Research On Video Semantic Extraction

With the rapid development of multimedia and network, the number of digital videohas increased dramatically. How to understand these large amounts of video automaticallyby the computer and extract its semantic information so that users can query and retrievethem efficiently, is the urgent problem to resolve currently. At present, video semanticextraction technology has got widely attention and become hot research.Tag is an effective way to express semantic, according to the characteristics of thevideo data, the noun semantic tag which corresponds to the motion object in the video andthe verb semantic tag which corresponds to the specific action of the motion object can beused as a combination to express the semantic information. The noun semantic tag and theverb semantic tag can be combined to obtain the video semantic tag.To extract video semantic tag, machine learning method can be used, construct thevideo training set first, and then analyze the test video from two aspects which are nounsemantic tag and verb semantic tag. In the aspect of noun semantic tag, do shotsegmentation and key frame extraction to get the key frames first, and then do objectrecognition to get noun semantic tags. In the aspect of verb semantic tag, get the trajectorydata through motion object detection and tracking algorithm first, and then compute thesemantic similarity between the trajectory data of the test video and the trajectory data ofthe training set，contact the noun semantic tag of test video and the annotation of thetraining set, analyze all of the above information comprehensively to get the verb semantictag. Finally, the semantic information of the video can be obtained by following after theChinese subject-predicate structure and using the noun semantic tag to express the subjectand using the verb semantic tag to express the predicate.