Semantic Concept Learning through Massive Internet Video Mining

摘要

Semantic concept learning is one of the most challenging problems in video retrieval. The key barrier for semantic concept learning is lack of annotated training data. Internet videos are different from ordinary videos: massive, rich information, customized, non-uniform format, uneven quality, little descriptive text, only a few shots with limited length etc. Therefore, Internet is a potential repository to provide a reliable source for concept learning. In this paper, we focus on the semantic concept learning through known Internet video sources mining. Starting from the video-sharing websites, an automatical graph model generator for concepts relationship learning based on known ontology such as LSCOM, WordNet and ConceptNet is discussed. An automated source discovery method is addressed which prove to be useful in concept detection from the massive Internet videos. Experimental results prove that the addressed method is effective and efficient in semantic concept detection and learning through massive Internet video mining.