Tex-Lex: Automated generation of texture lexicons using images from the world wide web

A method for automatic creation of a semantic texture database is introduced, which exploits the cumulative knowledge that exists in the image tags on the World Wide Web. In the first step of the method, a number of images are retrieved from the Web using the text search option provided by search engines by querying simple notions (e.g. sky, grass water, etc.). These images are segmented into a number of predefined regions using standard clustering and each region is described by a set of image features. The descriptors of the extracted regions of the whole set of images are compared based on the Bhattacharyya distance and the ones that are more similar are considered to be entries of a dictionary associated with the initial keyword used for the query. Moreover, the corresponding regions are parts of the visual lexicon describing the keyword. Also, an already existing lexicon may be iteratively updated by new features that may not match the existing dictionary entries but they are represented over a significant number of query results. Early results on common keywords representing landscapes indicate that the method is promising and may be extended to describe composite structures and objects.