Abstract

An invariant moment method is used to achieve image recognition and classification that is invariant under changes of position, rotation, contrast, and scale. The generalized image descriptors are calculated from circular Fourier radial Mellin transforms, which are radial moments of circular harmonic functions. The radial moments are redundant, and only one or a few orders need be used. The selection of the harmonic order depends on the geometrical properties of the image. The normalization procedure used to obtain scale and contrast invariance yields a weighting effect on the invariant features. Multiclass pattern recognition invariance under changes of position, orientation, scale, and contrast was achieved on some letters using the nearest neighbor rule. Experimental results are shown, including using the classification of images degraded by high levels of noise.

© 1985 Optical Society of America