MIF: Toward Semantic-Aware Representation for Video Retrieval

Semantic Concept-Based Video Retrieval (SCBVR) has been widely studied recently, which exploring semantic representations of videos to execute user’s retrieval requests. However, the effectiveness of video retrieval often depends on the accuracy of semantic concepts, but the semantic concept is often imprecise or contrary to the ground-truth. In order to solve the above-mentioned problem, we propose a novel multi-information fusion (MIF) approach, and it is beneficial to improving the performance of video retrieval. Firstly, we infer the most relevant semantic concept corresponding to query keywords by using the inherent association information of videos. Secondly, we fuse the probabilities of the candidate semantic concept by minimizing the potential function. We conduct the extensive experiments on real-world datasets which demonstrate the effectiveness and efficiency of the proposed approach for enhancing the performance of video retrieval.