Face Information Structuring Based on Data Mining from Large-Scale Broadcast Video Archive

在这个项目中，我们实现了一种基于视觉信息和广播元数据的更高层次的人脸识别和结构化技术。该技术被应用于一个广播元数据种类繁多的大规模广播视频档案。在提出一种对光照条件、姿态和表情变化具有鲁棒性的人脸图像匹配算法的同时，我们将经典的人脸识别任务描述为一个定量预测问题，该问题估计伴随着网络路径的因果关系的事件概率。该解决方案以一种新的人脸分类方法的形式实现，该方法基于建立在广播元数据之上的贝叶斯网络，该方法以有效和高效的方式利用广播元数据和广播视频之间的相关性。

In this project, we achieved a higher-level facial recognition and structuration technique based on the integration of visual information and broadcasting metadata. The technique was applied to a large-scale broadcast video archive with broadcasting metadata of great variety. While proposing a facial image matching algorithm robust against the variation in illumination condition, pose, and expression, we formulated the classic facial recognition task into a quantitative prediction problem that estimates the event probability of cause-effect relationship accompanied by network path. The solution was achieved in the form of a novel facial classification approach based on a Bayesian network constructed on top of broadcasting metadata, which leverages the correlation between the broadcasting metadata and the broadcast video in a both effective and efficient manner.