Disclosed is an extraction method for optical character recognition systems in which the necessary character features for recognition process are extracted from a character information detected and digitized by an optical scanner. In accordance with this method, the local features of every adjacent two rows or columns of a digitized character pattern on a two-dimensional plane are extracted row by row or column by column and are successively integrated so as to extract in a global manner the respective feature types such as concavity and convexity, loop and connectivity of the character pattern. At the same time, the global feature regions or segments are separated and each of the separated feature segments is coded without any loss of the information. The feature quantity of each segment is obtained on the basis of the coded representation. Any local concavity or convexity noises are eliminated simultaneously with the extraction of the global features.