Research Of News Event Retrieval Based On Caption

News video is one of the most important media for users to get the information which are related to the people's daily life. It is an urgent demand for managing and retrieving the interested news from a huge amount of news video data efficiently and correctly. In news video, the captions provide the key for the related news story, and they can be used for the effective retrieval. This thesis proposes a caption-based news video retrieval technique, and develops a prototype system for news event retrieval. The prototyped system consists of two modules:①news event segmentation and (2) news event retrieval.The news event segmentation module is to partition the news video into several individual news stories by detecting special events (such as text event, anchorperson shot). An improved algorithm for text event detection is proposed and an anchorperson shot detection algorithm is developed in this thesis. Experimental results demonstrate that the accuracy of the text event detection algorithm is achieved, and the anchorperson shot detection algorithm is effective. The news event retrieval module is to locate and extract the caption in the news event. Then, the recognized text information is stored in the database as index for user to retrieve the related news. In this thesis edge detection and connected region labeled methods are used to locate the caption, and then the threshold segmentation method is used to extract the caption. Finally, the Chinese optical character recognition system developed by Han Wang Company is employed for the segmented characters recognition. The precision of the caption detection is higher than 85%, while the recognition rate is not high enough. The algorithm of caption extraction needs to be improved in the future research.