基于深度学习算法的凡纳滨对虾生长表型测定系统研发及应用  

Development and application of a deep learning algorithm-based growth phenotypes measurement system of the Pacific white shrimp(Litopenaeus vannamei)

作　　者：张士薇 代平[1,2] 高广春 孟宪红 罗坤[1,2] 隋娟 谭建[1,2] 傅强 曹家旺[1,2] 陈宝龙 李旭鹏 强光峰 邢群 戚云辉[4] 孔杰 栾生[1,2] ZHANG Shiwei;DAI Ping;GAO Guangchun;MENG Xianhong;LUO Kun;SUI Juan;TAN Jian;FU Qiang;CAO Jiawang;CHEN Baolong;LI Xupeng;QIANG Guangfeng;XING Qun;QI Yunhui;KONG Jie;LUAN Sheng(State Key Laboratory of Mariculture Biobreeding and Sustainable Goods,Yellow Sea Fisheries Research Institute,Chinese Academy of Fishery Sciences,Qingdao 266071,China;Laboratory for Marine Fisheries Science and Food Production Processes,Qingdao Marine Science and Technology Center,Qingdao 266237,China;College of Fisheries and Life Sciences,Dalian Ocean University,Dalian 116023,China;Hangzhou Feirui Technology Co.,Ltd.,Hangzhou 311100,China;School of Information and Electrical Engineering,Hangzhou City University,Hangzhou 310015,China;BLUP aquabreed Co.,Ltd.,Weifang 261312,China)

机构地区：[1]中国水产科学研究院黄海水产研究所,海水养殖生物育种与可持续产出全国重点实验室,山东青岛266071 [2]青岛海洋科技中心海洋渔业科学与食物产出过程功能实验室,山东青岛266237 [3]大连海洋大学水产与生命学院,辽宁大连116023 [4]杭州飞锐科技有限公司,浙江杭州311100 [5]浙大城市学院信息与电气工程学院,浙江杭州310015 [6]邦普种业科技有限公司,山东潍坊261312

出　　处：《水产学报》2025年第5期196-208,共13页Journal of Fisheries of China

基　　金：国家重点研发计划(2022YFD2400202);财政部和农业农村部:国家现代农业产业技术体系专项(CARS-48);中国水产科学研究院科技创新团队项目(2020TD26);泰山学者工程;广东省“十四五”农业科技创新十大主攻方向揭榜挂帅项目(2022SDZG01);山东省科技型中小企业创新能力提升工程项目(2023TSGC0744)。

摘　　要：【目的】为解决凡纳滨对虾生长表型人工测量效率低、误差大等问题。【方法】本研究设计了对虾专用图像采集箱,获取标准化、高质量的对虾侧面图像。在此基础上,利用高分辨率网络(HRNet)模型识别凡纳滨对虾9个关键特征点,实现对体长等体尺相关性状的测量;基于掩膜卷积神经网络(Mask R-CNN)进行凡纳滨对虾的轮廓分割,实现对虾体表面积的计算;最后复合体长以及体表面积构建回归模型预测对虾体重。通过开发配套的图像处理与数据管理软件,建立凡纳滨对虾生长表型精准测定系统。【结果】HRNet模型对9个特征点的识别率均超过98%,其中7个特征点的识别率超过99%。使用直尺人工测量和图像人工标注特征点测量两种方法测定体长和腹节长的真实值,计算体长和腹节长的预测准确性分别为0.91~0.97和0.91~0.93,平均相对误差分别为1.39%~4.63%和2.46%~4.59%。以人工分割虾体轮廓方式获取体表面积作为参考,评估Mask R-CNN模型对体表面积的预测准确性为0.98,平均相对误差为1.73%。以体长、体表面积、性别为变量,构建了4种回归模型来预测体重,所有模型的准确性均在0.94以上,其中以同时包含体长和体表面积的模型的预测准确性最高(0.97)。【结论】利用深度学习算法可以较为准确地获得凡纳滨对虾体长和体表面积等生长表型并预测体重。本研究结果可为凡纳滨对虾生长表型性状的准确、快速测量提供高效工具。To address the low efficiency and high error rates associated with manual measurement of growth phenotypes in the Pacific white shrimp(Litopenaeus vannamei),this study developed a dedicated image acquisition box capable of capturing standardized,high-quality side-view images of the shrimp.Utilizing this system,a High-Resolution Network(HRNet)model was employed to identify nine key feature points of the shrimp,enabling the measurement of traits such as body length.Additionally,a Mask Region Convolutional Neural Network(Mask R-CNN)model was utilized for shrimp contour segmentation to calculate body surface area.Regression models incorporating body length and body surface area were subsequently developed to predict body weight.An integrated image processing and data management software was also developed to establish a precise measurement system for the growth phenotypes of L.vannamei.The study found that the HRNet model achieved recognition rates exceeding 98%for all nine feature points,with rates exceeding 99%for seven points.The true values of body length and abdominal segment length were measured using two methods:manual measurement with a ruler and measurement from manually tagged feature points in the images.The predictive accuracy of body length and abdominal segment length was calculated to be 0.91-0.97 and 0.91-0.93,respectively,with average relative errors of 1.39%-4.63%and 2.46%-4.59%.Evaluation against manually segmented shrimp body contours showed that the Mask R-CNN model predicted body surface area with an accuracy of 0.98 and an average relative error of 1.73%.Regression models incorporating variables such as body length,body surface area,and gender were developed to predict body weight,achieving accuracies above 0.94,with the model incorporating both body length and body surface area achieving the highest prediction accuracy(0.97).These results demonstrate that computer vision technology combined with deep learning algorithms can accurately measure growth phenotypes,such as body length and body surf

关 键 词：凡纳滨对虾 生长表型 深度学习 计算机视觉 测定系统 

分 类 号：S917.4[农业科学—水产科学] TP181[自动化与计算机技术—控制理论与控制工程]