Scholarlink

<p>Accurate detection of dead fish is critical for enhancing aquaculture productivity and mitigating environmental pollution. In recirculating aquaculture systems, existing machine vision studies have focused inadequately on lightweight models for identifying dead fish in mixed fish populations. To address this, we propose YOLO-LD, which incorporates three key improvements: (1) Replacing YOLOv8n’s C2f module with Light UIB-SAM to simplify feature extraction channels, reduce resource consumption, and boost inference speed; (2) incorporating a Global–Local Interactive Feature Fusion Module into the neck to enhance multi-scale feature fusion, mitigate boundary blurring and water-body noise caused by underwater light refraction; (3) introducing the Depthwise-Residual Fusion Block in the head to enhance channel weighting of dead fish targets, boost the model’s attention to targets, further reduce computational cost, and improve recognition ability. Validated on a co-cultured largemouth bass and crucian carp dataset, YOLO-LD achieves 0.903 ± 0.006 mAP@50 and 396 ± 2.1 FPS. Compared with the baseline model YOLOv8n, it reduces parameters by 13.3%, GFLOPs (Giga Floating-Point Operations per Second) by 25.9%, and model size by 11.1%. This study provides technical references for RAS dead fish detection and supports lightweight aquaculture monitoring.</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 scene-adapted lightweight YOLO-LD model for dead fish detection in recirculating aquaculture systems

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Weihong Wei,
  • Maochun Wei,
  • Hongcheng Li,
  • Guobin Hong

摘要

Accurate detection of dead fish is critical for enhancing aquaculture productivity and mitigating environmental pollution. In recirculating aquaculture systems, existing machine vision studies have focused inadequately on lightweight models for identifying dead fish in mixed fish populations. To address this, we propose YOLO-LD, which incorporates three key improvements: (1) Replacing YOLOv8n’s C2f module with Light UIB-SAM to simplify feature extraction channels, reduce resource consumption, and boost inference speed; (2) incorporating a Global–Local Interactive Feature Fusion Module into the neck to enhance multi-scale feature fusion, mitigate boundary blurring and water-body noise caused by underwater light refraction; (3) introducing the Depthwise-Residual Fusion Block in the head to enhance channel weighting of dead fish targets, boost the model’s attention to targets, further reduce computational cost, and improve recognition ability. Validated on a co-cultured largemouth bass and crucian carp dataset, YOLO-LD achieves 0.903 ± 0.006 mAP@50 and 396 ± 2.1 FPS. Compared with the baseline model YOLOv8n, it reduces parameters by 13.3%, GFLOPs (Giga Floating-Point Operations per Second) by 25.9%, and model size by 11.1%. This study provides technical references for RAS dead fish detection and supports lightweight aquaculture monitoring.