Image similarity detection in large visual data bases

摘要

A method of similarity clusters detection in large visual databases is described in this work. Similarity clusters have been defined on the basis of a general concept of similarity measure. The method is based also on the properties of morphological spectra as a tool for image presentation. In the proposed method similarity of selected spectral components in selected basic windows are used to similarity of images evaluation. Similarity clusters are detected in an iterative process in which non-perspective subsets of images are step-by-step removed from considerations. In the method similarity graphs and hyper-graphs also play an auxiliary role. The method is illustrated by an example of a collection of medical images in which similarity clusters have been detected.