On Research And Prototype Implementation Of Video Semantic Annotation

In recent years, with the rapid development of information technology, video applications service rapid development. Digital video resources both from traditional television media, including the Internet video portal. How massive video effective organization, management, and the use of video content-based retrieval techniques CBVR (Content-based Video Retrieval) came into being. The content retrieval major based on low-level visual features of video, static characteristics, including color, texture, shape, and understanding of the people on the video description is based on the semantic level. Video semantic complexity, vagueness and subjectivity make the "semantic gap" between low-level features and high-level semantics, this problem has not been a good solution. Therefore, to provide users with a video annotation system consistent with human perception, so that the low-level features and image Advanced semantic Contact associated more closely, narrowing the "semantic gap" problem of low-level features and high-level semantics, a video retrieval research hotspot.The automatic video semantic annotation method based on machine learning theory, to adopt the annotation method Authority, with some success. However, to obtain a large number of annotated video samples need to pay a lot of artificial contrast, unlabeled video samples are very easy to get. Therefore, a small amount of labeled samples and a large number of unlabeled samples for classification of semi-supervised learning method known as the focus of the study. Semi-supervised learning process, and how to effectively use the unlabeled sample potential is called the key.This paper discusses the theory and practice. Video analysis, feature extraction, video annotation system based on machine learning techniques to achieve high-level semantic concept mapping technology discussed. Standard from the MPEG-7descriptors perspective, video low-level features of color, texture and shape feature extraction methods. Based on the semi-supervised learning method, machine learning, semi-supervised learning for semantic annotation. Tri-SVM collaborative training algorithm for LS-SVM algorithm. The experiments show that, compared to the Tri-SVM method since the training methods and the Co-Training methods have greatly increased in the classification error rate.This paper constructs semantic annotation framework for the design of the prototype system, the LS-SVM Tri-Training algorithm efficiency and sample requirements provide the feasibility of the application. System with manually labeled modules, training modules, automatic annotation module. The prototype system for the practical application of the automatic semantic annotation system reference.