基于LoFTR特征匹配算法的羊只身份识别  

Sheep Identification Based on LoFTR Feature Matching Algorithm

作　　者：崔家赫 宣传忠[1] 宋耀邦 王光普 CUI Jiahe;XUAN Chuanzhong;SONG Yaobang;WANG Guangpu(College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,nner Mongolia Engineering Research Center for Intelligent Facilities in Prataculture and Livestock Breeding,Hohhot 010018,China)

机构地区：[1]内蒙古农业大学机电工程学院/内蒙古自治区草业与养殖业智能装备工程技术研究中心,呼和浩特010018

出　　处：《内蒙古农业大学学报（自然科学版）》2024年第2期51-61,共11页Journal of Inner Mongolia Agricultural University(Natural Science Edition)

基　　金：内蒙古直属高校基本科研业务费项目(BR221032);内蒙古自治区科技计划项目(2021GG0111)。

摘　　要：在现代农牧业生产中,自动化家畜身份识别是提高育种选择和行为研究效率的关键环节。为了解决传统羊只身份识别方法耗时、成本高和准确性不足的问题,近年来学者们开始运用深度学习技术对羊只脸部生物特征进行学习,进一步实现身份识别。然而,基于羊脸图像的识别方法存在稳定性不足、识别性能波动等问题。因此,本研究开发了一种基于特征点匹配的羊只身份识别方法,即搭建LoFTR算法对羊脸特征点进行学习和匹配,该模型通过学习羊只面部的生物特征,实现了通过特征点匹配来识别个体身份的技术。针对以往研究中羊脸图像采集角度单一、识别效果不稳定的局限,本研究采集了89只小尾寒羊的多角度面部图像以提取更高质量的羊脸特征,并进一步构建了一个羊脸数据集用于模型的训练和识别。通过与经典的特征点匹配模型进行性能比较,试验结果表明LoFTR羊脸特征点匹配模型展现出更高的识别精度和效率。该模型在羊脸图像数据集上达到93.53%的平均识别准确率,而且平均匹配速率也达到了2.39 s。这一方法的应用有助于提高羊场的管理效率,推动羊场管理向现代化迈进。此外,提高羊只面部身份识别的准确性和效率,可以有效提高育种选择和行为研究的精度和速度,为农牧业生产提供更好的技术支持,促进行业的可持续发展。In modern agricultural and animal husbandry production,automated livestock identification is a crucial factor in enhanc-ing breeding selection and behavioral research efficiency.To address the issues of time consumption,high costs,and insuffi-cient accuracy associated with traditional sheep identification methods,scholars have recently begun to use deep learning technolo-gy to learn the sheep facial biological features,to further realize the identity recognition.However,identification methods based on the sheep facial images has some problems,such as the instability and fluctuation in recognition performance.Therefore,this study developed a sheep identification method based on matching the feature points.Specifically,the LoFTR algorithm was em-ployed to learn and match the facial feature points of sheep.By learning the biological features of sheep faces,this model achieved the technology of identifying individual identities through matching the feature points.In view of the limitations of single sheep face im-age acquisition angle and unstable recognition effect in previous studies,this study collected multi-angle facial images of 89 small-tailed Han sheep to extract higher-quality sheep facial features and further constructed a sheep facial dataset for model training and recognition.Comparative performance experiments with classical feature point matching models indicated that the LoFTR sheep fa-cial feature point matching model showed higher recognition accuracy and efficiency.This model achieved an average recognition accuracy of 93.53%on sheep facial image datasets,respectively,and the average matching rates also reached 2.39 s.The applica-tion of this method contributed to improve the management efficiency of sheep farms and promote the modernization of sheep farm management.By improving the accuracy and efficiency of sheep facial identification,it effectively enhanced the precision and speed of breeding selection and behavioral research,providing better technical supports for agricultural and animal husband

关 键 词：羊脸识别 LoFTR 深度学习 特征点匹配 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]