Adjustable invariant features by partial Haar-integration

A very common type of a-priori knowledge in pattern analysis problems is invariance of the input data with respect to transformation groups, e.g. geometric transformations of image data like shifting, scaling etc. For enabling most general analysis techniques, this knowledge should be incorporated in the feature-extraction stage. In the present work a method for this, called Haar-integration, is generalized to make it applicable to more general transformation sets, namely subsets of transformation groups. The resulting features are no longer precisely invariant, but their variability can be adjusted and quantified. Experimental results demonstrate the increased separability by these features and considerably improved recognition performance on a character recognition task.