Scholarlink

<p>To accurately evaluate the feeding intensity of pearl gentian grouper and overcome the vulnerability of single-modal features in complex aquaculture environments, this study proposes a multi-feature fusion strategy combined with an improved lightweight multi-dilated ghost shuffle network (MDGSNet), model. Acoustic data were collected from groupers of three size groups (200&#xa0;g, 400&#xa0;g, and 700&#xa0;g), denoised, and transformed into Log-Mel spectrograms, gammatone frequency cepstral coefficient (GFCC) maps, and root mean square (RMS) energy envelopes. These features were fused through RGB channels to form composite inputs. The proposed MDGSNet, built upon ShuffleNetV2, integrates multi-scale dilated convolutions, squeeze-and-excitation (SE) attention, and the multi-dilation attention receptive field–GhostBottleneck (MDAR-GhostBottleneck) structure, enhancing feature extraction while maintaining efficiency. Experimental results show that MDGSNet achieves an accuracy of 98.04% and an F1 score of 0.977, outperforming mainstream models such as ResNet-18 and GhostNet, while requiring only 2.46&#xa0;MB of parameters and 0.23 GFLOPs. This study demonstrates that integrating multi-dimensional acoustic features with a lightweight model offers a robust solution for accurate feeding intensity assessment, providing a foundation for intelligent feeding management in factory 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

Evaluation of feeding intensity in pearl gentian grouper under factory aquaculture environment based on multi-feature fusion and MDGSNet model

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Zeshuang Ma,
  • Hongpo Wang,
  • Hong Zhou,
  • Yunchen Tian,
  • Jianing Quan,
  • Haohui Liu,
  • Zhanshuo Sun,
  • Hongjun Yang

摘要

To accurately evaluate the feeding intensity of pearl gentian grouper and overcome the vulnerability of single-modal features in complex aquaculture environments, this study proposes a multi-feature fusion strategy combined with an improved lightweight multi-dilated ghost shuffle network (MDGSNet), model. Acoustic data were collected from groupers of three size groups (200 g, 400 g, and 700 g), denoised, and transformed into Log-Mel spectrograms, gammatone frequency cepstral coefficient (GFCC) maps, and root mean square (RMS) energy envelopes. These features were fused through RGB channels to form composite inputs. The proposed MDGSNet, built upon ShuffleNetV2, integrates multi-scale dilated convolutions, squeeze-and-excitation (SE) attention, and the multi-dilation attention receptive field–GhostBottleneck (MDAR-GhostBottleneck) structure, enhancing feature extraction while maintaining efficiency. Experimental results show that MDGSNet achieves an accuracy of 98.04% and an F1 score of 0.977, outperforming mainstream models such as ResNet-18 and GhostNet, while requiring only 2.46 MB of parameters and 0.23 GFLOPs. This study demonstrates that integrating multi-dimensional acoustic features with a lightweight model offers a robust solution for accurate feeding intensity assessment, providing a foundation for intelligent feeding management in factory aquaculture.