Neural network based object recognition in images

Abstract

An investigation of object recognition in images is described. It is based on the following idea: the image is viewed not unlike the transparency obtained by overlaying several transparencies, each containing a single object. The view is taken that any image is a composition of several atomic images. The atomic images contain only one object and they have the same size as composite images. It is shown that the neural networks trained on a small set of atomic images can recognize a very large set of all possible composite images, including overlapping objects, with reasonable recognition rates. Also briefly discussed is the research prototype of the postrelational database management system CHINOOK being developed at the University of Colorado. CHINOOK is intended to manage a database of digitized images and digitized one-dimensional data, as well as text and tables.< >