Automated summarization of broadcast news video

Abstract

Nowadays, time is a luxury not everyone can afford. Many people despise having to consume an entire news bulletin when only a few stories interest them. The solution in this project creates a summarized version of news in the form of an informative index allowing individual story playback, such that the user can randomly consume individual stories. As highlighted in the literature, complications arise due to inconsistent news structure. To combat this, many approaches have been developed to index news videos using the visual cues, audio cues, textual cues and mathematical models. The solution developed in this project (using an existing framework by Prof. Ing. V. Buttigieg, capable of video search) is specific to a local station (TVM), and uses the general structure to identify the eight o'clock news and extract individual news stories. A video searching algorithm is used to identify scenes with a news anchor in order to detect new stories. For each story detected, subsequent thumbnails are displayed to provide insight on the content of the story using a shot change detector. Whilst training the system, recognition results range from 88.9% to 73.3%, which is relatively good considering the simplicity of the algorithm principle. The algorithm success rate ranged between 89.5% and 73.3%. The results are slightly higher since in theory, the current algorithm is restricted to news anchor detection rather than actual story detection, and thus, it can never be fully successful regardless how well the algorithm performs. Recognition for a proper test video decreased to 41.7%. Nevertheless, the results highlighted possible improvements which will significantly improve results. Particularly for this video, it is safe to assume that another 50% of the news stories can be detected simply by searching for yet another image representing a common news anchor scene. Thus, further training is necessary.