Classification of machine-printed and handwritten addresses on Korean mail piece images using geometric features

We propose an effective method for classifying machine-printed and handwritten addresses on Korean mail piece images. It is of vital importance to know if an input image is machine-printed or handwritten in such applications as address reading, form processing, FAX routing, and etc., since approaches for handwritten images are developed quite differently from those for machine-printed images. Our method consists of three blocks: valid connected component grouping, feature extraction and classification. A set of features related to width and position of groups of valid connected components is used for the classification based on a multi-layer perceptrons network. The experiment done with address images extracted from Korean live mail piece images has demonstrated the superiority of the proposed method. The correct classification rate for 3,147 testing images was about 98.9%.