Similarity-based Ordering of Images in Search Engines

Search engines such as Google and Yahoo allow internet users to search through huge collections of images. These existing search engines are all text-based, i.e. the searching is based on the comparison of keywords with textual annotations. Content-based image retrieval (CBIR) is a more natural approach in which the searching is based on the actual content of the images. However, CBIR systems are still not capable of searching through collections consisting of millions of images. Therefore they are not suitable yet as a basis for a search engine on the internet. As a solution to that problem, we introduce a hybrid approach that combines text-based image retrieval (TBIR) with CBIR, without sacrificing the scalability of the system. More precisely we propose an extension of TBIR that adds the ability to sort the search results according to their similarity with a given set of examples. Such an extension requires measures that are able to express the similarity between two images. Because similarity is a gradual concept, we use fuzzy set theory to model these measures. Keywords—similarity measures, colour images, fuzzy sets, content-based image retrieval