| Nowadays, with the rapid development of data compression, communication,storage technology and the continuously improve most of computer performance, therecomes extensive applications of multimedia information. Especially, news video hasbecome an important source of information. Facing so much mass information, how toobtain people needed information accurately become a key research issue in multimediafield. Video summary and video retrieval are the main method to solve these problems,and video structure analysis is one of the important basic work.A new framework for news video structure analysis by using multi-modal features,such as visual features, audio features and theme caption, is proposed in this thesis. Onthe basis of news video special structure, we design the framework to segment newsvideo into a series story units. Firstly, news video is divided into audio stream and videostream. Secondly, For audio stream, mute intervals are detected and choose as audiocandidate points; for video stream, the shot segmentation for news video are detectedand shot boundary points are chosen as video candidate points, then anchorperson shotand topic caption are detected and extraited. Finally, story units are detected andextraited by fusing audio candidate points, video candidate points, anchorperson shotand topic caption based on time axis.Experimented result on a mid-size news video from CCTV shows that theframework is effective and gets83.92%in precision and83.18%in recall for newsvideo story segmentation. |
PDF Full Text Request