Organizing image databases as visual-content search trees

摘要

An unsupervised algorithm for arranging an image database as a visual-content binary search tree is described. Tree nodes are associated with image subsets, maintaining the property that the similarity among the images associated with the children of a node is higher than the similarity among the images associated with the parent node. Visual-content search trees can be used to automate image retrieval, and help a human to interactively search for images. Experiments with datasets of hundreds and thousands of images show that shallow trees produce clustering into `meaningful' classes.