Simple and Effective Feature Extraction for Optical Character Recognition*

A new representation method for recognition of handwritten charcters, called LLF (Local Line Fitting), is presented. The method, based on simple geometric operations, is very efficient and yields a relatively low-dimensional and distortion invariant representation. An important feature of the approach is that no preprocessing of the input image is required. A black & white or gray-scale pixel representation is directly used without thinning, contour following, binarization, etc. Therefore, high recognition speed can be achieved. Experiments using this parametrization method and several classification procedures on handwritten digits and letters are reported.