M Mining Images for Structure

Most data warehousing and mining involves storing and retrieving data either in numerical or symbolic form, varying from tables of numbers to text. However, when it comes to everyday images, sounds, and music, the problem turns out to be far more complex. The major problem with image data mining is not so much image storage, per se, but rather how to automatically index, extract, and retrieve image content (content-based retrieval [CBR]). Most current image data-mining technologies encode image content by means of image feature statistics such as color histograms, edge, texture, or shape densities. Two wellknown examples of CBR are IBM’s QBIC system used in the State Heritage Museum and PICASSO (Corridoni, Del Bimbo & Pala, 1999) used for the retrieval of paintings. More recently, there have been some developments in indexing and retrieving images based on the semantics, particularly in the context of multimedia, where, typically, there is a need to index voice and video (semantic-based retrieval [SBR]). Recent examples include the study by Lay and Guan (2004) on artistry-based retrieval of artworks and that of Benitez and Chang (2002) on combining semantic and perceptual information in multimedia retrieval for sporting events. However, this type of concept or semantics-based image indexing and retrieval requires new methods for encoding and matching images, based on how content is structured, and here we briefly review two approaches to this.