Similarity-based image organization and browsing

Abstract

Users do not always know what they want, in which case traditional query-based image retrieval approaches fail. This thesis serves to amend this shortcoming with a novel approach to organize and browse large image collections based on visual similarities in a way that extends the user's natural search sense. Starting with an unordered set of images, salient color and gradient information are extracted into feature vectors. A self organizing map (SOM) then projects these high-dimensional vectors onto a 2D canvas so that similar ones are grouped together. When browsing around on the canvas through intuitive operations like pan and zoom, a dynamic collage is generated that shows the most pertinent images. To make organizing larger image collections practical, a parallel SOM training algorithm is designed that runs on graphics processing units. The results of using a variety feature vectors are also evaluated.