Research On Extracting Semantic Features For Videos

With rapid development of computer network and the storage technology, video and other multimedia data are into a geometric growth. How to retrieval useful data from a broad array of resources in the video data is becoming a issue of concern.Most video retrieval techniques are low-level features based and no-semantic. These features are abstract and quite different from the semantic concepts in human thought and also affect the retrieval results from video data. It is hoped that the retrieval based on semantic can be used to instead of retrieval based on visual content.As the complexity, ambiguity and subjectivity of video semantic content, there exists the "semantic gap" between low-level visual features and high-level semantic. How to beyond the "semantic gap" is still not fully resolved. The experiments show that if we could extract the high-level human cognitive semantics from the video, which describing video information as retrieval feature, we will find an effective way to solve "semantic gap" between low-level features and the high-level semantics. There are two main ways to extract semantics from video: rules-based approach and statistical theory-based method. Rule-based method is using the fields of knowledge to define the perception rules which is to test the semantic concepts in video. Because of its dependence for the strong domain knowledge, it is difficult to extend to other areas. Whereas statistical theory-based method is mainly about probability and statistics learning theory, through training samples to find their semantics' probability relations. It is an effective remedy from low-level characteristics to a higher-level semantics.In this paper, Method of based on the theory of statistical is the main research. We hope to extract cognitive semantics from the video information, namely, semantic tagging for video key frame in order to try to cross "semantic gap." Therefore, the paper starts from the reasons of "semantic gap" and proceeding on the current solution to this problem. Then introduce a number of classical algorithms such as K approaching, naive Bayesian, Gauss kernel and support vector machines aspect to algorithm principles, steps and experimental process. On this basis, we put forward a new algorithm called SID algorithm which is considering the importance of semantic. It focuses on the SID algorithm background, purpose and experiment. Finally, we conclusion that through experiments SID algorithm is superior to the above-mentioned other algorithms and has made better result in semantic extraction from video data.