Image classifier utilizing class distribution maps for character recognition

Claim 1

A method of character recognition, comprising the steps of: receiving an input image that is to be assigned to one of a plurality of image classes; extracting a respective numerical value of each of plural features of the input image, wherein at least one of said features is defined according to a rule that relates to the shapes of images of at least one said image class; recording the input-image feature values; providing a tabulation of those feature values that occur in a training set, wherein: (i) said training set comprises, for each image class, a plurality of training images that are known to belong to that class; and (ii) said tabulation associates each tabulated feature value with a corresponding class; comparing the input-image feature values to at least some of the tabulated training-image feature values; and based on a result of said comparing step, assigning the input image to one of the image classes; CHARACTERIZED IN THAT the step of providing a tabulation comprises providing a respective class distribution map for each image class; each said class distribution map contains, for each of the plural features, a record of each numerical value of said feature that was exhibited by at least one training image of the pertinent class; the comparing step comprises evaluating a respective class distance between each of the class distribution maps and a test map that tabulates the respective input-image feature values; in the assigning step, the input image is assigned to that class having the least distant class distribution map as determined in the comparing step; each class distance is evaluated by counting non-matched features; a feature is counted as non-matched if and only if the numerical value of said feature as evaluated on the input image fails to match any of the numerical values of said feature tabulated in the pertinent class distribution map; and in the counting of non-matched features, every feature is given equal weight.

Claim 2

A method of character recognition, comprising the steps of: receiving an input image that is to be assigned to one of a plurality of image classes; extracting a respective numerical value of each of plural features of the input image, wherein at least one of said features is defined according to a rule that relates to the shapes of images of at least one said image class; recording the input-image feature values; providing a tabulation of those feature values that occur in a training set, wherein: (i) said training set comprises, for each image class, a plurality of training images that are known to belong to that class; and (ii) said tabulation associates each tabulated feature value with a corresponding class; comparing the input-image feature values to at least some of the tabulated training-image feature values; and based on a result of said comparing step, assigning the input image to one of the image classes; CHARACTERIZED IN THAT the step of providing a tabulation comprises providing a respective class distribution map for each image class; each said class distribution map contains, for each of the plural features, a record of each numerical value of said feature that was exhibited by at least one training image of the pertinent class; the comparing step comprises evaluating a respective class distance between each of the class distribution maps and a test map that tabulates the respective input-image feature values; in the assigning step, the input image is assigned to that class having the least distant class distribution map as determined in the comparing step; each class distance is evaluated by counting non-matched features; a feature is counted as non-matched if and only if the numerical value of said feature as evaluated on the input image fails to match any of the numerical values of said feature tabulated in the pertinent class distribution map; and in the counting of non-matched features, every feature is given equal weight.