Understanding user intention in image retrieval: Generalization selection using multiple concept hierarchies

Abstract

Image retrieval is the technique that helps users to find and retrieve desired images from a huge image database. The user has firstly to formulate a query that expresses his/her needs. This query may appear in textual form as in semantic retrieval (SR), in visual example form as in query by visual example (QBVE), or as a combination of these two forms named query by semantic example (QBSE). The focus of this paper lies in the techniques of analyzing queries composed of multiple semantic examples. This is a very challenging task due to the different interpretations that can be drawn from the same query. To solve such a problem, we introduce a model based on Bayesian generalization. In cognitive science, Bayesian generalization, which is the base of most works in literature, is a method that tries to find, in one hierarchy of concepts, the parent concept of a given set of concepts. In addition and instead of using one single concept hierarchy, we propose a generalization so it can be used with multiple hierarchies where each one has a different semantic context and contains several abstraction levels. Our method consists in finding the optimal generalization by, firstly, determining the appropriate concept hierarchy, and then determining the appropriate level of generalization. Experimental evaluations demonstrate that our method, which uses multiple hierarchies, yields better results than those using only one single hierarchy.