集成学习和蚁群算法优化XGBoost的人脸检索及应用  

Ensemble Learning and Ant Colony Parameter Optimization of XGBoost for Face Retrieval and Applications

作　　者：张涛 高宇航 陈永俊 张家宝 郭红涛[3] ZHANG Tao;GAO Yu-hang;CHEN Yong-jun;ZHANG Jia-bao;GUO Hong-tao(Department of Big Data Center,Nanjing Public Security Bureau,Nanjing 210005,China;Nanjing Skytech Co.,Ltd,Nanjing 211800,China;People's Public Secutity University of China,Beijing 100038,China)

机构地区：[1]南京市公安局大数据中心,江苏南京210005 [2]南京擎天科技有限公司,江苏南京211800 [3]中国人民公安大学,北京100038

出　　处：《中国电子科学研究院学报》2023年第11期1021-1028,共8页Journal of China Academy of Electronics and Information Technology

基　　金：江苏省重点研发计划资助项目(BE2017616);南京市公安局科技信息化“十四五”重大资助项目(NJZC-2021GK0029)。

摘　　要：无约束人脸检索场景中,为了进一步提高识别精度,实践中往往会引入多个主流的深度学习人脸识别模型。针对每次从多个模型检索结果中确认人脸身份的工作量较大问题,以及考虑支持异质模型的插件式更换,使用基于Blending集成学习的人脸检索,将多个人脸识别模型作为基模型,并提出了一种基于蚁群算法优化的XGBoost作为元学习器对基模型的人脸相似度预测结果进行融合。实验结果表明:集成模型相对个体模型的检索性能均有不同程度提升,其中蚁群优化的XGBoost作为元模型的集成模型检索性能优于多项式回归、随机森林和GBDT作为元模型的性能,验证了集成模型和优化后的XGBoost的有效性。Unconstrained face retrieval scenarios often require the introduction of multiple mainstream deep learning face recognition models in order to further improve recognition accuracy.Face retrieval based on Blending ensemble learning is used for the problem of high workload in confirming face identity from multiple model retrieval results each time and for considering plug-in replacement that supports heterogeneous models.Multiple face recognition models are used as base models,and an ant colony algorithm optimization based XGBoost is proposed as a meta-learner to fuse the face similarity prediction results of the base models.The experimental results show that the retrieval performance of the ensemble model relative to the individual models are all improved to different degrees,in which the retrieval performance of the ensemble model with ant colony optimized XGBoost as the metamodel outperforms that of polynomial regression,random forest and GBDT as the metamodel,verifying the effectiveness of the ensemble model and the optimized XGBoost.

关 键 词：Blending集成学习 XGBoost 蚁群算法 预测结果融合 人脸检索 

分 类 号：TN929[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]