Font- and size-invariant character recognition with grey-value image features

A novel approach for character recognition based on grey-value features is proposed, as the usual approach based on binary images fails when very small characters are analyzed. Size-invariance is achieved in two ways. A coarse scaling is determined by the Laplacian pyramid. Taking the appropriate level, the variation of size is within a factor of two. Complete size-invariance is achieved by the specific choice of the features themselves, which are weighed distributions of directions and direction changes. These features are also comparatively invariant to different fonts but quite specific to the basic structure of a character.<>