Abstract

This paper investigates the applicability of Gabor functions to textural segmentation. Gabor functions are sinusoidal plane waves in 2-D Gaussian envelopes. The choice of parameters characterizing the geometry of an individual Gabor function affects its spatial extent as well as orientation and spatial frequency tuning. Daugman has indicated that these functions belong to a class of filters having optimal joint resolution in the 2-D space and 2-D frequency domains. They are, therefore, appropriate filter choices for tasks which require selective measurement in these domains. Textural segmentation appears to be one of those tasks. A set of Gabor functions of different frequencies and orientations is applied by computer program to images containing regions of different texture. This process produces a kind of localized and orientation selective frequency spectrum of various fields in the image. The program then attempts to delineate the boundaries of the textured regions by identifying spectrum differences between these fields. Gabor functions are effective in distinguishing between many of the textures used in psychophysical studies differing in first- or second-order statistics. Additional textures in which the difference is related to some aspect of the collinearity of the texture elements have also been tried with promising results.

© 1985 Optical Society of America