基于子空间增量学习的视频中人脸图像检索

提出一种以电影视频中人脸图像为依据的视频检索方法.首先通过AdaBoost检测视频序列中的人脸图像,将检测到的人脸做标准化处理后投影到增量特征人脸子空间中,得到人脸图像的向量表述;然后应用单类支持向量机进行训练和分类,根据分类的结果动态地调整前面得到的最优分类超平面,实现对电影视频中特定演员的检索功能.由于不同镜头中同一人的人脸图像通常差别很大,该方法随时间序列动态地调整特征人脸空间,以适应人脸分布的变化.对电影《小花》、《Notting hill》等的实验表明,该方法在视频环境下可以较准确地检索出特定人像.

A Subspace Incremental Learning Method for Face Retrieval In Feature-Length Films

This paper presents a novel method for face retrieval in feature-length films. At first, the AdaBoost is used to detect human faces in the video, then the detected faces are normalized and projected to the incremental eigenspace to obtain their vector representation. After that, the one-class support vector machine is trained for classification and the optimal classification plane is dynamically adjusted to account for different actors in the film. Experiments on a famous Chinese film "little flower" and an Oscar prize film "Notting hill" show the effectiveness of our proposed method.