| An important research in the field of multimedia databases is the detection of similar multimedia objects such as images, sound and videos. The dimensionality of feature vectors extracted from the multimedia objects is often very high, when traditional indexing methods face the difficulty of'curse of dimensionality'. For the efficiency of query in high-dimensional space, a new index structure is presented to speed up the similarity query of multimedia.In this paper, we analyze the development of high-dimensional indexing technology and characteristics of various high-dimensional index structures, especially one-dimensional mapping method. For k-nearest neighbor (KNN) search in the high-dimensional vector space, a new one-dimensional mapping approach based on Principal Component and distance, named pcDistance, is proposed. The new approach partitions the dataset, and selects a reference point for each partition. Then Principal Component Analysis is performed on each partition in the dataset, and the data points are transformed into a single dimensional value based on their distance to the reference point. And we use B+-Tree to index the single dimensional values. When performing k-nearest neighbor search, points can be pruned according to the relationship between the query and the axis determined by the first principal component, and partial distortion searching algorithm can also be used to speed up the query processing. Experiment results show that the new approach can reduce the number of data points accessed during the query, and can provide a faster search speed than existing high-dimensional indexing methods. Finally, we apply pcDistance to the Video Copy Detection. In this system, pcDistance is employed to manage and index the compacted video feature, BCS (Bounded Coordinate System). Our extensive performance study on a large database of more than 50,000 video clips demonstrates that pcDistance can improve the performance in detecting copy video. |
PDF Full Text Request