Fast Image Search by Trees of Keypoint Descriptors

Abstract

In this paper, we present a method for fast image searching tree-based representation of local interest point descriptors. Thanks to decreased number of steps needed to perform the search, such a representation of image keypoints is more efficient than the standard, frequently used list representation where images are compared in all-to-all manner. The proposed method generates a tree structure from a set of image descriptors, e.g., generated by the SURF algorithm. The descriptors are stored as leaves in the tree structure and other parent tree nodes are used to group similar descriptors. Each next parent node of the tree forms a wider, more general, group of descriptors. We store average values of the descriptor components in the nodes making it possible to quickly compare sets of descriptors by traversing the tree from the root to a leaf by choosing the smallest deviation between searched descriptor and values of nodes. With each next step of tree traversing we reduce the final number of descriptors that will be needed to compare. The proposed structure also allows to compare whole trees of descriptors what can speed up the process of images comparison, as it involves generating trees of descriptors for single images or for groups of related images accelerating the process of searching for similarities among others.