基于子空间聚类的视频人脸数据自动标注

Automatic annotation of face data from videos based on subspace clustering

针对人脸数据标注所需的人工和时间成本巨大,标注出的人脸数据集含有较多噪声问题,提出一种基于子空间聚类的视频人脸数据自动标注方法。首先,将海量视频作为人脸数据的采集来源,以满足多种人脸识别任务中不同的人脸数据需求,然后使用人脸识别模型将人脸数据映射到特征空间,使用改进K近邻算法把人脸数据划分到不同的子特征空间,最后在每个子特征空间内使用K均值算法分离人脸数据中的正样本、难正样本与负样本,收集难正样本构建人脸数据集。实验在公开数据集LFW与真实待标注数据上进行,实验结果表明子空间聚类法的F1度量得分比传统聚类算法分别提高了10%和7%,数据标注速度达到传统人工标注的10倍。使用该方法建立了一个包含200个ID、9 500张人脸照片的模糊人脸数据集,可用于多种人脸识别问题的研究。

In order to solve the problems of the huge labor and time costs on data annotation,lots of noises in annotated face data,we propose an automatic annotation method of face data from videos based on subspace clustering.Firstly,a huge number of videos are utilized as the source of face data,offering different kinds of data for different face recognition tasks.Secondly,map face data into feature space by using face recognition model,and divide feature space into subspaces by improved K-nearest neighbor algorithm.Finally,K-means clustering algorithm is utilized in every subspace to divide face data into positive samples,hard positive samples and negative samples,then hard positive samples are collected for data annotation.Experiments are conducted on LFW dataset and face data from videos without annotation.The results show that the proposed method gains 10%and 7%higher F1 measure score than traditional clustering algorithm and the proposed method makes face data annotation 10 times faster than manual annotation.In addition,a blur face dataset containing 200 ID with 9 500 faces is built by the proposed method and can be used for many face recognition tasks.