Abstract: Video-based face retrieval has become one of the hottest topics in the domain of face recognition and retrieval. A face retrieval algorithm based on affine hull and pseudo-zernike moments in video （FRIVAP） is proposed. The sets of images of faces are constructed by detecting and tracking human faces in videos, in which the pseudo-zernike moments of faces are extracted, and the affine hulls for the features in image sets are set up, the same or similar faces are retrieved by calculating the similarity. Experimental results on Honda/UCSD and self-constructed video databases show that the proposed algorithm fully utilizes both spatial and temporal information of faces in video, and robust under noise, pose variations.

Key words: face retrieval in video, face recognition, affine hull, pseudo-zernike moments

摘要： 近年来基于视频的人脸检索已成为人脸识别和检索领域最为活跃的研究方向之一。提出了一种基于仿射包结合伪Zernike矩特征的视频人脸检索算法（FRIVAP）。在视频中检测跟踪到人脸生成图像集，接着提取图像集中人脸的伪Zernike矩特征，建立特征的仿射包，通过相似度计算得到结果。经对Honda/UCSD视频数据库和自行构建的视频数据库的大量实验表明，该算法可以充分利用视频中人脸的时间和空间信息，并且对噪声、人脸姿势变化等条件下的人脸检索有较强的鲁棒性。

关键词: 视频人脸检索, 人脸识别, 仿射包, 伪Zernike矩