الفهرس 
News video clustering and annotation
ContentsOnly 14 pages are availabe for public view
 |  | from | 96 |  |  |
| | | from | 96 | | |
Abstract
In order to enable the users of news videos to gather the maximum amount of information contained in the news video in minimum time, we cluster the news video frames into shots using abrupt cut detection, summarize news video shots, and extract captions contained in the news video as annotation clues. A method to detect and localize all caption types in Arabic news videos is proposed. Moreover, different types of captions are considered including static, horizontal scrolling and vertical scrolling captions. Our method is able to deal with different patterns of appearance and disappearance of captions in news video. Also it can deal with news videos with multiple captions. The proposed method is based on edge feature and multiple frames integration. Canny edge map is computed for each frame. Horizontal lines detection is applied and frames are categorized into clusters. Finally, caption types are recognized from each cluster by observing the normalized inter-frame edge map difference. A new representation of images is proposed. We called that representation as 2dominant colors3. The dissimilarity of two images is defined as a vector contains the difference in order of each dominant color between the two images representations. The new image representation and dissimilarity measure are utilized to detect the abrupt cuts in news videos. A neural network is trained with the new dissimilarity measure to differentiate between two classes of news videos frames: cut frames, and non-cut frames. In addition, a key frame extraction method is proposed. The proposed method takes a confidence level as input from the user to satisfy the different needs