Image Classification by Image Matching

Abstract

Image classification is usually performed by making measurements on the image itself. For example, texture measurements are made to classify an image as “natural” or “urban.” This chapter describes an experiment that we carried out to determine whether an image could be classified by matching the image against databases of images, where each database belongs to a different class. Because we wanted a classification problem that could not be solved in a reasonable time by making measurements on the image, we selected the classes of interest as “benign images” and “objectionable images.” This classification problem is an extremely difficult one because although people can usually make the classification, it is almost impossible to formally define “benign” and “objectionable.”

It pays to have a healthy skepticism about the miracles of modern technology. — Dennis A. Hejhal (1948-)¹, commenting on the Internet

Hejhal is a professor of mathematics and a fellow in the Supercomputing Institute, at the University of Minnesota.