A Study Of Lecture Video Annotation Algorithm

With the rapid development of Internet and multimedia technologies, Online learning based on online learning platform(or E-learning)has become an important alternative way of traditional classroom learning. Now, there are abundant lecture videos of academic lectures, reports and other teaching activities. The diversity and richness of video makes those lecture videos has become an important material for online learning. Lecture video annotation by analyzing low-level features, mapping of a low-level features and highlevel semantic in order to cross the semantic gap, and assigning to lecture videos with key words. These key words can describe video in content, and provide great convenience to massive videos indexing, retrieval, classification as well as recommendation. Meanwhile, it also can facilitate the interaction of software and the diversity and efficiency of learning.In this paper, automatic annotation of lecture videos is studied based on v-Lecture platform. The main works include lecture video structuration, keyframes subtitle extraction, text preprocessing, semantic shot segmentation, and video annotation based on machine learning methods.Firstly, this paper presents a semantic shot segmentation method based on LDA(Latent Dirichlet Allocation) combined subtitle text information extracted by OCR(Optical Character Recognition). By the way, if the video without subtitles, ASR(Automatic Speech Recognition) will be used instead of OCR. Then, using LDA to train the keyframes text information, comparing the distance of keyframes to realize semantic shot segmentation. The experimental results show that this method is capable of realizing a reasonable semantic shot segmentation due to introducing text information, but also has better segmentation result.Secondly, this paper proposes a video annotation method based on S4VM(Safe SemiSupervised Support Vector Machine). This method trains a small amount of labeled shots of subtitle message at S4 VM learning algorithm to complete the annotation for unlabeled video shot.it also can achieve annotation in video level. The experimental results show that the proposed algorithm can complete the annotation better and possesses a high accuracy rate.Finally, this paper proposes a video annotation method based on user collaborative filtering. Firstly, calculate the shot of term frequency inverse document value and select more important words to establish the shot-keyword matrix; secondly, compare the similarity between video shots to complete the annotation in shot-level; then achieve the annotation in video level. The experimental results show that the proposed algorithm can complete the annotation better and possesses a high accuracy rate.