Pre-processing and feature extraction system for character recognition

Abstract

Character recognition apparatus and a method for identifying machine-printed and hand-printed characters. A character is optically scanned and converted to a rectangular matrix representation containing black and white points. The matrixed character is pre-processed by a non-iterative line-thinning technique so as to reduce the stroke width to no more than two points along the center line of the matrixed character. Isolated black points formed in the process of thinning the strokes are removed. Any gaps in the center-line of the matrixed character are filled and the edges of all long strokes are smoothed. The thinned character matrix is divided into a plurality of regions. One of a set of predetermined stroke features of the character matrix is extracted by scanning subregions of the character matrix such that the center position of the scanned subregions will have occupied all elements of the character matrix after one complete scan. A particular stroke feature is identified by an analysis of the arrangement of black and white points detected in each subregion. The feature extracted will be assigned to the region containing the center position of the particular subregion which was scanned. The features extracted in each of the regions are used to identify the character. In one embodiment, the sequence of stroke features detected for regions of the character matrix is compared with a table of stroke sequences corresponding to previously identified characters. The matrix is assigned to the character whose stroke sequences most closely match that of the matrixed character. Alternatively, a weighted value may be assigned to each of the possible stroke features in each region for every possible character, based on the relative importance of that feature to the character in that region, as determined by analysis of previously identified samples of the particular hand-printed or machine-printed characters. The matrixed character is identified by summing the weighted values of the detected features assigned to each region to produce a score for each character matrix. The matrix is then assigned to the character yielding the highest score.