Research About Semantic Marker Of Faces Extracted Form Video

For faces in the videos, if we want to get their semantic information, we can look up from the Web or books. But it will take a lot of time to get their information if we just by manual operation. If we can get their information automatically it will be meaningful.First, we have some faces whose semantic information are known, which is called prior library. We want to get semantic information of faces in the videos by prior library. But for each man in the prior library is little, so we can't get the true information just by a little faces in the prior library. Faces in the videos which want to be labeled are so much that we can use the faces in the video to be training samples. We combine Metric Learning and Active Learning to solve the problem of little samples, and propose a method so called Active Learning Online Method. The experiment is proved to be effective compared to Online Metric Learning. The main contributions in the thesis are as follows:First of all, we combine Metric Learning and Active Learning to propose a method so called Active Learning Strategy Based on Metric Learning. And we describe the detailed technology framework.Second, we study the four most popular Metric Learning Methods. And we experiment on the LFW databases using the four Metric Learning Methods. We show that when the training samples are small, the result of the methods is not effective. But when the training samples are sufficient, the four methods are effective.Third, we describe our Active Learning strategy amply, and we show its effectiveness through abundant experiments.Fourth, we experiment on the weak labeled face sequences by using our Active Learning strategy. Compared to online Metric Learning Method, our method can obtain higher labeled recognition rate. It shows the effectiveness of our method.