New primitives to reduce the effect of noise for handwritten features extraction

A method for feature extraction for a handwritten OCR system is presented. In order to reduce the effect of the noise which is either an original noise or obtained as a result of the preprocessing stages, there is a need to develop a feature extraction method invariant to the expected distortions, and less dependent on the locations of high probable appearance of noise and distortion. This method depends only on the two primitive features: straight lines and curves. A chain code has been built from the thinned shape of the character. Two rules have been introduced to cut this chain code into small segments. From each segment one feature is defined and for each input character, a feature vector will be built. The prototype system was tested for alphanumeric characters and the results were satisfactory