Row-by-row segmentation and thresholding for optical character recognition

Claim 1

An optical character recognition system which recognizes an individual character in a thresholded image generated from a gray level document image by associating said character with the one symbol in a predetermined symbol set having the greatest similarity to said character, and computes a corresponding confidence score which is a function of the similarity of the character to the symbol having the greatest similarity and the similarity of the character to another one of the symbols having the next greatest similarity, said document image comprising an array of columns and rows of pixels and characterized by a succession of characters and spaces separating adjacent ones of said characters, said system comprising: thresholding means for discounting those pixels in said document image having a gray level below a certain threshold, whereby to generate said thresholded image from said document image; gray level threshold prediction means for determining a minimum value of said threshold at which the number of pixels not discounted by said thresholding means in a portion of one of said spaces is less than a background noise level and for reducing said threshold to said minimum value; and gray level threshold correction means responsive to the confidence score corresponding to the thresholded image of one of said characters for increasing said threshold from said minimum value to a new threshold whenever the corresponding confidence score is below a predetermined score, whereby said thresholding means generates another thresholded image of said one character corresponding to a higher confidence score.