JP2004178569A - Data classification device, object recognition device, data classification method, and object recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve a problem where, although inputted data are required to be classified into appropriate categories, the categories are classifications required for an application, or types defined by a user, so that the distributions of the categories are partially overlapped depending on the inputted data, and the optimum classification can not be executed. <P>SOLUTION: This data classification device is so structured that the number of classes and the number of dimensions are previously calculated in order to classify a model data group comprising a plurality of model data into classes of separated distributions; a class configuration is determined according to the number of classes and the number of dimensions having been obtained; the model data group is classified class-by-class; a representative value of each class is compared with inputted data to determine the class having a representative value having the highest degree of similarity; and the category of the inputted data is found by finding the category to which the determined class belongs. <P>COPYRIGHT: (C)2004,JPO

Description

  The present invention classifies a category into which data is classified, a data classification device and a data classification method for constructing a class having a distribution smaller than the category, and classifies input data into a class constituted by the data classification device. The present invention relates to an object recognizing device and an object recognizing method for recognizing an object included in data.

  2. Description of the Related Art Conventionally, there is an object recognition apparatus that classifies sequentially input data into categories and recognizes an object included in the data.

  First, a conventional object recognition apparatus obtains a vector of input data and a representative vector of each category when classifying which category the input data falls into. Here, the representative vector of the category is a center-of-gravity vector of vectors of all model data included in the category.

Next, the conventional object recognition device compares the input data vector with the representative vector of the category, and detects a category having a representative vector having the highest similarity to the input data vector. Then, the conventional object recognition device classifies the input data into a category having a representative vector having the highest similarity, and recognizes an object included in the data (for example, Patent Document 1).
JP-A-2000-242784 (pages 4 to 6, FIGS. 2 and 3)

  Here, the object recognition device is required to classify the input data into an appropriate category. However, since the category is a classification required by the application or a type determined by the user, the input data is In some cases, there is a problem that the distribution of categories partially overlaps, and it is not possible to perform optimal classification.

  Here, it is required to classify the categories into classes with a smaller distribution, but the classification of the data is inappropriate, the number of classes is too large, or the class configuration has no generalization ability. There was a problem.

  The present invention has been made in view of such a point, and in the category distribution, when the classification is inappropriate, an appropriate class is configured, the input data is classified into an accurate category, and included in the input data. An object of the present invention is to improve the accuracy of object recognition.

  In order to solve this problem, the present invention calculates a class number and a dimension number in advance in order to classify a model data group including a plurality of model data into classes that are separated distributions. The class configuration is determined according to the number of dimensions, the model data group is classified for each class, the representative value of each class is compared with the input data, and the class having the highest similarity is determined. The category of the input data is obtained by obtaining the category to which the determined class belongs.

In other words, by classifying a model data group consisting of a plurality of model data according to predetermined items in advance, the model data group is classified again into a class having a distribution smaller than the category required by the application and having no redundancy, and representing the class. The value and the sequentially input data are compared, the class having the representative value having the highest similarity to the input data is determined, and the category to which the determined class belongs is determined to determine the category of the input data. That is.

  In other words, the class is not a class composed of several model data, and is not specialized for a group of model data. The class configuration is also effective for sequentially input data.

  As described above, since sequentially input data can be classified into an accurate class and into an accurate category, the accuracy of data classification can be improved, and practical data classification can be performed.

  As described above, according to the present invention, the number of classes and the number of dimensions that can accurately perform class classification are obtained, and then the class configuration according to the number of classes and the number of dimensions is obtained. The class having the highest similarity to the data to be detected is detected, and the category to which the detected class belongs can be obtained. Thereby, regardless of the distribution of the categories, the sequentially input data can be classified into correct categories, and the advantageous effect of reducing the probability of misclassifying the data by the conventional method can be reduced. can get.

  According to a first aspect of the present invention, there is provided a method for determining the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution smaller than the category in which a plurality of model data are classified, and determining the number of classes and the number of dimensions. Unit, a class configuration is calculated using the number of classes and the number of dimensions, a class configuration construction unit that classifies the model data into the class, and the model data classified into the class, the intra-class variance is small, A feature extraction matrix calculation unit that calculates a feature extraction matrix used when converting the inter-class variance into a large distribution, and calculates an average value of a feature vector for each class from the model data classified into the class and the feature extraction matrix. It is a data classification device that has a model feature vector calculation unit. Classification class configuration can be obtained which is suitable for the data can be classified on the exact class, an effect that can be classified into and accurate categories.

  According to a second invention of the present invention, the class number / dimension number determination unit is configured to determine the number of classes having a class configuration that minimizes an information amount criterion that is a criterion for model selection using a model probability distribution estimated from a plurality of model data. The first invention calculates a class number and a dimension number, and the class configuration constructing unit constructs a class by clustering, which is a technique for reducing intra-class variance, using the class number and the dimension number with the minimum information amount criterion. The use of the information amount criterion and the clustering has an effect that a class having a distribution smaller than a category can be easily constructed.

In a third aspect of the present invention, the class number / dimension number determining unit converts the plurality of model data using a covariance matrix which is a unit matrix calculated from the plurality of model data, and Calculates the number of classes and the number of dimensions that minimize the information amount criterion using the model data of, and the class configuration construction unit reduces the intra-class variance by using the number of classes and the number of dimensions that minimizes the information amount criterion A class is constructed by clustering, and the feature extraction matrix calculation unit converts the plurality of model data converted by the covariance matrix of the unit matrix into a distribution having a small intra-class variance and a large inter-class variance of the class. A data classification apparatus according to the first aspect of the present invention, which calculates a patent extraction matrix used in a case, converts a plurality of model data using a covariance matrix, and An effect that it is possible to easily construct a class is smaller distribution.

  According to a fourth aspect of the present invention, the class number / dimension number determining unit determines the class number by using a judgment evaluation criterion including a similarity between the distribution of the feature vector calculated from the plurality of model data and the distribution of the category or the class. And a data classification device according to the first aspect of the present invention, which calculates the number of dimensions, and uses an evaluation criterion including a similarity between the distribution of the model data and the distribution of the category or the class to categorize the category without redundancy. Classification is performed into classes, and pattern vectors are extracted for each of the classified categories, and the classification vector can be easily classified based on the distribution of the pattern vectors.

  A fifth invention of the present invention is the data classification device according to the fourth invention, wherein the judgment evaluation criterion is a criterion including a Kullback-Leibler information amount, and the judgment evaluation criterion F is minimized. , The distribution is similar to the class configuration of the sample derived from the Fisher criterion, but has the effect that the number of classes and the number of dimensions that minimize the AIC can be obtained.

  According to a sixth aspect of the present invention, in the class configuration constructing section, the first evaluation amount criterion calculated using the model data already classified into the class, the first class number and the dimension number, Is compared with the second evaluation criterion calculated using the second class number and the dimension number different from the first class number and the dimension number, and the number of classes having the smaller evaluation criterion is compared. And a data classifying apparatus according to any one of the first to fifth aspects of the present invention, which reconstructs a class using the number of dimensions and the number of dimensions. Class composition can be determined, data can be classified into accurate classes, and classified into accurate categories, data classification accuracy can be improved, practical data classification can be performed, and objects included in data can be obtained. of An effect that identification accuracy is improved.

  According to a seventh aspect of the present invention, there is provided the data classification device according to any one of the first to sixth aspects, wherein input data including an object to be recognized is converted into an input feature vector using the feature extraction matrix. An input feature vector calculation unit, an input for recognizing the object by comparing the average value of the feature vector for each class with the input feature vector and obtaining the category to which the class most similar to the input data belongs; It is an object recognition device including a data determination unit, and even if the distribution of the category is not separable, a class configuration suitable for classification can be obtained, and the data can be classified into an accurate class. Object recognition can be performed using a database that can be classified into accurate categories, and the accuracy of object recognition can be improved. It is having an effect.

  An eighth aspect of the present invention is a first step of calculating the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution smaller than the category into which the plurality of model data are classified, A second step of calculating a class configuration using the number of classes and the number of dimensions, and classifying the model data into the classes; and Is a third step of calculating a feature extraction matrix used when converting to a large distribution, and a fourth step of calculating an average value of feature vectors for each class from the model data classified into the class and the feature extraction matrix. Classification suitable for classification even when it is not separable with the distribution of the category as it is Mel it possible, data can be classified into correct class, an effect that can be classified into and accurate categories.

A ninth aspect of the present invention is a first step of calculating the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution smaller than the category into which the plurality of model data are classified, A second step of calculating a class configuration using the number of classes and the number of dimensions, and classifying the model data into the classes; and Is a third step of calculating a feature extraction matrix used when converting to a large distribution, and a fourth step of calculating an average value of feature vectors for each class from the model data classified into the class and the feature extraction matrix. And input data including an object to be recognized is input, and the input data is converted into an input feature vector using the feature extraction matrix. A fifth step of recognizing the object by comparing the average value of the class with the input feature vector and obtaining the category to which the class most similar to the input data belongs; It is a method of recognizing objects that has a class distribution suitable for classification even if it is not separable with the distribution of categories as it is, and it is possible to classify data into accurate classes, and to classify accurate data Since the object recognition can be executed by using the database that can be classified, there is an effect that the accuracy of the object recognition can be improved.

  According to a tenth aspect of the present invention, a computer calculates a class number and a dimension number which are parameters used when constructing a class having a distribution smaller than the category by classifying a plurality of model data. And a second step of calculating a class configuration using the number of classes and the number of dimensions, and classifying the model data into the class. A third step of calculating a feature extraction matrix used when converting the inter-class variance into a large distribution, and calculating an average value of a feature vector for each class from the model data classified into the class and the feature extraction matrix. This is a program for executing the data classification method including the fourth step. Possible even if it is not, classification class structure can be obtained which is suitable for the data can be classified on the exact class, an effect that can be classified into and accurate categories.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(Embodiment 1)
FIG. 1 shows a block diagram of the object recognition device according to the first embodiment. Here, a configuration in which data is image data will be described.

  Here, 101 indicates a computer system that performs data classification, 102 indicates a CPU that performs predetermined processing based on a program, 103 indicates a memory that stores a program and serves as a work area of the CPU 102, and 104 Reference numerals 107 denote interface units (Interface Units: I / F) for exchanging data with the outside, and 108 is provided outside the computer system, which is configured by a secondary storage device, a magneto-optical disk, or the like. A model data database for storing model data, which is image data constituting a category, is shown. Reference numeral 109 denotes a model data memory for temporarily storing model data in the model data database.

  Reference numeral 110 denotes a video camera for capturing an image of an object, 111 denotes an image memory for temporarily storing image data sequentially input from the video camera 110, and 112 denotes a keyboard serving as input means. 114 denotes a pattern vector memory that stores a pattern vector expressing a pixel value of image data as a one-dimensional vector. 115 denotes a model pattern vector that is a pattern vector of model data. 4 shows a category configuration model data database stored for each category.

Reference numeral 116 denotes a class configuration model data database in which model data is stored for each class configured according to the number of classes and the number of dimensions. Reference numeral 117 denotes a CPU bus that connects each configuration inside the computer system 101; Denotes an output terminal for outputting the result of classification by the object recognition device.

  Further, reference numeral 120 denotes a class number / dimension number memory storing the number of classes and dimensions, 121 denotes a feature extraction matrix memory storing a feature extraction matrix for converting a coordinate system into a feature coordinate system, and 122 denotes a feature extraction matrix memory. A model feature vector database that stores a model feature vector that is a feature vector of model data is shown. Reference numeral 123 denotes a feature vector memory that stores an input feature vector that is a feature vector of image data.

  The object recognition device configured as described above constitutes a class that classifies model data classified in each category stored in advance into finer classes without redundancy, and represents a representative value of a feature vector indicating a feature of each class. Is compared with the image data sequentially input from the video camera 110 and the representative value of the feature vector obtained in the offline processing, the corresponding class is determined, the image data is classified, and included in the image data. It is roughly divided into online processing for recognizing objects.

  First, the offline processing of the object recognition device will be described in detail with reference to FIGS.

  FIG. 2 is a functional block diagram for offline processing of the object recognition device according to the first embodiment.

  Here, reference numeral 201 denotes a pattern vector creating unit that creates a pattern vector from a model pattern, and 202 denotes a class number / dimension number calculation that calculates the number of classes and the number of dimensions in a class configuration for finely classifying categories without redundancy. 203, a class configuration constructing unit that reconfigures the category configuration into a class configuration, 204, a feature extraction matrix calculating unit that is a learning unit that creates a feature vector from model data, and 205, a feature extracting matrix calculating unit. A model feature vector calculation unit that calculates a pattern vector by converting a pattern vector into a model feature vector using an extraction matrix corresponds to a program stored in the memory 103 and is executed by the CPU 102.

  The offline processing includes I / Fs 104 and 106, a model data memory 109, a keyboard / display 112, a category configuration model data database 115, a class configuration model data database 116, a class / dimension memory 120, a feature extraction matrix memory 121, It comprises a pattern vector creation unit 201, a class / dimension determination unit 202, a class configuration construction unit 203, a feature extraction matrix calculation unit 204, and a model feature vector calculation unit 205.

  Note that FIG. 2 can be used as a feature vector generation device that classifies model data belonging to a category into classes and generates feature vectors of the classes.

  FIG. 3 is an operation flowchart of the offline processing of the object recognition device according to the first embodiment.

  First, the relationship between the functional blocks and the operation flowchart will be described, and then a detailed description of each functional block will be described.

  The model data database 108 stores a plurality of model data. The model data is stored in the model data memory 109 via the I / F 104 (S301).

The pattern vector creating unit 201 arbitrarily extracts a model pattern that is a rectangular partial image from the model data, and creates a pattern vector that is a vector expressing a pixel value as a one-dimensional vector (S302).

  Next, the pattern vector creation unit 201 creates a label describing the category name classified into the pattern vector from the category information corresponding to the model pattern input from the keyboard 112 by the user.

  Then, the pattern vector creating unit 201 assigns a label to the created pattern vector, and stores the pattern vector in the category configuration model data database 115 in a state of being classified for each category input by the user (S303).

  Next, the number-of-classes / dimensions determination unit 202 refers to the category configuration model data database 115, extracts a pattern vector for each of the classified categories, and sets the number of classes and the number of dimensions that can be easily classified based on the distribution of the pattern vectors. Is calculated by the AIC method and stored in the class number / dimension number memory 120 (S304).

  Next, the class configuration construction unit 203 extracts the number of classes and the number of dimensions from the number-of-classes / dimensions memory 120, obtains the class configuration that minimizes the intra-class variance by the ISODATA method, and classifies the pattern vectors into classes. . Then, a label of the class classified based on the obtained class configuration is attached to the pattern vector, and the pattern vector is stored in the class configuration model data database 116 (S305).

  Next, the feature extraction matrix calculation unit 204 calculates a feature that the pattern vectors included in the class are most united from each pattern vector stored in the class configuration model data database 116, and further, the class is most separated. Calculate the extraction matrix. Then, the feature extraction matrix calculation unit 204 stores the calculated feature extraction matrix in the feature extraction matrix memory 121 (S306).

  Next, the model feature vector calculation unit 205 converts the pattern vector stored in the class configuration model data database 116 into a model feature vector using the feature extraction matrix obtained by referring to the feature extraction matrix memory 121 (S307). ).

  Next, the model feature vector calculation unit 205 calculates the average value of the model feature vector for each class, and registers the average value of the model feature vector for each class in the model feature vector database 122 (S308).

  The above is the process executed by the object recognition device offline.

  Next, a detailed description of each functional block will be described using mathematical expressions.

  First, the pattern vector creation unit 201 will be described.

  The pattern vector creation unit 201 extracts a model pattern from model data and creates a pattern vector.

  The size and the shape of the model pattern may be any shape, such as a vectorized edge image, a vector expressing coordinate information of a binary image in a one-dimensional vector, or model data as a model pattern. .

Next, the pattern vector creation unit 201 creates a label describing the category name from the category information corresponding to the model pattern input from the keyboard 112.

  Note that the user may view the corresponding model pattern displayed on the display 112 and input category information corresponding to the model pattern.

  The label is information expressing the type, affiliation, and the like of the model pattern.

  The category information input by the user is the category information required by the application of the object recognition device.

  Then, the pattern vector creation unit 201 assigns a label to the pattern vector and stores the label in the category configuration model data database 115.

  In the first embodiment, the label indicating the category of the pattern vector is assigned after being classified into the category. However, the label indicating the category may be added to the model data from the beginning. In this case, the pattern vector creating unit 201 classifies the pattern vector into a category by referring to the category label, and stores the category in the category configuration model data database 115.

  Next, the configuration of the category configuration model data database 115 will be described with reference to FIG. FIG. 4 is a diagram showing a configuration of the category configuration model data database according to the first embodiment.

  As shown in FIG. 4, the category configuration model data database 115 stores a plurality of pairs of a category 401 required by an application and model data corresponding to the category 401.

  Note that the model data stored in the category configuration model data database 115 is a pattern vector.

  The categories required by the application include categories such as "Mr. A", "Mr. B", and "Mr. C" indicating a person, and categories such as "Sedan", "Truck", and "RV" indicating a car type. And various categories such as information indicating the distance to the object. The model data corresponding to the category 401 is provided with a label 402 having the name of the category 401.

  In the example of FIG. 4, the label 402 is indicated by 3 bits, the label 402 of “001” is assigned to the images a1 and a2 of the category A, the label 402 of “010” is assigned to the images b1 and b2 of the category B, The label 402 of “011” is attached to the header part of the image data on the images c1 and c2 of C.

  Next, the class number / dimension number determining unit 202 will be described.

  The number-of-classes / dimensions determination unit 202 refers to the category configuration model data database 115, extracts a pattern vector for each category, calculates the number of classes and the number of dimensions, and stores it in the class number / dimension number memory 120.

Specifically, the number of classes and the number of dimensions are calculated using AIC (Akaike Information criterion), which is an example of an information amount criterion that is a criterion for model selection when estimating a certain model probability distribution from given data. It is possible. AIC
Is described in Yoshiyuki Sakamoto, Makio Ishiguro, Genshiro Kitagawa, "Information Statistics", First Edition, Kyoritsu Shuppan, 1983, p42. The AIC is represented by (Equation 1).

  Assuming that the model probability distribution represented by the coordinate system of the pattern vector is a mixture distribution based on a multidimensional normal distribution, it is represented by (Equation 2).

Here, x is, shows a pattern vector, n _i denotes the number of samples of the class i, K denotes the number of classes, d represents the number of dimensions.

  The number of classes and the number of dimensions, which are parameters of the model probability distribution in the coordinate system of the pattern vector estimated so that the AIC of (Equation 2) is minimized, are the values of the number of classes and the number of dimensions. What is necessary is just to store it in the memory 120.

  In this example, a mixture distribution based on a multidimensional normal distribution is assumed as the model probability distribution, but may be assumed in accordance with model data such as a Poisson distribution or a polynomial distribution.

  In the first embodiment of the present invention, the number of classes and the number of dimensions are obtained. However, the same effect can be obtained with only the number of classes.

  Next, the class configuration construction unit 203 will be described.

  The class configuration constructing unit 203 extracts the number of classes and the number of dimensions, obtains a class configuration by the ISODATA method, assigns a label indicating the class to the pattern vector, classifies the class, and stores it in the class configuration model data database 116.

  Here, the class label describes the class number and the class name, which are information for identifying the class to which the model data belongs, and the contents of the items used for constructing the class configuration. At this time, the class configuration construction unit 203 may calculate a representative value of each class and store the representative value in the class configuration model data database 116.

  It should be noted that the class label may be configured to describe only the information to which the class to which it belongs, for example, only the class number, and to add the information after the label indicating the category.

  As described above, a plurality of sets of the class number indicating the class type, the category corresponding to the class and the item used for constructing the class configuration, and the model data classified into the class are stored in the class configuration model data database 116. ing.

  Here, the representative value of the class may be stored and used for calculating the representative value of the model feature vector later.

  Here, the ISODATA method used in the class configuration construction unit 203 is described in Kenichi Mori, “Pattern Recognition”, Third Edition, IEICE, 1988, p122. More specifically, the ISODATA method is a clustering method for reducing cluster variance by repeatedly dividing and merging clusters.

  Note that the class configuration construction unit 203 employs the ISODATA method as a clustering method, but may use a clustering method other than the ISODATA method, for example, a K-mean method.

  Next, the feature extraction matrix calculation unit 204 will be described.

  The feature extraction matrix calculation unit 204 calculates, from the pattern vectors of each class, a feature extraction matrix in which the pattern vectors included in the class are the most integrated and the class is most separated, and stored in the feature extraction matrix memory 121. Store.

Note that, in the first embodiment, the feature extraction matrix A ^T is defined as Fisher. R. A. , "The use of multiple measurements in taxonomic problems", Ann. Eugenics. 7. Part II, pp. 179-188, 1936. ) Was calculated according to the concept of Fisher's discriminant function.

Here, the feature extraction matrix A ^T, the coordinate system is a space representing the pattern vector created from the model data, between classes of classes obtained from the model data distribution C _B is as large as possible, and within-class variance C _W is This is a matrix for conversion into a model feature vector that is a feature coordinate system having features that are as small as possible.

First, the feature extraction matrix calculation unit 204 refers to the class configuration model data database 116 and extracts a set of pattern vectors classified into the same class. Next, the feature extraction matrix calculation unit 204 calculates the in-class covariance matrix _CW from the set of pattern vectors according to (Equation 3) for each class.

Next, the feature extraction matrix calculation unit 204 calculates an inter-class covariance matrix C _B from each set of pattern vectors according to (Equation 4).

Further, the feature extraction matrix calculation unit 204 calculates the covariance matrix C _T of all classes according to (Equation 5).

Here, the feature extraction matrix calculating section 204, as a discrimination criterion in accordance with the idea of Fisher, as large as possible a dispersion in which the inter-class covariance matrix C _B between classes, intraclass covariance is the variance of pattern vectors simultaneously in class so that only smaller possible covariance matrix C _W, obtains the transformation vector a such that the variance ratio of the inter-class covariance matrix C _B and within-class variance matrix C _W (number 6) to a maximum.

A case where the category 900 is composed of three classes will be described with reference to FIG. The intra-class variance 902 converted into a one-dimensional feature coordinate system by multiplying the intra-class variance 901 of each class, which is the variance of the pattern vector displayed in two dimensions, by a feature extraction matrix is small.
And by increasing the inter-class covariance matrix C _B, distribution region between classes leaves.

  That is, since the feature coordinate system moves by adjusting the feature extraction matrix, the one-dimensional intra-class variance 902 converted into the feature coordinate system can be adjusted, and the features of the class can be clarified. .

  Further, finding the transformation vector a that maximizes (Equation 6) is equivalent to solving the eigenvalue problem (Equation 7).

Therefore, the feature extraction matrix calculation unit 204 obtains ^AT obtained by transposing A from (Equation 7), and stores the obtained ^AT as a feature extraction matrix in the feature extraction matrix memory 121.

  Next, the model feature vector calculation unit 205 will be described.

  The model feature vector calculation unit 205 converts the pattern vector into a model feature vector using the feature extraction matrix, and registers the average value of the model feature vector for each class in the model feature vector database 122 as a representative value.

  Note that the representative value of the model feature vector for each class may be a value other than the average value, for example, an intermediate value.

  The model feature vector data database 122 stores a plurality of pairs of classes, model feature vectors of model data corresponding to the classes, and representative values of model feature vectors in the classes. However, the model feature vector of the model data can be omitted.

  Further, the data processing apparatus 100 can repeatedly calculate the number of classes and the number of dimensions a plurality of times.

  FIG. 6 shows an operation flowchart of the offline processing when the feature extraction matrix is repeatedly calculated a plurality of times.

  After processing the step of S303 shown in FIG. 3, the number of classes and the number of dimensions are set to the same number as the categories classified as model data, and the process proceeds to S305 (S1001).

  Next, the processes from S305 to S307 are performed, and the threshold value obtained from the data to be classified is compared with the AIC value obtained from the model feature vector, and the execution of the process for obtaining the number of classes and the number of dimensions is determined again. However, when the number of times of processing for obtaining the number of classes and the number of dimensions is the first time, the process may shift to S1003 without comparing the threshold value and the AIC value (S1002).

  Next, the first AIC value is calculated from (Equation 2) without changing the number of classes and the number of dimensions from the distribution state of the model feature vector calculated in S1002.

  Next, the second AIC value is calculated by increasing the number of classes and the number of dimensions. Here, the increase in the number of classes and the number of dimensions is determined by the data to be classified.

  Here, the first AIC value is compared with the second AIC value, and if the second AIC value exceeds the first AIC value, the number of classes and the number of dimensions to be obtained are changed to the number of classes for which the first AIC value is obtained. If the first AIC value and the second AIC value are the same or the first AIC value exceeds the second AIC value, the number of classes and the number of dimensions are determined to be the second AIC value. Number of classes.

  That is, when the number of classes and the number of dimensions are increased when the number of classes and the number of dimensions are obtained, it is desirable that the calculated AIC is reduced (S1003).

  Next, the processes of S305 to S307 are executed using the number of classes and the number of dimensions obtained in S1003, and in S1002, the threshold determined by the data to be classified is compared with the model feature vector obtained in S307. If it is, the process proceeds to S308, and if it is equal to or more than the threshold, the process proceeds to S1003, and the number of classes and the number of dimensions are calculated again (S1002).

  As described above, according to the first embodiment, an easily separable class configuration in which model data previously classified for each category is set to the number of classes and dimensions according to the shape of the category by offline processing Can be reconstructed.

  Furthermore, the class configuration that can be easily separated can be compared with the image data and the class while the distribution of the model data in the class is reduced and the class distribution between the classes is increased. That is, since the comparison between the image data and the class can be performed in a state where the characteristics of the distribution of the class are clear, the comparison between the image data and the class can be accurately performed.

  Next, online processing of the object recognition device will be described. The purpose of the online processing is to determine the class to which the image data is most similar and to recognize the objects included in the image data.

  First, a configuration for performing online processing of the object recognition device will be described with reference to FIGS.

  FIG. 7 is a functional block diagram for online processing of the object recognition device according to the first embodiment.

  Here, reference numeral 501 denotes an input feature vector calculation unit input that converts an input pattern vector into an input feature vector using a feature extraction matrix and stores the input feature vector in the feature vector memory 121, and 502 denotes an input feature vector and a model feature vector. 503 shows an input pattern vector determining unit that calculates a similarity by comparing a representative value of the input pattern vector with a representative value of the input pattern vector. Indicates a part.

  The online processing includes an I / F 105, an image memory 111, a pattern vector memory 114, a feature extraction matrix memory 121, a model feature vector database 122, a feature vector memory 123, a pattern vector creation unit 201, an input feature vector calculation unit 501, an input A pattern vector determining unit 502 and an input data determining unit 503 are configured. The pattern vector creating unit 201, the input feature vector calculating unit 501, the input pattern vector determining unit 502, and the input data determining unit 503 are configured such that the CPU 102 is stored in the memory 103. This is done by loading a program.

  Next, FIG. 8 shows an operation flowchart of online processing of the object recognition device according to the first embodiment of the present invention.

  First, the relationship between the functional blocks and the operation flowchart will be described, and then a detailed description of each functional block will be described.

  First, the object recognition device inputs image data to be recognized from the video camera 110. Next, the object recognition device sends the image data to the image memory 111 via the I / F 105 and stores the image data (S601).

  Next, the pattern vector creating unit 201 extracts a part of the data from the image data once stored in the image memory 111, and furthermore, an input pattern in which the pixel values of the part of the data are expressed as a one-dimensional data sequence. A vector is created and stored in the pattern vector memory 114 (S602).

  Next, the input feature vector calculation unit 501 calculates an input feature vector using the input pattern vector stored in the pattern vector memory 114 and the feature extraction matrix stored in the feature extraction matrix memory 123 and obtained during off-line ( S603).

  Then, the input feature vector calculation unit 501 outputs the input feature vector to the input pattern vector determination unit 502.

  Next, the input pattern vector determining unit 502 extracts the representative value of each class stored in the model feature vector database 122, and calculates the similarity between the input feature vector and the representative value of each class (S604).

  Next, the input data determination unit 503 compares the similarity between the input feature vector determined by the input pattern vector determination unit 502 and the representative value of each class, and determines the class with the highest similarity as a recognition result ( S605).

  Next, the input data determination unit 503 determines a category to which a class having the highest similarity to the input pattern vector belongs, and recognizes an object included in the image data (S606).

  Note that it is also possible to further search in the category by using image recognition or the like.

  The above is the process executed by the object recognition device online.

  Next, a detailed description of each functional block will be described.

  First, the pattern vector creation unit 201 will be described.

  The pattern vector creation unit 201 creates an input pattern vector from the image data and stores it in the pattern vector memory 114.

  When the pattern vector of the model data stored in the class configuration model data database 116 is created by a different method for each class, the pattern vector creation unit 201 determines the pattern determined for each class from the image data. Create an input pattern vector according to the vector creation method.

Further, the pattern vector creation unit 201 may store the representative value of the coordinates of the input pattern vector in the image, for example, the coordinates of the center point of the image data in the image in the pattern vector memory 114.

  Next, the input feature vector calculation unit 501 will be described.

The input feature vector calculation unit 501 calculates an input feature vector using (Equation 8) from the input pattern vector and the feature extraction matrix ^AT (S603).

Here, b is an input pattern vector, A ^T is a feature extraction matrix, and f is an input feature vector.

  Then, the input feature vector calculation unit 501 stores the input feature vector in the feature vector memory 121.

  Next, the input pattern vector determining unit 502 will be described.

The input pattern vector discriminating unit 502 calculates the similarity between the input feature vector and the representative value of each class. The similarity between the image data and each class is calculated by the Euclidean distance between the input feature vector and the representative value of each class. Deciding. The similarity increases as the Euclidean distance between the input feature vector and the representative value of each class decreases.

  Note that another index such as a correlation distance may be used as the similarity.

  Next, the input data determination unit 503 will be described.

  The input data determination unit 503 determines the class with the highest similarity as a recognition result, determines the category to which the class with the highest similarity to the input pattern vector belongs, and determines the class, the type in the class, the category, and the input pattern vector, Output from the output terminal 118.

  Note that the data processing apparatus 100 can increase the data classification accuracy by repeatedly performing the above-described offline and online processing.

  That is, when the input data determination unit 503 determines that there is a plurality of categories to which the image data belongs in the online processing, the processing of the class number / dimension number determination unit 202 and the class configuration construction unit 203 is re-executed to change the class number. Then, by classifying the model data, the distribution of the classes is reconstructed into a class configuration that can be classified.

  Note that a plurality of images photographed at different distances are employed as model data, and distances are employed as categories. By recognizing a category having the highest similarity to the input image data, the real space of the input image data is recognized. It can be applied to a distance detection system that recognizes the distance above

  Also, by changing the type of image data and model data, the object recognition device can be applied to gate systems, iris authentication systems, fingerprint authentication systems, character recognition systems, human detection systems, car detection systems, car recognition systems, etc. it can.

  In the first embodiment, the image data is classified using the image data as the sequentially input data. However, the sequentially input data is divided into the concept of a category such as audio data. Any data may be used as long as it is data for performing recognition processing as to which category it belongs to.

  Further, in the first embodiment, the model data is stored in the category configuration model data database 115 and the class configuration model data database 116 after the model data is converted into a pattern vector. The model data may be stored as it is.

  In this case, as necessary, a model pattern or model data is extracted from the category configuration model data database 115 and the class configuration model data database 116 to create a pattern vector.

  In the first embodiment, a matrix considering the intra-class variance and the inter-class variance is used as the feature extraction matrix. However, any matrix may be used as long as the matrix is converted into a coordinate system in which the feature of each class clearly appears. It may be.

  For example, a matrix that is transformed into a coordinate system that performs a principal component analysis method, a subspace method, or the like may be used as the feature extraction matrix. The coordinate system for performing the principal component analysis method is a coordinate system of an eigenspace that widens the interval between model data that is sample data.

  Further, the coordinate system for performing the subspace method is a coordinate system in which the dimensions of the coordinate system are reduced and the features of the class are expressed. The principal component analysis method and the subspace method are disclosed in Kenichi Mori, “Pattern Recognition”, Third Edition, IEICE, 1988, p13.

  As in the present embodiment, as an off-line process, a category is formed into a pattern vector, an information amount criterion such as AIC is used, a class is formed, and a pattern vector included in each class is most united. A feature extraction matrix that can be separated most is calculated, converted into a feature coordinate system of a pattern vector in a class unit, a class representative value is calculated, and, as online processing, data sequentially input from the video camera 110 and offline processing are performed. By comparing the representative values of the obtained feature vectors, determining the corresponding class, and classifying the data that is sequentially input, model data classified in advance into each category can be obtained by using only the categories. Classify the data that cannot be classified into the most appropriate category by classifying the data into more detailed classes without redundancy. It can be.

  Therefore, it is possible to construct a class configuration that makes it easy to classify data by finding the number of classes and dimensions that match the characteristics of the data, such as the shape of the category distribution, the degree of variance, the presence or absence of noise data, and the degree of noise data. The effect is obtained.

(Embodiment 2)
In the offline processing, the object recognition device according to the first embodiment constructs a class using model data of a category as a pattern vector, converts the coordinates of the pattern vector classified into the class, and calculates a representative value of the class. In the online processing, classification of input data is executed, whereas in the offline processing, the object recognition device according to the second embodiment converts the coordinates of the pattern vector obtained from the model data of the category and converts the class into the class. It constructs, calculates a representative value of a class, and executes classification of input data in online processing.

Hereinafter, the operation of the object recognition device according to the second embodiment will be described with reference to FIGS. Since the difference from the first embodiment is the offline processing, the online processing is omitted.

  FIG. 9 illustrates a class number / dimension number determining unit 801 that determines the class number and the dimension number for constructing a class from the model probability distribution of the coordinate-transformed pattern vector, and 802 obtains from all model data. A class configuration constructing unit that transforms the coordinates of a pattern vector and constructs a class from a category using the number of classes and the number of dimensions is shown. A feature extraction matrix calculation unit 803 calculates a feature extraction matrix for calculating a representative value of the class. Indicates a part.

  Further, the offline processing of the object recognition apparatus according to the second embodiment includes the I / Fs 104 and 106, the model data memory 109, the keyboard / display 112, the category configuration model data database 115, the class configuration model data database 116, It comprises a dimension number memory 120, a feature extraction matrix memory 121, a pattern vector creation unit 201, a model feature vector calculation unit 205, a class number / dimension number determination unit 801, a class configuration construction unit 802, and a feature extraction matrix calculation unit 803.

  The offline processing of the object recognition apparatus will be described with reference to the offline processing operation flowchart of FIG.

  First, the object recognition device performs the processing of S301 to S303 as in the first embodiment, and builds the category configuration model data database 115.

Next, the class configuration construction unit 203 outputs pattern vectors of all model data from the category configuration model data database 115, and performs covariance of all pattern vectors using (Equation 5) in order to perform coordinate transformation of the pattern vectors. A matrix C _T is obtained (S701).

Next, the class configuration constructing unit 802 converts the coordinate system into a coordinate system in which the covariance matrix _CT is a unit matrix in order to enable the calculation of the feature extraction matrix ^{AT at} the subsequent stage. Specifically, the class configuration construction unit 802 converts the coordinate system into a coordinate system in which the covariance matrix _CT is a unit matrix using the matrix U shown in (Equation 9) (S702).

  Here, U indicates a transformation matrix, and I indicates a unit matrix.

  Next, the class number / dimension number determination unit 801 estimates the model probability distribution that minimizes the AIC in the coordinate system transformed by the matrix U, as in the first embodiment, and determines the class number and the dimension number. It is obtained (S703).

Next, the class configuration constructing unit 802 generates a class configuration that minimizes the intra-class variance C _W in the coordinate system transformed by the matrix U according to the number of classes and the number of dimensions obtained by the number-of-classes / dimensions determining unit 801. , By the ISODATA method.

  Next, the class configuration constructing unit 802 creates the class configuration model data database 116 based on the obtained class configuration (S704).

Here, since the covariance matrix _CT is constant, a class configuration that minimizes the intra-class variance C _W is obtained from (Equation 10), thereby obtaining a class configuration that maximizes the determination criterion J _1. .

Thus, the object recognition device is capable of calculating a discrimination criterion J ₁ is calculated, and the discrimination standard J ₁ By maximizing feature extraction matrix A ^T shown in equation (10) (S705).

  Thereafter, the object recognition apparatus performs the processing of S307 and S308 based on the obtained class configuration model data database 116 in the same manner as in the first embodiment, and obtains a representative of each class from the model feature vector that can perform class identification with high accuracy. The value can be calculated.

Note that the feature extraction matrix ^AT can also be calculated using (Equation 3) to (Equation 8), as in the first embodiment.

  As in the first embodiment, as shown in FIG. 11, the number of classes and the number of dimensions are calculated again by comparing the threshold value obtained from the data to be classified with the AIC value obtained from the model feature vector. It is also possible to determine the execution of the process.

  FIG. 11 is a flowchart illustrating the operation of the offline processing of the object recognition device according to the second embodiment. The operations in steps S301 to S303, S307, S308, S701, S702, S704, and S705 are the same as those in FIG. 6, and a description thereof will be omitted.

First, the processes from S301 to S702 are performed to convert the coordinate system into a coordinate system in which the covariance matrix _CT is a unit matrix. Then, in S1101, the number of classes and the number of dimensions are changed to the category in which the model data is classified. The number is determined to be the same, and the process proceeds to S305.

  Next, the processes of S704, S705, and S307 are performed, and in S1102, the threshold value obtained from the data to be classified is compared with the AIC value obtained from the model feature vector, and the process of obtaining the number of classes and the number of dimensions is performed again. decide. However, when the number of processes for obtaining the number of classes and the number of dimensions is the first time, the process may proceed to S1103 without comparing the threshold value and the AIC value.

  Next, in S1103, the first AIC value is calculated from (Equation 2) without changing the number of classes and the number of dimensions from the calculated distribution state of the model feature vector. Next, the second AIC value is calculated by increasing the number of classes and the number of dimensions. Here, the increase in the number of classes and the number of dimensions is determined by the data to be classified.

  Here, the first AIC value is compared with the second AIC value, and if the second AIC value exceeds the first AIC value, the number of classes and the number of dimensions to be obtained are changed to the number of classes for which the first AIC value is obtained. If the first AIC value and the second AIC value are the same or the first AIC value exceeds the second AIC value, the number of classes and the number of dimensions are determined to be the second AIC value. The class number is set, and the process proceeds to S704.

  In other words, when calculating the number of classes and the number of dimensions, when the number of classes and the number of dimensions are increased, it is desirable that the calculated AIC is reduced.

As described above, as in the present embodiment, the covariance matrix C _T is obtained from the pattern vector of the model data, and the optimal number of classes is determined from the model data converted into a coordinate system in which the covariance matrix C _T is a unit matrix. By classifying the model data into classes and determining the representative value of each class, the class to which the sequentially input data belongs can be determined in the online processing, and the category can be further determined.

  Therefore, it is possible to construct a class configuration that makes it easy to classify data by finding the number of classes and dimensions that match the characteristics of the data, such as the shape of the category distribution, the degree of variance, the presence or absence of noise data, and the degree of noise data. The effect is obtained.

  As described above, the evaluation experiment of the face image recognition was performed using the object recognition device of the present invention. The experimental subjects were 23 persons, the pixel data was 9 × 12 pixels, and 256 gradations. The model data is composed of about 620 face images taken under five different lighting conditions and about 620 face images taken under random lighting conditions two years later. However, 500 different face images were selected.

  Here, the category is set to 23, the number of subjects, and the model data is selected from specific image conditions of 620 pieces of image data, subjected to offline processing according to the embodiment of the present invention, and constructed an optimal class.

  Next, as online processing, object recognition was performed using, as input data, image data excluding lighting conditions of the model data used in the online processing. Here, since the category is the same as the number of subjects, the object is recognized by classifying the input data into classes.

  As a result of performing the above-described face image recognition, in the conventional object recognition apparatus, the misclassification rate, which is the probability of misclassification, was 18.5%. In the recognition device, the misclassification rate was 5.4%, which was an improvement of about 13% in accuracy.

  In addition, when image data having different photographing directions was used as model data, the accuracy was improved from 10.1% to 2.5%.

  The present invention can be easily implemented by another independent computer system by realizing the program, recording the program on a recording medium such as a flexible disk, and transferring the program.

  The recording medium can be similarly implemented as long as it can record a program, such as an optical disk, an IC card, and a ROM cassette.

(Embodiment 3)
The object recognition device according to the third embodiment constructs a class using an evaluation criterion including a similarity between two distributions of a feature vector distribution calculated from a plurality of model data and a category or class distribution. Is what you do.

  Hereinafter, the operation of the object recognition device according to the third embodiment will be described with reference to FIG. The difference from the first embodiment is the offline processing, and the online processing will be described later.

Here, the class number / dimension number determination unit 1101 classifies the categories into fine classes without redundancy using an evaluation criterion including a similarity between the distribution of the model data and the distribution of the categories or classes. A pattern vector is extracted for each classified category from the constituent model data database 115, and evaluation criteria are optimized so that classification can be easily performed based on the distribution of the pattern vectors. A feature space based on the Fisher criterion and an AIC This shows a method for simultaneously optimizing the number of classes and the number of dimensions based on, and the class configuration determined from these.

  The I / Fs 104 and 106, the model data memory 109, the keyboard / display 112, the category configuration model data database 115, the class configuration model data database 116, the number of classes / dimensions memory 120, the feature extraction matrix memory 121, the pattern vector creation unit 201, the class configuration construction unit 203, the feature extraction matrix calculation unit 204, and the model feature vector calculation unit 205 have the same operations as those in the first embodiment, and thus description thereof will be omitted.

  In order to find a compromise between the Fisher criterion and the AIC criterion, a criterion evaluation criterion FAIC shown in (Expression 11) is provided, and a class configuration and a feature space that minimize this are determined. The evaluation criterion FAIC is an expression obtained by adding a Kullback-Leibler information amount (KL information amount) obtained by multiplying the AIC by the weighting coefficient. Here, the KL information amount is used as a criterion evaluation criterion for measuring the closeness of distribution between the distribution determined by the AIC and the class configuration given when the feature space is optimized based on the Fisher criterion.

  Here, the distribution of the feature vector calculated from the plurality of model data corresponds to the distribution of the feature vector group on the feature space based on the Fisher criterion, and the distribution of the category or class corresponds to the distribution on the feature space based on the Fisher criterion. , The distribution of each feature vector group in each category or class corresponds to the similarity, and the KL information amount corresponds to the similarity.

  At this time, the density function of the class configuration of the sample on the feature space optimized based on the Fisher criterion is defined as a model distribution f (x), and the density function of the distribution determined by the AIC is defined as a true distribution g (x). For example, at this time, it is assumed that the distribution f (x) of the class configuration of the sample also forms a mixture normal distribution on the optimal feature space.

  From the above, by minimizing the evaluation criterion FAIC, it is possible to obtain the number of classes and the number of dimensions that minimize the AIC while having a distribution similar to the class configuration of the sample derived from the Fisher criterion.

  Note that the class configuration method of the object recognition device according to the third embodiment can be applied to the object recognition device according to the second embodiment. FIG. 13 shows a configuration diagram.

  The number-of-classes / dimensions determining unit 1301 shown in FIG. 13 optimizes an evaluation criterion including the similarity between the distribution of the coordinate-transformed pattern vector and the distribution of the category or the class, and is the same as in the second embodiment. In the coordinate system converted by the matrix U, the number of classes and the number of dimensions for minimizing the FAIC are obtained.

  Next, the flow of operation of the object recognition device according to the third embodiment in the offline processing will be described with reference to FIG.

  Here, FIG. 14 is a flowchart showing the operation of the offline processing by the object recognition device having the configuration of FIG.

  In the first embodiment, in the flowchart shown in FIG. 6, the step based on the Fisher criterion (S306) and the step of determining the number of classes and the dimension of the feature space using the AIC (S1003) are alternately repeated. Process.

  On the other hand, in the third embodiment, a process based on the Fisher criterion (S306) and a process of determining the number of classes and the dimension of the feature space using FAIC (S1302) are alternately repeated. ing.

  Here, the operations in S301 to S307 and S1000 are the same as those in FIG.

  First, the processes from S301 to S307 are performed, the model feature vector calculation unit 205 calculates the model feature vector, and then the class number / dimension number determination unit 1101 determines the FAIC value (first The FAIC value is compared with the threshold, and if it is equal to or greater than the threshold, the flow shifts to S1302, and if it is less than the threshold, to S308. In addition, the first time may be performed repeatedly (S1301).

  Next, in S308, the model feature vector calculation unit 205 calculates the average value of the model feature vector for each class, and registers the average value of the model feature vector for each class in the model feature vector database 122.

  When the process proceeds to S1302, the class number / dimension number determining unit 1101 increases the class number and the dimension number from the calculated distribution state of the model feature vector, and uses the (Equation 11) to perform the second Calculate the FAIC value. Here, the increase in the number of classes and the number of dimensions is determined by the data to be classified.

  Here, the first FAIC value and the second FAIC value are compared, and if the second FAIC value exceeds the first FAIC value, the number of classes and the number of dimensions are determined by the number of classes for which the first FAIC value is determined. If the first FAIC value and the second FAIC value are the same, or the first FAIC value exceeds the second FAIC value, the number of classes and the number of dimensions are calculated as the second FAIC value. Number of classes.

  That is, when calculating the number of classes and the number of dimensions, if the number of classes and the number of dimensions are increased, it is desirable that the calculated FAIC be reduced.

  Next, the processes of S305 to S307 are executed again using the number of classes and the number of dimensions obtained in S1302.

  FIG. 15 is a flowchart showing the operation of the offline processing by the object recognition device having the configuration shown in FIG.

  First, similarly to FIG. 11, the process of S307 is performed, and in S1401, the threshold value determined based on the data to be classified is compared with the FAIC value determined from the model feature vector, and the process of calculating the number of classes and the number of dimensions is performed again. To determine. However, when the number of times of processing for obtaining the number of classes and the number of dimensions is the first time, the process may shift to S1402 without comparing the threshold and the FAIC value.

Next, in S1402, the first FAIC value is calculated from (Equation 11) without changing the number of classes and the number of dimensions from the calculated distribution state of the model feature vector. Next, the number of classes and the number of dimensions are increased, and a second FAIC value is calculated. Here, the increase in the number of classes and the number of dimensions is determined by the data to be classified.

  Here, the first FAIC value and the second FAIC value are compared, and if the second FAIC value exceeds the first FAIC value, the number of classes and the number of dimensions are determined by the number of classes for which the first FAIC value is determined. If the first FAIC value and the second FAIC value are the same, or the first FAIC value exceeds the second FAIC value, the number of classes and the number of dimensions are calculated as the second FAIC value. The class number is set, and the process proceeds to S704.

  That is, when calculating the number of classes and the number of dimensions, if the number of classes and the number of dimensions are increased, it is desirable that the calculated FAIC be reduced.

  With the above configuration, the image recognition device according to the third embodiment optimizes one evaluation criterion by using (Equation 11) as compared with the image recognition devices according to the first and second embodiments, It is possible to obtain a feature space that can accurately classify the classes at once, and a class configuration based on the number of classes and dimensions.

  The data classification device, the object recognition device, the data classification method, and the object recognition method according to the present invention determine the number of classes and the number of dimensions so that class classification can be performed accurately, and then determine the number of classes and the number of dimensions. By obtaining the class configuration according to the above, the class having the highest similarity to the sequentially input data can be detected, and the category to which the detected class belongs can be obtained. Has the advantageous effect of being able to reduce the probability of misclassifying data in conventional ways, and to reduce the category in which the data is classified A data classification device and a data classification method for constructing a class having a small distribution, and classify input data into a class constituted by the data classification device Useful objects contained in the data as an object recognition device and an object recognition method or the like to recognize.

FIG. 1 is a diagram illustrating a configuration of an object recognition device according to a first embodiment of the present invention. The figure which shows the structure of the offline processing of the object recognition apparatus by the 1st Embodiment. Flowchart of offline processing of the object recognition device according to the first embodiment FIG. 4 is a diagram showing a category configuration model data database according to the first embodiment. FIG. 7 is a diagram showing coordinate conversion of a pattern vector according to the first embodiment. Flowchart of offline processing of the object recognition device according to the first embodiment The figure which shows the structure of the online processing of the object recognition apparatus by 1st Embodiment Flow chart of online processing of the object recognition device according to the first embodiment The figure which shows the structure of the offline processing of the object recognition apparatus by the 2nd Embodiment. Flowchart of offline processing of the object recognition device according to the second embodiment Flowchart of offline processing of the object recognition device according to the second embodiment The figure which shows the structure of the offline processing of the object recognition apparatus by the 3rd Embodiment. The figure which shows the structure of the offline processing of the object recognition apparatus by the 3rd Embodiment. Flowchart of offline processing of the object recognition device according to the third embodiment Flowchart of offline processing of the object recognition device according to the third embodiment

Explanation of reference numerals

REFERENCE SIGNS LIST 100 object recognition device 101 computer system 102 CPU
108 Image Database 110 Video Camera 112 Keyboard / Display 114 Pattern Vector Memory 115 Category Configuration Model Data Database 116 Class Configuration Model Data Database 120 Class / Dimension Memory 121 Feature Extraction Matrix Memory 122 Model Feature Vector Database 123 Feature Vector Memory 201 Pattern Vector Creation unit 202, 801, 1201, 1301 Class / dimension determination unit 203, 802 Class configuration construction unit 204, 803 Feature extraction matrix calculation unit 205 Model feature vector calculation unit 501 Input feature vector calculation unit 502 Input pattern vector discrimination unit 503 Input data judgment unit

Claims (10)

A class number / dimension number determination unit for calculating the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution smaller than the category into which the plurality of model data are classified, and the number of classes and the number of dimensions And a class configuration constructing unit that classifies the model data into the class, and converts the model data classified into the class into a distribution with small intra-class variance and large inter-class variance. Data having a feature extraction matrix calculation unit for calculating a feature extraction matrix used in the case, and a model feature vector calculation unit for calculating an average value of feature vectors for each class from the model data classified into the class and the feature extraction matrix. Classifier. The number-of-classes / dimensions determination unit calculates the number of classes and the number of dimensions of the class configuration that minimizes the information amount criterion that is a criterion for model selection using a model probability distribution estimated from a plurality of model data, and 2. The data classification device according to claim 1, wherein the constructing unit constructs a class by clustering, which is a technique for reducing intra-class variance, using the number of classes and the number of dimensions having the smallest information amount criterion. The class number / dimension number determining unit converts the plurality of model data using a covariance matrix which is a unit matrix calculated from a plurality of model data, and an information amount criterion using the converted plurality of model data. Is calculated, and the class configuration constructing unit constructs a class by clustering to reduce the intra-class variance by using the number of classes and the number of dimensions in which the information amount criterion is the smallest, and performs feature extraction. The matrix calculation unit calculates a patent extraction matrix used when converting the plurality of model data converted by the covariance matrix of the unit matrix into a distribution having a small intra-class variance and a large inter-class variance of the class. The data classification device according to claim 1, wherein The class number and dimension number determining unit calculates the class number and the dimension number using a judgment evaluation criterion including a similarity between the distribution of the feature vector calculated from the plurality of model data and the distribution of the category or the class. Data classification device as described. The data classification device according to claim 4, wherein the judgment evaluation criterion is a criterion including a Kullback-Leibler information amount. The class configuration constructing unit includes a first evaluation criterion calculated using the model data already classified into the class, the first class number and the dimension number, and the model data already classified into the class and the first evaluation amount criterion. 1 is compared with a second evaluation criterion calculated using a second class number and a dimension number different from the class number and the dimension number, and a class is re-used using the class number and the dimension number having a smaller evaluation criterion. The data classification device according to any one of claims 1 to 5, which is constructed. A data classification device according to any one of claims 1 to 6, an input feature vector calculation unit that converts input data including an object to be recognized into an input feature vector using the feature extraction matrix, An object recognition apparatus comprising: an input data determination unit that recognizes the object by comparing an average value of a feature vector with the input feature vector and obtaining the category to which the class most similar to the input data belongs. A first step of calculating the number of classes and the number of dimensions, which are parameters used when constructing a class in which the categories in which the plurality of model data are classified are smaller than the categories, and using the number of classes and the number of dimensions A second step of calculating a class configuration and classifying the model data into the class; and converting the model data classified into the class into a distribution having a small intra-class variance and a large inter-class variance. A data classification method comprising: a third step of calculating a feature extraction matrix; and a fourth step of calculating an average value of a feature vector for each class from the model data classified into the class and the feature extraction matrix. A first step of calculating the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution smaller than the category in which the plurality of model data are classified, and using the number of classes and the number of dimensions A second step of calculating a class configuration and classifying the model data into the class; and converting the model data classified into the class into a distribution having a small intra-class variance and a large inter-class variance. A third step of calculating a feature extraction matrix; a fourth step of calculating an average value of a feature vector for each class from the model data classified into the class and the feature extraction matrix; and an object to be recognized. A fifth step of receiving input data and converting the input data into an input feature vector using the feature extraction matrix; It compares the average value and the input feature vector of the class, by obtaining the category to which the class belongs to the most similar to the input data, an object recognition method and a sixth step for recognizing the object. A first step of calculating the number of classes and the number of dimensions, which are parameters used when constructing a class having a distribution in which a plurality of model data are classified into categories smaller than the category, and the number of classes and the number of dimensions; A second step of calculating a class configuration and classifying the model data into the class, and converting the model data classified into the class into a distribution having a small intra-class variance and a large inter-class variance. A data extraction method including: a third step of calculating a feature extraction matrix used in the case; and a fourth step of calculating an average value of a feature vector for each class from the model data classified into the class and the feature extraction matrix. The program to execute.

Images