Abstract

Methods for rotation-invariant pattern recognition using matched filters can be based on recognizing some invariant feature from the object of interest. Various techniques using circular harmonic components as the invariant features have been proposed. A new approach introduced here is to find a set of angular principal components q_k(θ) for the object of interest f(r, θ). The principal components q_k(θ) are found by defining vectors v_rk(θ) which are the values along constant radii r_k of the object and finding their principal components. Matched filters derived from those principal components may be used for pattern recognition that is invariant under rotations and changes of scale. If the object itself is used as the input to the filter, the scale invariance does not function well. Scale invariance is obtained when the projection of the object on the eigenvectors is used as the input to the principal component matched filter. This avoids calculation of the principal components for each target of interest, a much more time-consuming task than calculating the projections. Experimental results are shown.

© 1987 Optical Society of America