a low-cost, lightweight, and robust solution suitable for real-time fish feeding recognition, providing practical support for precision feeding.</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

Fish feeding intensity recognition in aquaculture based on MFCC and an improved CNN

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Chenlong Li,
  • Hongyu Pan,
  • Mingrui Kong,
  • Dingshuo Liu,
  • Qingling Duan

摘要

Recognition of fish feeding intensity enables precision feeding, which is beneficial for enhancing economic returns and protecting the aquatic environment. Vision-based methods are often hindered by water turbidity and surface disturbances, while existing acoustic approaches face high cost, model complexity, and noise sensitivity. To address low-cost, lightweight, and robust requirements, we propose FTPG-CNN6, an improved Convolutional Neural Network (CNN) with Frequency-Time Attention (FT-Attention), Frequency-Prioritized convolution (FP-Conv), and GhostConv. Raw audio signals collected by low-cost hydrophones are processed to generate Mel-frequency cepstral coefficient (MFCC) feature maps, reducing data dimensionality while preserving key acoustic information. These feature maps are then input into the FTPG-CNN6 network, which incorporates the three modifications to enhance recognition: GhostConv to reduce computational cost, FP-Conv to capture critical frequency features and adapt to the uneven frequency distribution of signals collected by low-cost hydrophones, the FT-Attention mechanism to emphasize important regions and mitigate the impact of environmental noise. The resulting lightweight model is deployed on embedded devices, enabling real-time recognition of fish feeding intensity. On a self-built dataset, the method achieved 97.30% accuracy, outperforming EfficientNet and ResNet18 by 5.37% and 3.35%, respectively, while reducing memory use by 32.16% and 71.17%. These results demonstrate a low-cost, lightweight, and robust solution suitable for real-time fish feeding recognition, providing practical support for precision feeding.