Structure tensor series-based large scale near-duplicate video retrieval

With the huge amount of video data and its exponential growth in recent years, many new challenges, like storage, search and navigation, have arisen. Among these challenges, near-duplicate video retrieval aims to find clips that are identical or nearly identical in content to a query clip. This has attracted much attention due to its wide applications including copyright detection, commercial monitoring and news video tracking. In this paper, we propose a practical solution based on 3-D structure tensor model for this problem. We first propose a novel video representation, adaptive structure video tensor series, together with a robust similarity measure, to improve the retrieval effectiveness. Then, we design a dimensionality reduction technique for tensor series to improve the search efficiency. Finally, we prove the effectiveness and efficiency of the proposed method by extensive experiments on hundreds of hours of real video data.