基于图像匹配的宽基线牛只面部图像识别研究  

Facial recognition of cattle based on wide-baseline image-matching

作　　者：翁智[1,2] 刘永兴 刘科 郑志强[1,2] WENG Zhi;LIU Yongxing;LIU Ke;ZHENG Zhiqiang(School of Electronic Information Engineering,Inner Mongolia University,Hohhot 010021,China;State Key Laboratory of Reproductive Regulation&Breeding of Grassland Livestock,Inner Mongolia University,Hohhot 010030,China)

机构地区：[1]内蒙古大学电子信息工程学院,呼和浩特010021 [2]内蒙古大学草原家畜生殖调控与繁育国家重点实验室,呼和浩特010030

出　　处：《中国农业大学学报》2025年第3期49-59,共11页Journal of China Agricultural University

基　　金：国家自然科学基金项目(61966026);内蒙古自治区高等学校青年科技英才支持计划(NJYT23063)。

摘　　要：为提升识别模型对非标准化数据的适应性,本研究设计了一种基于图像匹配的宽基线牛只面部识别方法,采用SuperPoint与SuperGlue构建图像匹配算法,并针对性改进以提高识别准确率。在特征点提取过程中,引入了基于图像质量评估的动态阈值调整策略,对图像的清晰度、边缘密度及纹理特性进行量化评估,并据此调整SuperPoint的阈值,实现高质量特征点提取。同时,通过评估特征点的分布熵与空间覆盖性,调整SuperGlue的匹配阈值,以实现高效匹配。为验证方法有效性,在自建牛只面部数据集上与多种图像匹配方法进行了对比试验。结果表明:1)算法在窄基线数据集上的宏平均精确率、宏平均召回率及宏平均F1分别为92.1%、90.4%和91.2%。2)算法在宽基线数据集上的宏平均精确率、宏平均召回率及宏平均F1分别为82.8%、81.0%和81.9%。3)算法在公开数据集上的微平均F1得分为86.4%。各项结果均显著优于传统图像匹配算法。综上,本研究提出了一种全新的图像匹配算法,有效提升了牛只面部识别模型对非标准化数据的适应性,为牛面部识别方法的实际应用提供技术参考。To enhance the adaptability of the recognition model to non-standardized data,this study designed a Wide-Baseline cattle facial recognition method based on image matching,using SuperPoint and SuperGlue to construct the image matching algorithm,and made targeted improvements to improve the recognition accuracy.In the process of feature point extraction,a dynamic threshold adjustment strategy based on image quality assessment was introduced to quantitatively assess the clarity,edge density and texture characteristics of the image,and adjust the threshold value of SuperPoint accordingly to realize high-quality feature point extraction.Meanwhile,the matching threshold of SuperGlue was adjusted by evaluating the distribution entropy and spatial coverage of feature points to achieve efficient matching.In order to verify the effectiveness of the method,a comparison test was conducted with multiple image matching methods on the self-constructed cow face dataset.The results showed that:1)The macro average precision,macro average recall and macro average F1 of the algorithm on the Narrow-Baseline dataset were 92.1%,90.4%and 91.2%,respectively.2)The macro average precision,macro average recall and macro average F1 of the algorithm on the Wide-Baseline dataset were 82.8%,81.0%and 81.9%.3)The micro F1 score of the algorithm on the public dataset was 86.4%,and all the results were significantly better than the traditional image matching algorithm.In summary,this study aims to propose a new image matching algorithm,which effectively improves the adaptability of the cattle facial recognition model to non-standardized data,and provides technical references for the practical application of cattle facial recognition methods.

关 键 词：牛个体识别 特征匹配 动态阈值调整 面部识别 

分 类 号：S816[农业科学—饲料科学]