Fishery Modernization ›› 2026, Vol. 53 ›› Issue (1): 131-143. doi: 10.26958/j.cnki.1007-9580.2026.01.012
Previous Articles Next Articles
HE Bingqing1,2 ,ZANG Zhaoxiang 3,4 ( 1 Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering, China Three Gorges University,Yichang 443002,Hubei,China;#br# 2 College of Computer and Information Technology,China Three Gorges University,Yichang 443002,Hubei,China;3 Key Laboratory of Artificial Intelligence in Yunnan Province,Kunming University of Science and Technology, Kunming 650504,Yunnan,China;4 Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,Yunnan,China)
何冰晴1,2 ,臧兆祥3,4
( 1 三峡大学水电工程智能视觉监测湖北省重点实验室 ,湖北 宜昌 443002;
2 三峡大学计算机与信息学院 ,湖北 宜昌 443002;
3 昆明理工大学云南省人工智能重点实验室 ,云南 昆明 650504;
4 昆明理工大学信息工程与自动化学院 ,云南 昆明 650504)
Abstract: To address challenges in underwater fish detection caused by turbid water and uneven illumination—such as blurred edges,high missed-detection rates,and the high computational cost of existing models,this paper proposes CRL-YOLO11,an improved lightweight detection algorithm based on YOLOv11n. Firstly,an Edge -Aware and Context -Guided Attention Block is proposed to enhance the model ’s perception of weakly represented targets,thereby improving the detection of small fish targets. Secondly,a lightweight and efficient aggregation module is designed,which leverages a re-parameterized multi-branch structure to enable cross-scale feature fusion and reduce information loss during feature propagation. In addition,to address the issue of high computational cost,a selective channel down-sampling module and a lightweight asymmetric detection head are introduced. Results obtained on the self-built dataset demonstrate that, compared with the baseline model YOLOv11n,the proposed CRL-YOLO11 achieves a 2. 1% improvement in mAP50 and a 2. 9% increase in recall,while reducing the model's weights and parameters to 86% and 82% of the original,respectively. Furthermore,on the Kaggle-Fish dataset,the mAP50 increased by 0. 7% and recall by 2. 4%. On the URPC2019 dataset,the mAP50 improved by 1. 1%. The experimental results demonstrate that the proposed model offers a balanced trade-off between detection accuracy,generalization,and deployment efficiency,rendering it well - suited for real-time underwater object detection in smart fisheries and aquaculture environments.
Key words:
underwater object detection,
YOLOv11,
edge awareness,
context guidance,
multi-scale feature fusion,
lightweight network,
fish detection
摘要： 针对水下图像中由于水体浑浊、光照不均而导致的鱼类目标边缘模糊、漏检率高以及现有模型计算复杂度大、部署困难的问题 ,本研究提出了一种基于改进 YOLOv11n 的水下鱼类检测算法 CRL-YOLO11。首先 ,设计了边缘感知与上下文引导注意力模块 ,以增强模型对弱特征目标的感知能力 ,从而提升小目标鱼类的检测效果 ;其次 ,构建了轻量化高效聚合模块 ,利用重参数化多分支结构实现跨尺度特征融合 ,降低信息传递损失;此外 ,针对计算成本较高的问题 ,引入了选择性通道下采样和轻量化非对称检测头 。在 自建数据集上试验结果显示 :与基准模型 YOLOv11n 相比 ,CRL-YOLO11 模型的 mAP50 提高了 2. 1% ,召回率提升了 2. 9% ; 同时权重和参数量分别减至原模型的 86%和82%;此外 ,在 Kaggle- fish 数据集中 ,mAP50 提高了 0. 7% ,召回率提升了 2. 4% ,在 URPC2019 数据集上 ,mAP50 提升了 1. 1%。研究表明 ,该模型具备良好的检测精度、泛化能力与部署效率 ,适用于智慧渔业与养殖场景下的实时水下目标检测任务。
