In image search field, many local feature based algorithms do very good jobs but they are not quite applicable to the real world applications interms of big image sets. Some algorithms such as SIFT and SURF extract many features and makes their searches using those many features. But that becomes an obstacle when searching in big datasets. In addition to features obtained from "SURF" and "SIFT" algorithm, for the purpose of increasing complexity of problem and increasing accuracy level in search result we use a method whats called "Localized Global Features"[4]. In this method we apply global features along the windows sizes of SURF Key points. We also change the size of the window and see that those new features diffrent from default size of the window. By the help of optimization we find optimum window size and apply global features on that areas and get feature what is we called "Extended Localized Global Features" By the help of clustering we reduced multi-dimensional features into one-dimensional search problem, and by the help of the search power of tree algorithms in onedimensional data, the descibed method reduced the time it takes, even in large datasets, to very short amounts. By expanding number of different local descriptors we achieve very good result both in search time and search results. On the other hand despite the big decrease in search time we did not see much precision recall ratio decrease in this approach.
展开▼