基于 Chinese Whispers 的人脸动态聚类-《计算机技术与发展》

文章信息/Info

作者:: 张陈欢¹; 2 ; 史燕中³; 1. 北京航天长峰科技工业集团有限公司,北京 100039; 2. 中国航天科工集团第二研究院,北京 100039; 3. 北京航天长峰股份有限公司,北京 100039

Author(s):: ZHANG Chen-huan ¹; 2 ; SHI Yan-zhong ³; ZHANG Chen-huan 1,2 ,SHI Yan-zhong 3

Keywords:: Chinese Whispers; dynamic clustering; face cluster; representative point; data mining

摘要:: 针对 Chinese Whispers 算法对于小规模数据聚类随机性大,对于大规模数据聚类速度缓慢的问题,提出了一种改进的 Chinese Whispers 算法用于人脸的动态聚类。新增了一个阈值 P ,用Chinese Whispers 算法对数据规模为 P 的数据进行聚类时,既可以保证聚类结果的稳定性,又可以保证聚类算法的高效性;利用代表点而不是所有点完成聚类更新,能够有效减少对增量数据聚类时的数据量,从而达到提升聚类速度的目的。采用 CNN+ArcFace Loss 方法提取人脸特征,采用余弦距离作为相似性度量的方式,采用类中心作为代表点来描述类别信息,采用增量聚类的算法架构实现对于大规模数据的人脸动态聚类,并完成在 LFW、VGGFace2 和 CASIA-Webface 三个公开人脸数据集的测试。实验结果表明,基于Chinese Whispers 人脸动态聚类算法可有效提高聚类的时间效率,时间复杂性由原来的 O(n 2 ) 变为 O(n*p) 。

Abstract:: Aiming at the problem of large randomness of small-scale data clustering and slow speed of large-scale data clustering,an improved Chinese Whispers algorithm is proposed for dynamic face clustering. A threshold P is added. When using Chinese Whispers agorithm to cluster data with a data size of P,it can not only guarantee the stability of clustering,but also ensure the efficiency of the clustering algorithm. Using representative points instead of all points to complete clustering update can effectively reduce the amount of data in incremental data clustering,so as to improve clustering speed. The CNN+ArcFace Loss method is used to extract facial features, and the cosine distance is used as the similarity measure. The class center is used as the representative point to describe the category information,and the incremental clustering algorithm architecture to realize the face dynamics for large-scale data. Test is completed in three public face data sets of LFW,VGGFace2 and CASIA-Webface. The experiment shows that the Chinese Whispers algorithm can effectively improve the time efficiency of clustering,and the time complexity is changed from O(n2) to O(n*p) .