Application Of Semi-Supervised Learning Algorithm Based On Kernel Density In Video Semantic Annotation

Since the 1990s, with the fast explosion of multi-media information, Content-Based Video Retrieval has become a research hotspot. How to use the Machine Learning theory to make the computer get the semantic information of video automatically and use them to retrieve videos efficiently has become an urgent problem in the research of multi-media. So, how to do video annotation efficiently is the purpose of this paper.At present, most of video semantic annotation methods are based on statistic theory. The methods use supervised learning method to do semantic label. These methods take more than enough labeling samples for study, so as to establish classifier to label unknown samples. As we all know, obtaining a lot of labeling samples is a time-consuming and laborious work, but obtaining large number of unlabeled samples is a very easy task. So, the semi-supervised learning method by learning a small number of labeling samples and a large number of samples to establish classifier came into being. If the ultimate effect of the semi-supervised method is the same or close to the result of supervised learning method, the semi-supervised learning is more advantages in labor costs and achievement. How to make the effect of semi-supervised learning closed to or the same to the supervised learning by unlabeled samples information is the key of semi-supervised learning method.In this paper, we discuss the problem of using semi-supervised learning method to do video semantic annotation. In the paper, we use the statistical theory to calculate the probability of video semantics by Bayesian formula, choose the semantic of maximal probability to label the unlabeled samples. In the process, we calculate the posterior probability of semantics by unlabeled samples information. By adding the impact factors of unlabeled samples information in the Bayesian probability formula, we introduce the semi-supervised learning method into video semantic annotation work, combing the labeling and unlabeled samples information to calculate the semantics probability.The kernel estimation method analyzes the data distribution by not using the prior knowledge of data distribution. This method avoids the impaction of model and parameters estimation. To make the video semantic more accurate and more effective, to achieve the purpose of this paper, the semi-supervised learning method is based on the Gaussian kernel in this paper. The method uses the optimal factor of unlabeled samples, combing the superiority of semi-supervised learning and kernel density estimation method, to label unlabeled video samples. In the paper, we design an analysis system to analyze the performance of this method. By the result of a lot of experiments, we find that the semi-supervised learning method we discuss in the paper could get the performance which is similar to or equal to the supervised learning method. It will reduce the manual working, improve the performance of video semantic labeling, is conductive to large-scale video semantic labeling work.