Obtaining medial responses from steerable filters

In building up a multiscale, medial axis representation, the generation of medial responses from a set of boundary responses at multiple scales can be a computationally intensive process. Polar separable medial weighting kernels are applied to the computation of medial responses. An approximate method for the synthesis of rotated kernels is proposed. Because of the need to not only rotate, but to displace boundary responses by an amount dependent on the image scale, the traditional steerable filter approach must be modified to yield correct medial responses. Comparisons between integral evaluation and the adaptive filtering approximations show that 100-fold increases in modestly sized images at a single scale are possible. The error in the approximation is insignificant