Abstract

We present a method to synthesize filters tuned for spatial frequency and orientation as linear combinations of a set of basis filters, allowing an efficient architecture for filtering and analyzing images with oriented filters. Using this technique, one can design filters which allow adaptive control of both the orientation and phase of a filter simply by changing the linear weightings in a sum of basis filters. The number of basis filters required depends on the polynomial order n of the filter, after modulation by a windowing function. Filter functions of even or odd parity require n + 1 basis filters; functions of mixed parity require 2n + 1 basis filters. Filters similar to those found in the visual cortex form a good basis set for image analysis. Steerable filters have application in the analysis of texture and local orientation; they can also be adapted for shape-from-shading analysis in a variant of Pentland's linear shape-from-shading scheme.

© 1989 Optical Society of America