On the Role of Context-Awareness in Binary Image Comparison

Abstract

The quantification of image similarity has been a key topic in the computer vision literature for the past few years. Different mathematical theories have been used in the development of these measures, which we will refer to as comparison measures. An interesting aspect in the study of comparison measures is the natural requirement to replicate human behavior. In almost all cases, it is appropriate for a comparison measure to produce results that are consistent with how humans would perform that assessment. However, despite accepting this premise, most of the proposals in the literature ignore a fundamental characteristic of the way in which humans carry out this evaluation: the context of comparison. In this work we present a comparison measure for binary images that incorporates the context of comparison; more precisely, we introduce an approach for the generation of ultrametrics for the context-aware comparison of binary images.