Pattern recognition system with statistical classification

Claim 2

A pattern recognition system comprising: a training subsystem that receives during a training operation a plurality of training input patterns of a data type from a plurality of subject classes, that forms a set of categories of the training input patterns, that assigns each category a category definition according to training input patterns received within the category, that counts the training input patterns received for each class within each category and that generates for each category a training histogram of the training input patterns received within the category, the training histogram including counts of training input patterns received for each class within the category; and a classifier that receives during a testing operation at least one test input pattern of the data type from the subject, that accesses the set of categories and computes a correlation between a category definition and each test input pattern, that forms a category association between each test input pattern and a category based on the correlation and that forms an observation histogram to classify the subject, the observation histogram being formed from each training histogram of each category of each category association and representing counts of training input patterns received by the training subsystem during the training operation, classification of the subject being determined by a peak class of the observation histogram, the peak class representing the highest training input pattern count of the observation histogram.

Claim 3

The pattern recognition system of claim 1 wherein the training subsystem assigns to each category a label that is determined by a peak class of the category.

Claim 4

The pattern recognition system of claim 2 wherein the set of categories includes a category labeled as unknown such that the subject can be classified as unknown.

Claim 5

The pattern recognition system of claim 1 wherein during the training operation, the training subsystem computes a correlation between each training input pattern and a category definition vector for each category.

Claim 6

The pattern recognition system of claim 4 wherein if the correlation is below a threshold, a new category is formed.

Claim 7

The pattern recognition system of claim 4 wherein if the correlation is above a threshold, an existing category definition vector is modified by an incoming training input pattern according to a learning rule.

Claim 8

The pattern recognition system of claim 6 wherein the learning rule comprises creating a new category definition vector by multiplying a learning rate by a difference between the incoming training input pattern and the existing category definition vector and adding that product to the existing category definition vector.

Claim 9

The pattern recognition system of claim 4 wherein the correlation includes a vector dot product.

Claim 10

The pattern recognition system of claim 1 wherein, during the training operation, an existing category definition vector is modified by an incoming training input pattern by multiplying a learning rate by a difference between the incoming training input pattern and the existing category definition vector and adding that product to the existing category definition vector.

Claim 11

The pattern recognition system of claim 1 wherein for each category the training subsystem computes a category class contrast indicative of the degree to which a peak class of the category dominates the other classes of the category.

Claim 12

The pattern recognition system of claim 10 wherein if the category class contrast exceeds a threshold, the category is labeled as the peak class.

Claim 13

The pattern recognition system of claim 10 wherein if the category class contrast is below a threshold, the category is labeled as unknown.

Claim 14

The pattern recognition system of claim 1 further comprising: a second training subsystem that receives a second plurality of training input patterns of a second data type from the subject classes and forms a second set of categories; a second classifier that receives at least one test input pattern of the second data type from the subject during the testing operation and forms a category association between the at least one test input pattern of the second data type and a category of the second set of categories and forms a second observation histogram; and a processor for combining the first and second observation histograms into a cumulative histogram to produce a cumulative classification of the subject within a class.

Claim 15

The pattern recognition system of claim 13 wherein at least one data type comprises visual data.

Claim 16

The pattern recognition system of claim 13 wherein at least one data type comprises sound data.

Claim 17

The pattern recognition system of claim 13 wherein the data types comprise visual and sound data.

Claim 18

The pattern recognition system of claim 13 wherein at least one data type comprises voice data.

Claim 19

The pattern recognition system of claim 1 wherein the observation histogram for a single test input pattern is a single-frame k-nearest neighbor histogram of a predetermined number k of categories which have the closest correlations to the single test input pattern.

Claim 20

A pattern recognition system comprising: a source for generating input patterns of multiple data types; a plurality of training subsystems, each training subsystem receiving a plurality of training input patterns of a single corresponding data type during a training operation, forming a set of categories of the training input patterns of the data type, generating a category definition for each category according to training input patterns received within the category, counting the number of training input patterns of each class within each category of the data type, and generating a training histogram for each category of the data type, each training histogram including counts of training input patterns of each class within the respective category; a plurality of classifiers, at least one classifier receiving during a testing operation at least one test input pattern of its corresponding data type, accessing the categories and computing a correlation between a category definition and each test input pattern, forming a category association between each test input pattern and a category based on the correlation and forming for each data type an observation histogram to classify the subject, the observation histogram being formed from each training histogram of each category of each category association and representing counts of training input patterns received by the training subsystem during the training operation; and a processor that combines each observation histogram for each data type into a cumulative histogram to form a cumulative classification of the subject within a class.

Claim 21

A pattern recognition system comprising: a source for generating input patterns of multiple data types; a plurality of training subsystems, each training subsystem receiving a plurality of training input patterns of a single corresponding data type during a training operation, forming a set of categories of the training input patterns of the data type, generating a category definition for each category according to training input patterns received within the category, counting the number of training input patterns of each class within each category of the data type, and generating a training histogram for each category of the data type, each training histogram including counts of training input patterns of each class within the respective category; a plurality of classifiers, at least one classifier receiving during a testing operation at least one test input pattern of its corresponding data type, accessing the categories and computing a correlation between a category definition and each test input pattern, forming a category association between each test input pattern and a category based on the correlation and forming for each data type an observation histogram to classify the subject, the observation histogram being formed from each training histogram of each category of each category association and representing counts of training input patterns received by the training subsystem during the training operation; and a processor that combines each observation histogram for each data type into a cumulative histogram to form a cumulative classification of the subject within a class.

Claim 22

The pattern recognition system of claim 19 wherein at least one data type comprises visual data.

Claim 23

The pattern recognition system of claim 19 wherein at least one data type comprises sound data.

Claim 24

The pattern recognition system of claim 19 wherein the data types comprise visual and sound data.

Claim 25

The pattern recognition system of claim 19 wherein at least one data type comprises voice data.

Claim 28

The method of claim 24 further comprising assigning to each category a label that is determined by a peak class of the category.

Claim 29

The method of claim 25 wherein assigning a label includes assigning a label of unknown to a category such that the subject can be classified as unknown.

Claim 30

The method of claim 24 wherein the training operation includes computing a correlation between each training input pattern and a category definition vector of each category.

Claim 31

The method of claim 27 wherein if the correlation between a training input pattern and a best match category definition vector is below a threshold, a new category is defined.

Claim 32

The method of claim 27 wherein if the correlation between a training input pattern and a best match category definition vector is above a threshold, the best match category definition vector is modified by the training input pattern according to a learning rule.

Claim 33

The method of claim 29 wherein the learning rule comprises creating a new category definition vector by multiplying a learning rate by a difference between the training input pattern and the existing category definition vector and adding that product to the existing category definition vector.

Claim 34

The method of claim 27 wherein the correlation includes a vector dot product.

Claim 35

The method of claim 24 wherein, during the training operation, an existing category definition vector is modified by an incoming training input pattern by multiplying a learning rate by a difference between the incoming input training pattern and the existing category definition vector and adding that product to the existing category definition vector.

Claim 36

The method of claim 24 wherein the training operation includes computing for each category a category class contrast indicative of the degree to which a peak class of the category dominates the category.

Claim 37

The method of claim 33 wherein if the category class contrast exceeds a threshold, the category is labeled as the peak class.

Claim 38

The method of claim 33 wherein if the category class contrast is below a threshold, the category is labeled as unknown.

Claim 39

The method of claim 24 wherein forming an observation histogram comprises generating for a single test input pattern a single-frame k-nearest neighbor class histogram of a predetermined number k of categories which have the closest correlation with the single test input pattern.

Claim 40

A pattern recognition system comprising: a memory having stored therein a set of categories of training input patterns received from a plurality of subject classes during a training operation, each category having a category definition according to training input patterns received within the category, and each category being associated with a training histogram of the training input patterns received within the category, the training histogram including counts of training input patterns received for each class within the category; and a classifier that receives during a testing operation at least one test input pattern from the subject, that accesses the set of categories and computes a correlation between a category definition and each test input pattern, that forms a category association between each test input pattern and a category based on the correlation and that forms an observation histogram to classify the subject, the observation histogram being formed from each training histogram of each category of each category association and representing counts of training input patterns received by the training subsystem during the training operation, classification of the subject being determined by a peak class of the observation histogram, the peak class having the highest training input pattern count of the observation histogram.

Claim 41

A pattern recognition system comprising: a memory having stored therein a set of categories of training input patterns received from a plurality of subject classes during a training operation, each category having a category definition according to training input patterns received within the category, and each category being associated with a training histogram of the training input patterns received within the category, the training histogram including counts of training input patterns received for each class within the category; and a classifier that receives during a testing operation at least one test input pattern from the subject, that accesses the set of categories and computes a correlation between a category definition and each test input pattern, that forms a category association between each test input pattern and a category based on the correlation and that forms an observation histogram to classify the subject, the observation histogram being formed from each training histogram of each category of each category association and representing counts of training input patterns received by the training subsystem during the training operation, classification of the subject being determined by a peak class of the observation histogram, the peak class having the highest training input pattern count of the observation histogram.

Claim 42

The pattern recognition system of claim 37 wherein the set of categories includes a category labelled as unknown such that the subject can be classified as unknown.

Claim 43

The pattern recognition system of claim 37 wherein the observation histogram for a single test input pattern is a single-frame k-nearest neighbor histogram of a predetermined number k of categories which have the closest correlations to the single test input pattern.