Hyponymy-Based Peculiar Image Retrieval

Abstract

Most researches on Image Retrieval (IR) have aimed at clearing away noisy images and enabling users to retrieve only acceptable images for a target object specified by its objectname. We have become able to get enough acceptable images of a target object just by submitting its object-name to a conventional keyword-based Web image search engine. However, because the search results rarely include its uncommon images, we can often get only its common (maybe similar) images and cannot easily get exhaustive knowledge about its appearance (look and feel). As next steps of IR, it is very important to discriminate between “Typical Images” and “Peculiar Images” in the acceptable images, and moreover, to collect many different kinds of peculiar images exhaustively. This paper proposes novel methods to retrieve peculiar images from the Web by expanding or modifying a target object-name (as an original query) with its hyponyms, which are based on hand-made concept hierarchies such as WordNet and Wikipedia, or which are extracted from the Web by text mining techniques, and validates their precision by comparing with Google Image Search.