Scholarlink

<p>Accurate measurement of fish morphological parameters is essential for precision aquaculture, supporting biomass estimation, feeding optimization, and health monitoring. However, underwater imaging is challenged by light attenuation, turbidity, and dynamic fish postures, which degrade image quality and reduce detection accuracy. To address these issues, this paper proposes FishMorphNet, an improved deep learning model built on YOLOv12 that incorporates three key innovations: (1) a novel bounding box regression loss function, Fish-IoU, specifically designed for high-aspect-ratio fish targets; (2) integration of Convolutional Block Attention Modules (CBAM) into the backbone to enhance feature extraction in complex underwater scenes; and (3) replacement of standard convolutions with Depthwise Separable Convolutions (DWConv) in the neck to reduce computational cost while maintaining accuracy. The core contribution of this study lies in the proposal of Fish-IoU. By fundamentally decoupling the distance loss term from the shape penalty term, its mathematical design is specifically tailored to address the critical challenges of regressing high-aspect-ratio fish targets. Experiments were conducted on two fish species, largemouth bass and spotted knifejaw, using a binocular stereo vision system in real aquaculture environments. The proposed method achieves superior detection performance, with a mean Average Precision at 50-95% IoU (mAP@50-95) of 0.800,&#xa0;and significantly improves body length and height estimation accuracy, reducing Root Mean Square Error (RMSE) by over 60% compared to the baseline. Fish-IoU consistently outperforms existing IoU-based loss functions across all evaluation metrics. This work contributes to the advancement of intelligent aquaculture systems by enabling accurate, efficient, and automated fish morphometric measurement.</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

FishMorphNet: A YOLOv12-based deep learning model with Fish-IoU loss for accurate fish morphometric measurement in underwater vision systems

  • 期刊论文

  • 发表日期：

    2026年8月1日

查看全文
  • Yufei Ren,
  • Lin He,
  • Tianye Zhang,
  • Daoliang Li,
  • Chen Shi

摘要

Accurate measurement of fish morphological parameters is essential for precision aquaculture, supporting biomass estimation, feeding optimization, and health monitoring. However, underwater imaging is challenged by light attenuation, turbidity, and dynamic fish postures, which degrade image quality and reduce detection accuracy. To address these issues, this paper proposes FishMorphNet, an improved deep learning model built on YOLOv12 that incorporates three key innovations: (1) a novel bounding box regression loss function, Fish-IoU, specifically designed for high-aspect-ratio fish targets; (2) integration of Convolutional Block Attention Modules (CBAM) into the backbone to enhance feature extraction in complex underwater scenes; and (3) replacement of standard convolutions with Depthwise Separable Convolutions (DWConv) in the neck to reduce computational cost while maintaining accuracy. The core contribution of this study lies in the proposal of Fish-IoU. By fundamentally decoupling the distance loss term from the shape penalty term, its mathematical design is specifically tailored to address the critical challenges of regressing high-aspect-ratio fish targets. Experiments were conducted on two fish species, largemouth bass and spotted knifejaw, using a binocular stereo vision system in real aquaculture environments. The proposed method achieves superior detection performance, with a mean Average Precision at 50-95% IoU (mAP@50-95) of 0.800, and significantly improves body length and height estimation accuracy, reducing Root Mean Square Error (RMSE) by over 60% compared to the baseline. Fish-IoU consistently outperforms existing IoU-based loss functions across all evaluation metrics. This work contributes to the advancement of intelligent aquaculture systems by enabling accurate, efficient, and automated fish morphometric measurement.