Computational strategies for object recognition

This article reviews the available methods for automated identification of objects in digital images and proposes the simplest strategies, which work on data appropriate for feature vector classification, and methods that match models to symbolic data structures for situations involving reliable data and complex models.

Abstract

This article reviews the available methods for automated identification of objects in digital images. The techniques are classified into groups according to the nature of the computational strategy used. Four classes are proposed: (1) the simplest strategies, which work on data appropriate for feature vector classification, (2) methods that match models to symbolic data structures for situations involving reliable data and complex models, (3) approaches that fit models to the photometry and are appropriate for noisy data and simple models, and (4) combinations of these strategies, which must be adopted in complex situations. Representative examples of various methods are summarized, and the classes of strategies with respect to their appropriateness for particular applications.