基于YOLOv5模型的仔猪社交识别方法研究  

Research on Social Relationship of Piglets Based on YOLOv5

作　　者：冯兴尧 王海峰 朱君 孙想[1] 邱阳 李斌[1,2] FENG Xingyao;WANG Haifeng;ZHU Jun;SUN Xiang;QIU Yang;LI Bin(Intelligent Equipment Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China;College of Engineering,Huazhong Agricultural University,Wuhan 430070,China;Shandong Jialejia Agriculture and Animal Husbandry Technology Co.Ltd.,Weifang 262204,China)

机构地区：[1]北京市农林科学院智能装备技术研究中心,北京100097 [2]华中农业大学工学院,湖北武汉430070 [3]山东佳乐家农牧科技有限公司,山东潍坊262204

出　　处：《中国猪业》2024年第1期73-83,共11页China Swine Industry

基　　金：山东省重点研发计划项目(2022TZXD0014);北京市平谷区“博士”农场项目;北京市农林科学院智能装备技术研究中心开放课题项目(KFZN2020W011);2023年北京市农林科学院财政专项。

摘　　要：精准识别仔猪间社交关系对了解仔猪内部社交和预警异常仔猪具有重要意义。针对传统方法在仔猪社交识别时存在的人工依赖多、劳动强度大、观测效率低等问题,本研究借助机器视觉与深度学习技术,提出了一种基于改进的YOLOv5模型的仔猪社交识别研究方法。该研究以9头30~35日龄群养的长白二元杂交仔猪为研究对象,从顶部视角连续采集视频数据,经图像截取与数据增强共获得13389张图像作为数据集。首先,选取Faster R-CNN、SSD、YOLOv4和YOLOv5这4种典型目标检测算法对数据集进行训练,通过对比分析,确定用于仔猪个体身份识别最优模型;然后依据K-means聚类算法确定仔猪社交中心,通过计算仔猪与社交中心的欧氏距离量化仔猪社交值,利用位置信息构建仔猪社交网络,绘制仔猪运动轨迹,获得社交正常与社交异常仔猪的识别阈值;最后,利用该阈值对仔猪进行分类,识别社交异常仔猪个体并实现预警。经测试,改进的YOLOv5对群养仔猪个体身份识别的平均精度均值达99.29%,模型大小为13.71 MB,满足仔猪身份识别需求,与YOLOv5、YOLOv4、SSD和Faster R-CNN模型相比,改进的YOLOv5平均精度均值分别提高了0.26、1.97、12.74和4.31个百分点。通过统计仔猪社交值均值变化情况,发现正常与异常仔猪社交值均值差异明显,正常仔猪社交值均值范围[0.259~0.351],异常仔猪社交值均值范围[0.402~0.441],试验确定0.4为最佳社交判别阈值。该研究可为仔猪社交行为智能识别与异常早期预警提供方法参考。Accurately recognized the social relationship among piglets was important to understand the internal socialization of piglets and early warning signs of abnormal piglets.In response to the issues of high dependence on human intervention,high labor intensity and low observation efficiency in traditional piglet social relationship studied,this study proposed a research method of piglet social relationship based on the YOLOv5 model,utilized machine vision and deep learning technology.The study was conducted with nine Long White binary crossbred piglets reared in groups of 30 to 35 days of age,continuously collecting video data from a top-down perspective.By conducting image capture and data augmentation,a dataset of 13389 images was obtained.Firstly,four typical target detection algorithms,namely Faster R-CNN,SSD,YOLOv4 and YOLOv5,were selected to train the dataset,and the optimal model for individual piglet identification was determined through comparative analysis.Then,based on the K-means clustering algorithm to determine the social center of piglets,quantify the social value of piglets by calculating the Euclidean distance between the piglets and the social center,construct the social network of piglets by using the location information,drew the movement trajectory of piglets and obtained the identification thresholds of socially normal and socially abnormal piglets.Finally,piglets were classified to identify piglets with abnormal social relationships using the threshold and enable early warning.The experimental results showed that YOLOv5 achieved an average precision-recall of 99.29%for individual piglet identification in group housing.The model size was 13.71 MB which could satisfy the requirements for piglet identification.Compared with YOLOv5,YOLOv4,SSD,and Faster R-CNN models,the improved YOLOv5 increased average precision-recall by 0.26%,1.97%,12.74%and 4.31%,respectively.It was found that there was a significant difference between the mean social values of abnormal and normal piglets.The mean social value

关 键 词：仔猪 YOLOv5 K-MEANS 社交 深度学习 

分 类 号：S828[农业科学—畜牧学] TP391.41[农业科学—畜牧兽医]