Abstract

In a recent publication Wu and Stark¹ demonstrated a method for rotation-invariant pattern classification using optimum feature extraction. Here we present an alternative method for rotation-invariant classification. First, a filter function that is useful for classification of registered images (e.g., a maximum likelihood filter or an average filter) is determined from training sets consisting of objects from each of two classes. Circular-harmonic components of the filter are then computed. An input image is cross correlated with these circular-harmonic components to form a rotation-invariant vector as described in Ref.¹. Application of a linear transformation to this vector yields a feature that provides the basis for the classification decision. Photon-counting techniques² are used to implement the correlations in near real time. With photon-counting techniques, the input images for the two classes are acquired using a 2-D photon-counting detector and position-computing electronics. The photon statistics are used to compute the density function for the feature describing the two classes. The theory of hypothesis testing is used to construct a linear classifier on the basis of the calculated density function. Experimental results for various sets of object classes are presented; these results are in excellent agreement with theoretical predictions.

© 1986 Optical Society of America