Canon Inc.

Canon Research Center America, Inc.

Hadar I Avi-Itzhak

Harry T Garland

Thanh A Diep

Roger D Melen

Michael G Roberts

Kenneth M Hunter

An OCR system 10 classifies an input image vector of an unclassified symbol with respect to a library 14T of template image vectors of pre-classified characters. Each template vector is in the form of a sequence of elements representing the image intensity level of a pixel within the character defined by that template vector. Each template element is part of the image background, foreground, or transition ground between the background and foreground. Each input vector, like the template vectors, is also in the form of a sequence of elements. However, in the input vector, each element represents the sum or an image intensity level signal component defining the symbol within the image of the unclassified symbol plus a greyscale noise component. Each input element is also part of the background, foreground, or transition ground. The input vector and at least one of the template vectors are entered into a classifier device 18. The input vector is classified based on the backgrounds, foregrounds, and transition grounds. The presence of transition ground in the input vector and the template vector produces a robust classification response with a more uniform correlation coefficient between repeated classifications of the same input symbol. The classifier device may be a distance function classifier or a neural network classifier.

OCR classification based on transition ground data

Pattern recognition, for instance optical character recognition, is achieved by training a neural network, scanning an image, segmenting the image to detect a pattern, preprocessing the detected pattern, and applying the preprocessed detected pattern to the trained neural network. The preprocessing includes determining a centroid of the pattern and centrally positioning the centroid in a frame containing the pattern. The training of the neural network includes randomly displacing template patterns within frames before applying the template patterns to the neural network.

Training a neural network using centroid dithering by randomly

Method and system for selectively retrieving information contained in a stored document set using a non-literal, or 'fuzzy', search strategy. A text string query is transmitted (200) to a computer processor, and a dissimilarity value D.sub.i is assigned (208) to selected ones of stored text strings representative of information contained in a stored document set, based upon a first set of rules (106). A set of retrieved text strings representative of stored information and related to the text string query is generated (212), based upon a second set of rules (107). Each of the retrieved text strings has an associated dissimilarity value D.sub.i, which is a function of at least one rule R.sub.n from the first set of rules (106) used to retrieve the text string and a weight value w.sub.n associated with that rule R.sub.n. The retrieved text strings are displayed (216) preferably in an order based on their associated dissimilarity value D.sub.i. Once one or more of the retrieved text strings is chosen, the weight value w.sub.n associated with at least one rule of the first set of rules (106) is adjusted (220) and stored.

Adaptive non-literal text string retrieval

A general method of accelerating the classification of an unclassified input symbol S into a library of pre-classified templates employs a comparison inequality function. The correlation calculation is initiated between the input symbol S and templates retrieved from the library of templates. Templates with a low degree of correlation are excluded prior to completion of the correlation calculation to accelerate the correlation. Sorting by temple intensity permits the temple foreground pixels to be correlated first, which also accelerates the correlation calculation. A second sort by discrimination power permits critical conflict pixels to be correlated first for further accelerating the correlation calculation.

Comparison inequality function based method for accelerated OCR

A library of L unenhanced images (pixel bit maps) is enhanced for optical character recognition (OCR) with respect to a pre-existing group of G input symbol classes for creating a library of G recognition enhanced templates (pixel bit maps) of the G input symbol classes. The enhancement is accomplished by comparing each image of the library with each pixel version of each symbol class (forming V sets of L comparisons), and weighting the images with the highest potential for confusion. Divergent pixel version are further divided into sub-classes, each with a sub-image which evolves into a sub-template. A primary comparison C* and a secondary comparison C** are identified from the L comparisons within each of the V sets of comparisons. A recognition margin is determined between each pair of identified comparisons C* and C**. The single pair of identified comparisons C* and C** is selected forming the smallest recognition margin M*. The single pair of images I.sub.j * and I.sub.j ** underlying the pair of identified comparisons C* and C** is identified. Certain pixels of the closest pixel image I.sub.j * and the next closest pixel image I.sub.j ** are weighted in order to increase the recognition margin M* therebetween. The steps are iterated until the library of pixel images (and sub-images) has become a library of enhanced symbol templates (and sub-templates). The weighted pixel aberrations generated in the enhanced templates (and sub-templates) are not present in the corresponding symbol of the pre-existing group of G input symbol classes (and sub-classes).

Method of multi-font template enhancement by pixel weighting

High accuracy optical character recognition using neural networks with

A library of L unenhanced images (pixel bit maps) is enhanced for optical character recognition (OCR) with respect to a pre-existing group of G input symbols (pixel bit maps) for creating a library of G recognition enhanced templates (pixel bit maps) of the G input symbols. The enhancement is accomplished by comparing each image of the library with each symbol of the group, and weighting the images with the highest potential for confusion. A primary comparison C* and a secondary comparison C** are identified from the L comparisons within each of the G sets of comparisons. A recognition margin is determined between each pair of identified comparisons C* and C**. The single pair of identified comparisons C* and C** is selected forming the smallest recognition margin M*. The single pair of images I.sub.j * and I.sub.j ** underlying the pair of identified comparisons C* and C** is identified. Certain pixels of the closest pixel image I.sub.j * and the next closest pixel image I.sub.j ** are weighted in order to increase the recognition margin M* therebetween. The steps are iterated until the library of pixel images has become a library of sufficiently enhanced symbol templates. The weighted pixel abberations generated in the enhanced templates are not present in the corresponding pixel symbol of the pre-existing group of G pixel symbols.

Growing community of inventors

Mountain View, CA, United States of America

Hadar I Avi-Itzhak