Abstract

A moment-based pattern recognition system (in this case moment invariants) can be significantly improved by simply repeating the feature set in the Fourier domain. Moment-based feature sets in the space domain have these well-known properties: (1) they are best for distinguishing macroscopic regions of an image such as shapes; (2) high-order moments capture more information about the finer details contained in an image but are more susceptible to corruption by noise than are low-order moments. We show by theory and example that a feature set based on the low-order moments of the Fourier transform of a pattern has strengths that complement those of the usual space-domain moment-based feature sets. Improved results and a wider range of application result from using moment-based features from both domains. Basically, low-order Fourier-domain moments respond strongly to high-frequency details while ignoring the macroscopic shapes in an image. At the same time, their susceptibility to corruption by additive white noise is about the same as for low-order space-domain moments. Evidence shows that Fourier-domain moment invariants are far more useful for pattern recognition than the common ring-wedge-type features or other features based on the optical Fourier irradiance.

© 1987 Optical Society of America