Research On Vector Approximation Indexing Structure In Image Retrieval

Along with the digital and information age, a large number of multi-media information, including image data, is rapid growing. Image data are widely applied in weather, medical, transportation, military and so on fields. Meanwhile, how to retrieval the images with rapid and effective ways, and showing the required data to client become an urgency problem.In the traditional content-based image retrieval systems, because of the image features (color or texture, etc.) have high-dimensional nature, the one-dimensional data's index structure can't adapt to these data. For this reason, many algorithms have been proposed, such as tree-based index (KDB-tree, R-tree). However, we found the performance of those tree-based index structure sharply degraded as dimensionality increases, these algorithms even can't work as direct retrieval way when the data dimension above 10. This phenomenon is called as 'Dimension Curse'. For this cause, researchers have put forward many corresponding algorithms, such as VA-File, NB-tree, pyramid algorithm. However, these algorithms have their own shortcomings. For example, VA-File adopted vector approximation ways, and using the approximate vectors build the index file to filter high-dimensional data. It is one of algorithms which can solve the 'Dimension Curse' effectively. However, this method is proposed on the basis of independent data distribution, doesn't consider the relativity of the realistic data (this is the main problem of the pyramid algorithm). In addition, because of the approximate vectors' dimension doesn't be reduced, the actual amount which be calculated doesn't get effective decreasing, and thus affecting the efficiency of the index. The NB-tree method makes an effective decreasing of dimensions, but the way of decreasing dimensions did not take consideration of the distributing which the data situated in the space. As a result, this algorithm suffered from declining of the retrieval effect. In this paper, we proposed a new algorithm—'Target ways of Image Retrieval'. Through the experiment compared with other algorithms, it is convincingly proved that this index algorithm is efficient and applicable in CBIR.