文献类型：期刊文献

中文题名：基于改进YOLOv11的海洋牧场中鲍的检测方法

英文题名：Underwater abalone detection in sea ranch based on improved YOLOv11

作者：李坤达[1,2];刘侦龙[1,2];王骥[2,3]

机构：[1]广东海洋大学数学与计算机学院,广东湛江524088;[2]广东海洋大学,广东省智慧海洋传感网及其装备工程技术研究中心,广东湛江524088;[3]广东海洋大学电子与信息工程学院,广东湛江524088

年份：2026

卷号：50

期号：3

起止页码：237

中文期刊名：水产学报

外文期刊名：Journal of Fisheries of China

收录：北大核心2023、、北大核心

基金：广东省普通高校重点领域新一代信息技术专项(2020ZDZX3008)。

语种：中文

中文关键词：鲍;YOLOv11;水下目标检测;注意力模块;特征提取模块;下采样模块;海洋牧场

外文关键词：abalone;YOLOv11;underwater target detection;attention module;feature extraction module;downsampling module;marine ranch

中文摘要：【目的】针对海洋牧场中鲍栖息环境复杂、能见度低与图像存在大量噪声等问题,该研究提出了一种基于改进YOLOv11模型的水中鲍识别方法 YOLOv11-AMSTAR。【方法】该模型的核心优化包括3个方面:首先,基于该模型使用C3K2和StarNet构建新的增强型特征提取模块(C3Star),通过星操作增强高维特征表达,在保留原始特征信息的同时挖掘隐藏的高阶关联信息,从而提升模型的非线性表达和特征区分能力。其次,引入下采样模块(ADown),对输入特征图维度的重新排列和细粒度调整,提升了模型中深层网络对空间特征的捕捉能力。最后,在颈部网络中加入自注意力与卷积混合模型(ACmix),融合不同层次的语义信息,增强模型对特征的提取和整合能力,降低杂乱背景信息干扰。【结果】相比于原始模型,YOLOv11-AMSTAR的mAP@0.5、召回率、准确率和mAP@0.5:0.95等指标分别提升5.21%、2.06%、2.66%和1.79%。【结论】YOLOv11-AMSTAR能显著增强在低对比度、模糊等恶劣水下环境中对鲍的特征提取能力,显著提高了检测精度。本研究不仅为水下生物的自动化、精准捕捞提供了高效可靠的技术方案,其针对低质图像和伪装目标的复合改进策略,也为解决其他类似复杂场景下的目标检测问题提供了重要的学术参考与应用价值。

外文摘要：Aiming at the problems of complex abalone habitat,low visibility with a large amount of noise in the image in the marine pasture,this study proposes an in-water abalone recognition method YOLOv11-AMSTAR based on the improved You Only Look Once version 11(YOLOv11)model.The core optimization of the model consists of three aspects:Firstly,a new enhanced feature extraction module(C3Star)is constructed using Cross Stage Partial with kernel size 2(C3K2)and StarNet,which enhances the high-dimensional feature representation by star operation,and mines the hidden higher-order correlation information while preserving the original feature information,thus improving the nonlinear representation and feature differentiation ability of the model.Secondly,the downsampling module Adaptive Downsampling(ADown)is introduced,which rearranges the dimensionality of the input feature maps and adjusts the fine-grainedness to enhance the ability of the deep network in the model to capture spatial features.Finally,Self-Attention and Convolution mix(ACmix)is added to the neck network to fuse different levels of semantic information,enhance the model's ability to extract and integrate features,and reduce the interference of cluttered background information.The experimental results show that compared with the original model,YOLOv11-AMSTAR achieved increases of 5.21%in mAP@0.5,2.06%in recall rate,2.66%in accuracy,and 1.79%in mAP@0.5:0.95 compared to the original model.The study shows that YOLOv11-AMSTAR can significantly enhance the feature extraction ability of abalone in harsh underwater environments such as low contrast and blur,and significantly improve the detection precision.This study not only provides an efficient and reliable technical solution for automated and accurate fishing of underwater organisms,but also its composite improvement strategy for low-quality images and camouflaged targets provides an important academic reference and application value for solving other target detection problems in similar complex scenes.