A fisher discriminant approach to distortion-invariant pattern recognition using autocorrelations Available to Purchase

A new approach to geometrical or distortion-invariant multi-class pattern recognition is advanced. In this technique, a generalized chord transform of the input image is produced by a new optical auto-correlation technique and a wedge-ring detector-sampled auto-correlation observation space. The dimensionality of this feature vector is greatly reduced. We consider a two-class pattern recognition problem with different distorted versions of each class of object available. From training set images of each class of object, an optimal linear discriminant function is produced that maximizes the Fisher ratio. When a new test image is present, a weighted linear combination of the dominant features is produced in a digital post-processor and from the scalar output of this vector inner product, the class of the input object is determined. Excellent results have been obtained in initial experiments that demonstrate the ability of this technique to recognize object views not present in the training set and to discriminate between two classes of objects independent of geometrical distortions in each class.