基于聚类的人脸图像检索及相关反馈

提出了一种基于聚类的人脸图像检索算法. 首先利用归一化分割(Normalized cuts, NCuts)在每个时间段内分别对人脸聚类, 使同一个人在不同情况下的人脸图像聚为一类. 其次采用连续AdaBoost算法学习得到的人脸识别分类器度量人脸之间的相似度, 并进一步提出查询人脸与人脸聚类之间的相似度用于检索. 为了进一步提高性能, 用户可以在线标定错检和漏检的结果, 相关反馈环节把用户的交互标定结果作为约束条件重新对人脸聚类. 本文把人脸图像检索算法应用于自动的检索系统中, 在包含超过一千张人脸图像的家庭数码相册上, 通过与其他方法的对比实验证明了基于聚类的人脸图像检索算法是有效的.

Cluster-based Face Image Retrieval and Its Relevance Feedback

This paper proposes a novel cluster-based face image retrieval algorithm.By using normalized cuts(NCuts)to cluster faces in each time span into optimal partitions,various face images of the same character can be grouped together. A face recognition classifier learned by real AdaBoost is used for measuring similarity between two faces,and a similarity measure for the query face and the face cluster is further proposed for retrieval.To further improve the performance,we design an online step,in which users can interactively label false positives and missing retrieval,so that some constraints are involved to revise the clustering results.The algorithm is integrated into an automatic retrieval system,and the experiment on a family album containing over a thousand face images confirms its effectiveness in comparison to other alternative algorithms.