Abstract: A method and computer program product are disclosed for determining an optimal classifier model for a pattern recognition system and updating the determined model to recognize new output classes. An initial plurality of classifier models are selected from a set of generated classifier models. An optimal representative classifier for each classifier model is selected according to an iterative optimization routine. The representative classifier having the highest associated value for a fitness function is accepted.

Abstract: A method and computer product is disclosed for arbitrating the outputs of a plurality of recognizers to select a hypothesis associated with a given input from a plurality of hypotheses. Fuzzy membership states are assigned to respective candidate outputs of the plurality of recognizers. At least one rule is selected according to the assigned fuzzy membership states. A selected rule provides a determined amount of support mass for an associated one of the plurality of hypotheses and a determined amount of uncertainty mass. A support value for each hypothesis and an overall uncertainty are determined from the provided mass values. A hypothesis having a highest support value while having an acceptable low level of uncertainty is selected.

Abstract: Systems and methods are provided for determining the orientation of an envelope. A plurality of classification elements are each operative to analyze at least one image of the envelope and produce at least one output value indicative of the orientation of the envelope. An arbitrator determines an associated orientation for the envelope according to the plurality of output values provided by the plurality of classification elements.

Abstract: A system for recognizing and identifying postal indicia on an envelope. This includes an image acquisition element that acquires a first image, representing a first side of the envelope, and a second image, representing a second side of the envelope. A feature extractor, for each of the first and second image, extracts a plurality of numerical feature values from each image as respective first and second feature vectors that represent the envelope. An orientation classification element classifies the envelope into one of a plurality of output classes representing a plurality of possible orientations according to the first and second feature vectors.

Abstract: A system is presented for recognizing and identifying postal indicia on an envelope. The system includes an image acquisition element that acquires a first image, representing a first side of the envelope, and a second image, representing a second side of the envelope, and generates first and second candidate images from respective opposing corners of the first image and third and fourth candidate images from respective opposing corners of the second image. A feature extractor that, for each candidate image, divides the candidate image into a plurality of regions, extracts a plurality of numerical feature values from each of the plurality of regions, and recombines the plurality of feature values into a feature vector that represents the image. A classification element classifies the image into one of a plurality of output classes representing various types of postal indicia according to the numerical feature vector.

Abstract: A system is presented for recognizing and identifying postal indicia on an envelope. The system includes an image acquisition element that acquires at least one binarized image of an envelope and isolates at least one region of interest from the at least one image. A candidate locator attempts to locate at least one candidate object within the at least one region of interest according to the distribution of dark pixels within the at least one region of interest. A feature extractor extracts an associated set of numerical feature values from each of the at least one candidate object. A classification element classifies each candidate object according to its associated set of numerical feature values.

Abstract: A method and computer program product are disclosed for automatically encoding a complex system architecture for a pattern recognition classifier. A plurality of subclassifier states are defined as a plurality of sets of state variables, each set of variables corresponding to a subclassifier state. A set of rules are then determined for a state machine governing transitions between the plurality of subclassifier states. The plurality of sets of state variables and the determined rules are encoded into a configuration file. This configuration file is provided to a generic classifier system, including a state machine and a predefined generic classifier object.

Abstract: The present invention recites a method and computer program product for identifying and correcting systematic noise in a pattern recognition classifier. A plurality of input patterns that have been determined not to be associated with any of a set of at least one represented output class are rejected by the pattern recognition classifier. A subset of pattern samples is selected from the rejected input patterns based upon the similarity of each pattern to one of the represented output classes. The selected pattern samples are subjected to an independent review to determine if they were correctly rejected. The classifier is retrained based upon this independent review of the selected pattern samples.

Abstract: A method and computer program product are disclosed for identifying output classes with multi-modal dispersion in feature space and incorporating multi-modal structure into a pattern recognition system architecture. A plurality of input patterns, determined not to be associated with any of a set of at least one represented output class by a pattern recognition classifier, are rejected. The rejected pattern samples are grouped into clusters according to the similarities between the pattern samples. Clusters that contain samples associated with a represented output class are identified via independent review. The classifier is then retrained to recognize the identified clusters as output pseudoclasses separate from the represented output class with which they are associated. The system architecture is reorganized to incorporate the output pseudoclasses. The output pseudoclasses are rejoined to their associated class after classification.

Abstract: A method and computer program product are disclosed for use in classifying an input pattern into one of a plurality of output classes. A plurality of modular classifiers each represent a set of at least one associated output class. The modular classifiers are capable of being trained separately from the system. The classifiers select one of the associated output classes as a classification result and compute an associated confidence value. This confidence value possesses common significance across the plurality of modular classifiers. A processing stage processes the confidence values from the plurality of modular classifiers. As a result of this processing, the processing stage selects an associated classification result.

Abstract: The present invention recites a method and computer program product for classifying an input pattern into an associated output class. For each of a plurality of classification rungs, each rung representing at least one output class, the a priori probability that an input pattern will be associated with a class represented by that rung is determined. A sequential order of processing for the plurality of classification rungs is assigned, based upon the a priori probability associated with each classification rung. A confidence value, associated with a represented class, is computed for each classification rung, in the assigned order, until a class is selected or a termination event occurs. The associated confidence values are compared to a predetermined threshold, and a class associated with a confidence value that first exceeds the threshold is selected, if a confidence value exceeds the threshold.

Abstract: The present invention recites a method and computer program product for generating a set of training samples from a single ideal pattern for each output class of a pattern recognition classifier. A system equivalent pattern is generated for each of a plurality of classes from a corresponding ideal pattern. A noise model, simulating at least one type of noise expected in a real-world classifier input pattern, is then applied to each system equivalent pattern a set number times to produce, for each output class, a number of training samples. Each training sample simulates defects expected in real-world classifier input patterns.

Abstract: A method and computer program product are disclosed for generating training data over a plurality of feature variables for a new output class in a pattern recognition classifier representing a plurality of existing classes with previously calculated statistical parameters. Deviation measures are generated for each feature variable for each of the plurality of existing classes from the calculated statistical parameters. The deviation measures for each feature variable are averaged across the plurality of existing classes. Feature data is extracted from an ideal pattern, representing the new class, for each of the feature variables. Statistical parameters are approximated for each feature variable for the new class from the extracted feature data and the averaged deviation measures.

Abstract: A method is disclosed for classifying an input pattern into an associated class through use of a compound classifier. Data pertaining to preselected features present within the input pattern are extracted. A discriminant value for each of a plurality of classes is then determined via a first classification technique. This value reflects the relative likelihood that a class is the associated class. The class with the highest relative likelihood is selected. A confidence value is generated via a second classification technique. This confidence value is reflective of the a posteriori probability that the selected class is the associated class. The selected class is rejected if the determined confidence value is below a predetermined threshold value.

Abstract: A method and computer product is disclosed for arbitrating the outputs of a plurality of recognizers to select a hypothesis associated with a given input from a plurality of hypotheses. Fuzzy membership states are assigned to respective candidate outputs of the plurality of recognizers. At least one rule is selected according to the assigned fuzzy membership states. A selected rule provides a determined amount of support mass for an associated one of the plurality of hypotheses and a determined amount of uncertainty mass. A support value for each hypothesis and an overall uncertainty are determined from the provided mass values. A hypothesis having a highest support value while having an acceptable low level of uncertainty is selected.

Abstract: A method and computer program product are disclosed for determining an optimal classifier model for a pattern recognition system and updating the determined model to recognize new output classes. An initial plurality of classifier models are selected from a set of generated classifier models. An optimal representative classifier for each classifier model is selected according to an iterative optimization routine. The representative classifier having the highest associated value for a fitness function is accepted.

Abstract: A method and computer program product are disclosed for identifying output classes with multi-modal dispersion in feature space and incorporating multi-modal structure into a pattern recognition system architecture. A plurality of input patterns, determined not to be associated with any of a set of at least one represented output class by a pattern recognition classifier, are rejected. The rejected pattern samples are grouped into clusters according to the similarities between the pattern samples. Clusters that contain samples associated with a represented output class are identified via independent review. The classifier is then retrained to recognize the identified clusters as output pseudoclasses separate from the represented output class with which they are associated. The system architecture is reorganized to incorporate the output pseudoclasses. The output pseudoclasses are rejoined to their associated class after classification.

Abstract: A method and computer program product are disclosed for determining an efficient set of features and an optimal confidence threshold value for a pattern recognition system with at least one output class. An initial set of features is selected based upon an optimization algorithm. A plurality of pattern samples are then classified using the selected feature set. A threshold confidence value is optimized as to maximize the accuracy of the classification. The selected feature set and threshold confidence value are accepted if a cost function based upon classification accuracy meets a predetermined threshold cost function value. The feature set is changed, by adding, removing or replacing a feature within the set based upon the optimization algorithm, if the cost function does not meet the predetermined threshold cost function value.

Abstract: A method and computer program product are disclosed for automatically encoding a complex system architecture for a pattern recognition classifier. A plurality of subclassifier states are defined as a plurality of sets of state variables, each set of variables corresponding to a subclassifier state. A set of rules are then determined for a state machine governing transitions between the plurality of subclassifier states. The plurality of sets of state variables and the determined rules are encoded into a configuration file. This configuration file is provided to a generic classifier system, including a state machine and a predefined generic classifier object.

Abstract: A method and computer program product is disclosed for automatically establishing a system architecture for a pattern recognition system with a plurality of output classes. Feature data is extracted from a plurality of pattern samples corresponding to a selected set of feature variables. A clustering algorithm is then applied to the extracted feature data to identify a plurality of clusters, including at least one cluster containing more than one output class. The identified clusters are arranged into a first level of classification that discriminates between the clusters using the selected set of feature variables. Finally, the output classes within each cluster containing more than one output class are arranged into at least one sublevel of classification that discriminates between the output classes within the cluster using at least one alternate set of feature variables.