Multi-view News Video Topic Tracking Approach

Existing researches on tracking topics of news videos require lots of labeled examples. However, video labeling is too time-consuming to generate a large number of labeled videos in real applications. In this paper, a novel approach is proposed to track news video topics through using only a few labeled samples. The three main characters of proposed approach are: (1) Multi-view learning process is used for the lack of labeled training samples existing in the topic tracking processing. (2) Principal Component Analysis is used to simplify the feature vectors and speed up processing. And (3) a collaborative training process is proposed to track the related news videos with only one labeled sample. Experiment conducted on our collected Chinese news videos show that the proposed approach can get a good performance with 82.5% precision and 96.4% recall.