Scholarlink

<p>Accurate fish counting in pond aquaculture provides crucial scientific guidance for dynamic water quality regulation, disease early warning, precise feeding, and ecological benefit assessment. To address the challenges posed by multi-scale variation and dense occlusion in complex pond environments, a Multi-scale Adaptive Fusion Attention YOLO v11n (MAFA-YOLO v11n) model for robust fish detection and counting was proposed in this paper. First, a comprehensive dataset was constructed by collecting and preprocessing fish images from multiple ponds at various angles and time intervals. Secondly, the Convolutional Block Attention Module (CBAM) was introduced into the multi-scale feature fusion path of the YOLOv11 neck to enhance the perception capability for key target regions of fish bodies. Finally, the Fish Adaptive Spatial Feature Fusion Head (FASFFHead) was proposed to perform dynamic multi-level feature fusion expansion. This architecture effectively alleviates the problem of insufficient cross-scale integration by adaptively weighting features from different network layers. In the specific context of pond fish counting, this design ensures the accurate capture of both small-scale distant fish and large-scale near-field individuals while maintaining robust detection performance during high-density overlaps characteristic of feeding periods. In practical pond management, real-time surface fish counting at feeding zones serves as a critical proxy indicator for local aggregation density, as complete surface emergence of all fish during feeding cannot be guaranteed and a total population census of the entire pond is physically constrained. This numerical feedback provides a quantitative basis for density-responsive feeding management, where feed distribution can be adjusted according to the observed clustering of fish at the water surface. Evaluated on a self-built fish school dataset, the proposed MAFA-YOLO v11n model achieves a precision of 87.9%, a recall of 82.9%, an mAP@50 of 91.1%, and an mAP@50–95 of 56.3%, demonstrating competitive detection performance. Experimental results confirm that the method can effectively accomplish reliable fish counting in complex, real-world pond aquaculture scenarios, providing practical technical support for intelligent aquaculture management.</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 counting method based on MAFA-YOLO v11n for pond aquaculture

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Lu Zhang,
  • Hao Yang,
  • Shunshun Zhou,
  • Wenhui Ni,
  • Bin Li

摘要

Accurate fish counting in pond aquaculture provides crucial scientific guidance for dynamic water quality regulation, disease early warning, precise feeding, and ecological benefit assessment. To address the challenges posed by multi-scale variation and dense occlusion in complex pond environments, a Multi-scale Adaptive Fusion Attention YOLO v11n (MAFA-YOLO v11n) model for robust fish detection and counting was proposed in this paper. First, a comprehensive dataset was constructed by collecting and preprocessing fish images from multiple ponds at various angles and time intervals. Secondly, the Convolutional Block Attention Module (CBAM) was introduced into the multi-scale feature fusion path of the YOLOv11 neck to enhance the perception capability for key target regions of fish bodies. Finally, the Fish Adaptive Spatial Feature Fusion Head (FASFFHead) was proposed to perform dynamic multi-level feature fusion expansion. This architecture effectively alleviates the problem of insufficient cross-scale integration by adaptively weighting features from different network layers. In the specific context of pond fish counting, this design ensures the accurate capture of both small-scale distant fish and large-scale near-field individuals while maintaining robust detection performance during high-density overlaps characteristic of feeding periods. In practical pond management, real-time surface fish counting at feeding zones serves as a critical proxy indicator for local aggregation density, as complete surface emergence of all fish during feeding cannot be guaranteed and a total population census of the entire pond is physically constrained. This numerical feedback provides a quantitative basis for density-responsive feeding management, where feed distribution can be adjusted according to the observed clustering of fish at the water surface. Evaluated on a self-built fish school dataset, the proposed MAFA-YOLO v11n model achieves a precision of 87.9%, a recall of 82.9%, an mAP@50 of 91.1%, and an mAP@50–95 of 56.3%, demonstrating competitive detection performance. Experimental results confirm that the method can effectively accomplish reliable fish counting in complex, real-world pond aquaculture scenarios, providing practical technical support for intelligent aquaculture management.