A multimodal approach for extracting content descriptive metadata from lecture videos

摘要

The rapidly increasing availability of e-learning content and lecture videos over the internet, has brought forth an imperative need for developing effective content based retrieval systems. Comprehensive metadata extraction and support for topic-level search within videos are key factors in developing such systems. In this paper, we propose a multimodal metadata extraction system which extracts an optimal set of keyphrases and topic based segments that effectively summarize the content of a lecture video. The extraction process utilizes features from both audio transcripts and slide content in video streams. A hybrid approach combining a Naive Bayes classifier and a rule-based refiner is used for effective retrieval of the metadata in a lecture. The proposed content-descriptive metadata extraction technique has been evaluated using actual lecture videos from different sources, and our results show that our multimodal approach is effective in summarizing the lecture’s content, potentially improving the user experience during retrieval and browsing.