| Summary: | With the rapid development in technology, digital news video has become increasingly popular due to the convenience it can bring to the viewer. However, the linear structure of the video often poses difficulty for the viewer in accessing particular news that is of his/her interest. To make the news video more accessible, one possible way is to segment the video into small parts that are classified by their individual genres. However, manual segmentation is costly, thus automatic news segmentation is highly desirable.
Current automatic news segmentation methods cover two major areas: topic boundary and topic detection. These techniques rely on common features mainly visual, audio and texts. However, some of the methods such as LDA are not suitable for broadcast news as the amount of texts is too little for topic detection.
In this thesis, we experiment with a hybrid-based method, which is suitable for broadcast news video. This approach uses shot change detection, speaker change detection and natural language processing such as noun phrase extraction to generate a well-defined topic boundary. To materialize these concepts, a web based broadcast news navigation (BNN) is used. BNN provides web based retrieval tools from the multimedia database which support user searching. As part of the testing phase, our experiment results indicate that the segmentation technique outperforms other state-of-art techniques while using the hybrid-based method.
|
|---|