JP2005339186A - Pattern recognition learning device and identification device, pattern recognition learning processing method, identification processing method, pattern recognition program and recoding medium with its program recorded - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To handle with multi-class identification problem beyond 1,000 classes in a realistic processing time, and to realize highly precise identification. <P>SOLUTION: A leaning data inputting part 1 inputs preliminarily prepared learning data, and transmits the learning data to a pre-processing part 2. The pre-processing part 2 operates the normalization of the density values of data or noise removal from the inputted data, and transmits the data to a characteristic extracting part 3. The characteristic extracting part 3 extracts featured values from the data shaped by the pre-processing part 2. An identification function preparing part 4 prepares identification functions based on the extracted featured values, and stores the prepared identification functions in an identification function storing part 5. An image data inputting part 6 inputs image data for identification, and transmits the data through a pre-processing part 7 to a characteristic extracting part 8, and the characteristic extracting part 8 extracts featured values and transmits the values to an identifying part 9. The identifying part 9 reads a parameter stored in the identification function storing part 5, and calculates similarity, and outputs the result to an output part 10. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a pattern recognition learning device and an identification device included in an image, a pattern recognition learning processing method and an identification processing method, a pattern recognition program, and a recording medium on which the program is recorded.

  In recent years, the development of theories related to pattern recognition has been remarkable. Among them, there is a support vector machine (hereinafter referred to as SVM) (see Non-Patent Document 1).

  There is also a kernel nonlinear subspace method (see Non-Patent Document 2).

  This has nothing to do with the speeding up of the computer, and the identification performance is greatly improved by complex calculations that make the most of machine power. However, the calculation in the learning process requires a considerable calculation time even if the fastest existing computer is used.

SVM is considered to be one of the most excellent pattern recognition techniques among the currently known pattern recognition techniques. The SVM is a method for discriminating two classes, and for this purpose, an identification function is calculated using an optimization framework based on a margin maximization criterion.
VN Vapnik: “Statisical Learning Theory”, John Wiley & Sons, 1998 Eisaku Maeda, Hiroshi Murase: “Pattern Recognition by Kernel Nonlinear Subspace Method”, IEICE Transactions D-II, Vol.J82-D-II, No.4, pp600-612,1999

  In calculating the discriminant function described above, the convex quadratic programming problem must be solved, but generally the calculation time for this increases exponentially with the number of learning data.

In addition, when SVM is applied to multi-class identification exceeding 1000 classes such as Japanese character recognition, there are the following two methods.
(1) A method in which all combinations of two classes are made into an identification function, and the class selected most frequently is set as an identification class;
(2) An identification function for identifying a class and other classes is created for the number of classes, and a class having the highest similarity is used as an identification class.

  However, the former method needs to calculate 1000 × 999/2 = 499500 discriminant functions in the case of 1000 classes, for example, and is not very practical. On the other hand, in the case of the 1000 class, since it is sufficient to prepare 1000 discriminant functions, the number of discriminant functions to be calculated is reduced, but there is another problem here.

  In the case of 1000 classes, 999 classes constitute so-called “other” classes. It is clear that the number of data corresponding to this will be enormous. As already described, since the calculation time of SVM increases exponentially with respect to the number of data, the time required to calculate one discriminant function becomes unrealistic. Therefore, when dealing with multi-class identification problems exceeding 1000 classes, SVM is considered to be unsuitable due to the problem of calculation time.

  The present invention has been made in view of the above circumstances, and has discrimination performance comparable to that of SVM and other nonlinear discrimination functions, especially when dealing with a case where a multi-class discrimination problem exceeding 1000 classes is handled. It is an object of the present invention to provide a new pattern identification method that constitutes a classifier having an efficient learning process.

  In order to achieve the above-mentioned problem, the biggest problem in using SVM is the problem of the number of learning data. Generally, 100 to 100,000 pieces of learning data are given to one class. This enormous number of data has a decisive influence on the calculation time. However, if this is replaced with one piece of data for one class, the above problem should be solved.

In order to realize this, the distribution in the feature space of each class is approximated in the subspace, and the mutual similarity between the subspaces is used. Here, mutual subspace method (Kenichi Maeda, Sadaichi Watanabe, “Pattern matching method with local structure”), IEICE Transactions D, Vol.J68-D, No.3, pp.345-352, 1985). In the mutual subspace method, the similarity cos ² between two different subspaces is defined as the following equation.

Here, U and V represent subspaces approximating distributions of two different classes, and u and v represent basis vectors of U and V, respectively. (U, v) represents the inner product of u and v. The minimum angle formed by these subspaces corresponds to θ. The similarity cos ² θ can be calculated as the maximum eigenvalue of Q expressed by the following equations (2), (3), and (4).

(Equation 2) Q = (q _ij ) (2)

Here, u _i is the i-th base vector of the subspace U, v _i is the i-th base vector of the subspace V, u _ia is the a-th component of the i-th base vector of the subspace U, and v _i is a partial The a-th component of the i-th basis vector of the space V, D is the number of dimensions of the feature space, and M is the number of dimensions of the subspace V.

  Also, a nonlinear non-linear mutual subspace method (Nano Sakano, Naoki Takegawa, Taichi Nakamura: “Object recognition by nuclear non-linear mutual subspace method”, IEICE Transactions D-II, Vol.J84D- II, No. 8, pp.1549-1556, 2001) has also been proposed. This is a technique for calculating the similarity between subspaces in a high-dimensional feature space by efficiently calculating the inner product at the point mapped to the high-dimensional feature space by nonlinear transformation using a kernel function. . Specifically, it can be calculated as the maximum eigenvalue of the matrix Q represented by the following equations (5), (6), and (7).

(Equation 5) Q = (q _ij ) (5)

Here, A is a base vector of the subspace U on the high-dimensional feature space, B is a base vector of the subspace V on the high-dimensional feature space, x is a learning vector constituting the subspace U, and y is a subspace. Learning vectors constituting V, k (x, y) is a kernel function, M is the number of dimensions of the subspace V, N is the number of learning vectors x, and N ′ is the number of learning vectors y. Α _i is an eigenvector of the matrix W expressed by the following equations (8) and (9), and corresponds to the i-th eigenvector in descending order of eigenvalues.

(Equation 8) W = (w _ij ) (8)

( _Equation 9) w _ij = k (x _i , x _j ) (9)
Similarly, β _i can be calculated as an eigenvector of the matrix W replaced with the learning vector y. The kernel function includes a polynomial kernel of Expression (10) and a Gaussian kernel of Expression (11).

(Equation 10) k (x _i , x _j ) = (x _i ^t +1) ^P (10)

By using these methods, the similarity between subspaces can be calculated. Here, this similarity is considered as a kernel function k _subspace (U, V) as it is. Using this, the discriminant function s can be expressed by the following equation (12) in the framework of SVM.

Here, y is an identification target vector, and h is a threshold coefficient. U _i represents a subspace of class i, and Support represents a class for which coefficient γ _i is not “0”. T _i represents a correct label of class i and takes a value of “1” or “−1”. For example, when calculating a discriminant function for discriminating a class “1” from other classes, t ₁ may be “1” and t _{2 to} t _c (C is the number of classes) may be “−1”. In addition, k _subspace (U, y) takes a subspace and a vector as arguments, and it may seem that it cannot be calculated by the above formulation. However, as an original definition, k _subspace (U, V) is Since it is Formula (13),

For k _subspace (U, y), cos ² θ derived from the projection angle θ when the vector y is projected onto the subspace U may be calculated. (Note: Here, the base vector of the subspace U is calculated by KL expansion starting from the origin of the feature space.) Therefore, in the case of linearity, it can be calculated as the following equation (14).

Here, u _j represents the j th basis vector of the subspace U.

  In the case of non-linearity, it can be calculated as the following equations (15) and (16).

Here, u _j is an eigenvector of the matrix W expressed by equations (8) and (9), and corresponds to the j-th eigenvector in descending order of eigenvalues. Α _j is the jth largest eigenvalue, and x _{1 to} x _{n ′} are n ′ learning vectors constituting the subspace U.

  And parameter γ is a constraint condition

(Equation 18) 0 ≦ γ _i ≦ η, i = 1,..., C (18)
(However, η is a parameter for soft margin)
Therefore, γ that maximizes the objective function may be obtained by the following equation.

Here, U _i represents a subspace of class i. This can be calculated by the framework of the convex quadratic programming method (Hiroshi Konno, Hiroshi Yamashita “Nonlinear Programming”, Nikka Giren Publisher, 1978).

  When this method is adopted, the number of computations is the same as when the number of learning data per class is considered as 1 when calculating the discrimination function, and as a result, the learning time can be greatly shortened.

  In this method, the subspace of another class in the feature space becomes a support vector, and the positional relationship with the subspace of the own class is reflected in the discriminant function. Therefore, the conventional subspace method or kernel nonlinear subspace method is used. Higher accuracy can be expected.

  In order to achieve the above object, the present invention provides a pattern recognition device for identifying an arbitrary pattern included in image data, comprising learning data input means for inputting image data for learning, and a pattern from the learning data. Learning data feature extracting means for extracting feature quantities; and a discriminant function creating means for creating a subspace that approximates the distribution of the learning data in each class and calculating parameters of the discriminant function using the similarity between the subspaces; It is characterized by providing.

  Alternatively, in the discriminant function creating means, a subspace creating means for calculating a base vector of a subspace approximating the class distribution, and a discriminant function configured based on the similarity between the subspaces by the mutual subspace method Discriminating function parameter calculating means for calculating parameters using a learning process of a support vector machine (SVM).

  Alternatively, the discriminant function creating means is configured based on a subspace creating means for calculating a basis vector of a nuclear nonlinear subspace that approximates a class distribution and a similarity between subspaces by a nonlinear nuclear mutual subspace method. Discriminating function parameter calculation means for calculating parameters of the discriminating function using a learning process of a support vector machine (SVM).

  Alternatively, a pattern recognition apparatus for identifying an arbitrary pattern included in image data, which is an image data input unit that inputs digital image data to be identified, and a feature extraction unit that extracts a feature amount of a pattern from the input image data And an identification means for identifying the pattern based on the identification function read from the identification function storage means and its parameters, and an output means for outputting the identification result.

  Alternatively, a pattern recognition method for identifying an arbitrary pattern included in image data, a learning data input step for inputting image data for learning, and a learning data feature extraction step for extracting pattern feature amounts from the learning data And a discriminant function creating step of creating a subspace approximating the distribution of the learning data in each class and calculating a parameter of the discriminant function using the similarity between the subspaces.

  Alternatively, in the discriminant function creating step, a subspace creating step for calculating a basis vector of a subspace approximating the class distribution and And a discriminant function parameter calculation step for calculating parameters using a learning process of SVM.

  Alternatively, in the discriminant function creating step, a subspace creating step for calculating a basis vector of a nuclear nonlinear subspace that approximates a class distribution and a similarity between subspaces by a nonlinear nuclear mutual subspace method are configured. And a discriminant function parameter calculating step of calculating a discriminant function parameter using an SVM learning process.

  Alternatively, a pattern recognition method for identifying an arbitrary pattern included in image data, an image data input step for inputting digital image data to be identified, and a feature extraction step for extracting a pattern feature amount from the input image data And an identification step for identifying the pattern based on the identification function read from the identification function storage unit and its parameters, and an output step for outputting the identification result.

  Alternatively, the pattern recognition learning device and the identification device, and the pattern recognition learning processing method and the identification processing method are programs for causing a computer to execute.

  Alternatively, the pattern recognition learning device, the identification device, the pattern recognition learning processing method, and the identification processing method are programs for causing a computer to execute, and the program is recorded on a recording medium that can be read by the computer.

  A specific pattern recognition device has the following configuration. That is, a pattern recognition device for identifying an arbitrary pattern included in image data, learning data input means for inputting image data for learning, and learning data preprocessing for performing density value normalization and noise removal of learning data A learning data feature extracting means for extracting feature values of a pattern from learning data, a subspace that approximates the distribution of learning data in each class, and a parameter of an identification function that uses the similarity between the subspaces Discriminant function creating means for calculating discriminant functions, discriminant function storage means for storing discriminant functions and parameters thereof, image data input means for inputting digital image data to be discriminated, density value normalization of input image data, noise removal A pre-processing unit for performing the processing, a feature extracting unit for extracting the feature amount of the pattern from the input image data, and reading from the discrimination function storage unit Comprising identification function and the identification means for performing identification of patterns that parameter on the basis, and output means for outputting the identification result.

  In addition, in the discriminant function creating means, subspace creating means for calculating a base vector of a subspace approximating the class distribution and parameters of the discriminant function configured based on the similarity between the subspaces by the mutual subspace method Discriminating function parameter calculation means for calculating using the learning process of SVM.

  Alternatively, the discriminant function storage means is configured based on the subspace creation means for calculating the basis vector of the nuclear nonlinear subspace approximating the class distribution and the similarity between the subspaces by the nonlinear nuclear mutual subspace method. And a discriminant function parameter calculating means for calculating a discriminant function parameter using a learning process of SVM.

  A specific pattern recognition method has the following configuration. That is, a pattern recognition method for identifying an arbitrary pattern included in image data, a learning data input step for inputting image data for learning, learning data pre-processing for performing normalization of learning data, and noise removal A step, a learning data feature extraction step for extracting feature values of the pattern from the learning data, a subspace that approximates the distribution of the learning data in each class, and a parameter of an identification function that uses the similarity between the subspaces Discriminant function creation step for calculating discriminant function, discriminant function storage step for storing discriminant function and its parameters, image data input step for inputting digital image data to be discriminated, density value normalization of input image data, noise removal And a feature extraction step for extracting pattern features from the input image data. If, comprising an identification step of identifying a pattern based on the parameters and identification functions read from the identification function storing step and an output step of outputting the identification result.

  Also, in the discriminant function storage step, a subspace creation step that calculates a base vector of a subspace that approximates the class distribution, and a parameter of the discriminant function that is configured based on the similarity between subspaces by the mutual subspace method And a discriminant function parameter calculating step for calculating using a learning process of SVM.

  Alternatively, in the discriminant function storage step, the subspace creation step for calculating the basis vector of the nuclear nonlinear subspace that approximates the class distribution and the similarity between the subspaces by the nonlinear nuclear mutual subspace method are used. A discriminant function parameter calculating step of calculating a discriminant function parameter using a learning process of SVM.

  A specific pattern recognition program recording medium has the following configuration. That is, a recording medium on which a pattern recognition program for identifying an arbitrary pattern included in image data is recorded, a program code of a learning data input step for inputting image data for learning, and density value normalization of learning data, Create a learning data pre-processing program code that removes noise, a learning data feature extraction program code that extracts pattern features from learning data, and a subspace that approximates the distribution of learning data in each class, Program code for discriminant function creation process for calculating discriminant function parameters using similarity between the partial spaces, program code for discriminant function and discriminant function storage step for storing the parameters, and digital image data to be discriminated Program data input process program code and input image data Program code for pre-processing step for density value normalization and noise removal, program code for feature extraction step for extracting feature amount of pattern from input image data, discrimination function read from program code for discrimination function storage step and its The program code of the identification process which identifies a pattern based on a parameter, and the program code of the output process which outputs an identification result are provided.

  In addition, on the recording medium on which the discriminant function storage program is recorded, based on the program code of the subspace creation process for calculating the subspace basis vector approximating the class distribution and the similarity between the subspaces by the mutual subspace method. And a program code of a discriminant function parameter calculation step for calculating a discriminant function parameter constructed by using a learning process of SVM.

  Alternatively, in the program code of the discriminant function storage process, the program code of the subspace creation process that calculates the basis vector of the nuclear nonlinear subspace that approximates the class distribution and the degree of similarity between the subspaces by the nonlinear nuclear mutual subspace method A discriminant function parameter calculation process program code for calculating a discriminant function parameter based on the SVM learning process.

  As described above, by realizing the means and steps according to the present invention, the learning function of each class is approximated to one partial space, and the discriminant function based on the SVM framework with the mutual similarity as the kernel function is constructed. Therefore, the number of computations is the same as when the number of learning data per class is set to 1 when calculating the discrimination function, and as a result, a multi-class discrimination problem exceeding 1000 classes can be handled in a realistic processing time. become.

  Further, the means or step according to the present invention is based on the conventional subspace method or the subspace of the other class in the feature space, because the support vector is used as a support vector and the positional relationship with the subspace of the own class is reflected in the discriminant function. Higher accuracy than kernel nonlinear subspace method can be expected.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration of a pattern recognition apparatus according to an embodiment of the present invention, and FIGS. 2A and 2B are flowcharts of a pattern recognition method according to the embodiment of the present invention.

First, a description will be given with reference to FIG. 1 and FIG. FIG. 2A is a learning process flowchart for setting the coefficient γ _i and the threshold coefficient h in Expression (12).

  In FIG. 1 and FIG. 2A, a learning data input unit 1 inputs learning data prepared in advance using an image acquisition device such as a camera or an image scanner (S1). The data is transmitted to the processing unit 2.

  The learning data preprocessing unit 2 performs normalization of the density value of the image data (for example, normalizing the density value to 0 and normalizing the variance to 1) and noise removal from the input digital image data (S2). And transmitted to the learning data feature extraction unit 3. The learning data feature extraction unit 3 extracts feature amounts from the learning data shaped by the learning data preprocessing unit 2 (S3).

  Various features can be considered, for example, power spectrum of Fourier transform, color histogram feature (Hirase Murase, VVVinod “Fast object search using local color information-Active search method-”), IEICE Transactions D -II, Vol.J81-D-II, No.9, pp.2035-2042,1998), Local Direction Histogram Features (Tetsufumi Wakabayashi, Shinji Tsuruoka, Fumitaka Kimura, Koji Miyake “Handwritten figures by increasing the dimension number of features” "High accuracy recognition", IEICE Transactions D-II, Vol. J77-D-II, No. 10, pp. 2046-2053, 1994. Further, the luminance value itself of the image may be used as the feature amount without performing such feature extraction processing.

  The discriminant function creation unit 4 creates an discriminant function based on the feature amount extracted from the learning data (S4). The block diagram of FIG. 3 and the flowchart of FIG. 4 show the configuration of the arithmetic processing unit and the processing execution method in the discriminant function creation unit 4 at that time. In FIG. 3, the subspace creating unit 11 calculates a base vector of the class subspace from the feature quantity of each class (S10 in FIG. 4) and (S11 in FIG. 4).

  This is calculated as an eigenvector of a covariance matrix of feature amounts extracted from learning data. In the case of non-linearity, it can be calculated as an eigenvector of the matrix W expressed by the above-described equations (8) and (9).

  Here, technical ideas made in applying the concept of subspace or the mutual subspace method to SVM are as follows (a) and (b).

  (A) The kernel function of SVM must satisfy Mercer's condition, but the subspace similarity calculation method in the mutual subspace method or nuclear nonlinear mutual subspace method satisfies Mercer's condition. They can be used directly as kernel functions.

  (B) When calculating the subspace, the result differs depending on whether the eigenvector around the center of gravity or the eigenvector around the origin is taken. However, when applying to SVM, the eigenvector around the origin was adopted. As a result, even if one of the two subspaces that are arguments of the kernel function in the discrimination process is replaced with a vector to be discriminated, the generality is not lost (since the vector can be regarded as a one-dimensional subspace). It can be configured.

  Next, the subspace method will be described. The subspace method is a kind of statistical method for determining categories to be classified through projection onto a subspace formed from the distribution of feature vector components. In this case, for the eigenvector calculation of the vector component to be converted, for example, KL analysis by Karhunen-Loeve transform which is a quantization algorithm is employed. As typical techniques in the subspace method, the CLAFIC method and the average learning subspace method are known.

  ALSM belongs to an adaptive learning approach that also takes into account opposing categories, and the learning of category boundaries is advanced by re-adjusting the space repeatedly so as to minimize the misrecognition of a predetermined training pattern.

  First, sample data of a pattern to be identified is input in advance, a partial space representing the pattern is calculated, and stored in a learning pattern storage unit. The subspace can be obtained by calculating the eigenvalue and eigenvector of the covariance matrix of the feature vector obtained from the identification pattern.

  At the time of identification, the partial space data of all patterns is called from the learning pattern storage unit, and the pattern is identified based on the distance from these partial spaces. The distance is measured by a simple Euclidean distance in the feature space, and the pattern of the subspace with the closest distance is used as the identification result.

  The basic principle of the subspace method is described in detail in Kenichiro Ishii, Nobuyoshi Ueda, Eisaku Maeda, Hiroshi Murase, “Easy-to-understand pattern recognition”, Ohmsha (1998), etc. In addition to the subspace method, for example, a nearest neighbor method, Fisher's linear discriminant method, support vector machine, neural network, kernel nonlinear subspace method, or the like may be used.

  The discriminant function parameter calculation unit 12 in FIG. 3 calculates the parameters γ and h in Expression (12) based on the class subspace data (S12 in FIG. 4) and (S13 in FIG. 4). The parameters γ and h are calculated by calculating the equations (17), (18), and (19) by the convex quadratic programming method. The discriminant function storage unit 5 stores the discriminant function parameters in the memory (S5).

  Next, description will be made with reference to FIG. 1 and FIG. FIG. 2B is an identification process flowchart. 1 and 2B, the image data input unit 6 inputs image data for identification using an image acquisition device such as a camera or an image scanner (S6), and transmits the data to the preprocessing unit 7. . The preprocessing unit 7 performs the same processing as the learning data preprocessing unit 2 (S7), and transmits it to the feature extraction unit 8. The feature extraction unit 8 performs the same process as the feature extraction unit 3 (S8), obtains the vector y shown in the equation (12), and transmits it to the identification unit 9.

  The discriminating unit 9 reads out the parameters stored in the discriminant function storage unit 5, substitutes the discriminant target vector y obtained in the above equation (12) to obtain the discriminant function s, and equations (14) (15 ) The similarity of each class is calculated by (16). The result is compared, and the class with the highest similarity is set as the identification class (S9). The output unit 10 outputs the identification result calculated by the identification unit 9.

  Note that “preprocessing units 2 and 7” in the learning device L and the identification device D in FIG. 1 and “preprocessing S2 and S7” in the learning processing in FIG. 2A and the identification processing in FIG. 2B are omitted. It doesn't matter.

  In the present invention, a storage medium in which a program code of software for realizing the functions of the above-described embodiments is recorded is supplied to a system or apparatus, and the CPU (MPU) of the system or apparatus stores the program code stored in the storage medium. It can also be realized by executing reading. In that case, the program code itself read from the storage medium realizes the functions of the above-described embodiment, and a storage medium storing the program code, for example, a CD-ROM, DVD-ROM, CD-R, CD-RW, MO, HDD, etc. constitute the present invention.

Embodiments Japanese character recognition is taken up as one embodiment of the present invention. In Japanese character recognition, a pattern is image information input by a camera, an image scanner, or the like, or partial image information arbitrarily cut out from the image information, and a class is a character type concept such as “A”, “A”, or “A”. It corresponds to.

  The learning data is a plurality of types of character images prepared in advance. For example, a plurality of types of font characters such as Gothic fonts and Mincho fonts can be used.

  Also, the discriminant function is a function for discriminating whether or not it is a certain character type concept (for example, a function for discriminating whether it is “A” or “A”). Prepared. For example, the identification result may be output as a file as a text data code, displayed on a display, or output data processed by translation or the like.

  As described above, according to the embodiments of the present invention, it is possible to provide a pattern recognition apparatus and method capable of handling a multi-class identification problem exceeding 1000 classes in a realistic processing time and enabling high-precision identification. .

  Each of the function realizing means shown in FIG. 1 is configured by a computer program, or the processing steps of the flowcharts shown in FIGS. 2 and 4 are configured by a computer program, and the program is stored in the computer. Needless to say, the program for realizing the function by the computer or the program for executing the processing steps by the computer can be read by a recording medium such as a flexible disk or a CD. , DVD, MO, ROM, memory card, removable disk, etc., and can be stored or distributed.

  It is also possible to provide the above program through a network such as the Internet or electronic mail. The present invention can be implemented by installing the program from these recording media into a computer or by downloading the program from a network and installing the program into the computer.

It is a block diagram which shows embodiment of this invention. It is a flowchart for describing operation | movement of embodiment, (a) is a learning process, (b) is an identification process flowchart. It is a block diagram which shows the structure of the arithmetic processing unit in an identification function preparation part. It is a flowchart which shows the execution method of the arithmetic processing in an identification function preparation part.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Learning data input part 2 ... Learning data pre-processing part 3 ... Learning data feature extraction part 4 ... Discrimination function creation part 5 ... Discrimination function storage part 6 ... Image data input part 7 ... Pre-processing part 8 ... Feature extraction part 9 ... Discrimination unit 10 ... Output unit 11 ... Subspace creation unit 12 ... Discrimination function parameter calculation unit L ... Learning device D ... Discrimination device

Claims (10)

A pattern recognition device for identifying an arbitrary pattern included in image data,
Learning data input means for inputting image data for learning;
Learning data feature extraction means for extracting feature amounts of patterns from learning data;
A discriminant function creating means for creating a subspace that approximates the distribution of the learning data in each class and calculating parameters of the discriminant function using the similarity between the subspaces;
A pattern recognition learning apparatus comprising: In the discrimination function creating means,
Subspace creation means for calculating a base vector of a subspace approximating the class distribution, and parameters of a discriminant function configured based on the degree of similarity between the subspaces by the mutual subspace method are supported by a support vector machine (SVM). A discriminant function parameter calculating means for calculating using a learning process;
The pattern recognition learning device according to claim 1, comprising: In the discrimination function creating means,
Support vector machine with subspace creation means for calculating basis vectors of nuclear nonlinear subspaces approximating class distribution and discriminant function parameters based on similarity between subspaces by nonlinear nuclear mutual subspace method A discriminant function parameter calculating means for calculating using a learning process of (SVM);
The pattern recognition learning device according to claim 1, comprising: A pattern recognition device for identifying an arbitrary pattern included in image data,
Image data input means for inputting digital image data to be identified;
Feature extraction means for extracting feature amounts of patterns from input image data;
An identification means for identifying a pattern based on an identification function read from the identification function storage means and its parameters;
An output means for outputting the identification result;
A pattern recognition and identification device comprising: A pattern recognition method for identifying an arbitrary pattern included in image data,
A learning data input step for inputting image data for learning;
A learning data feature extraction step for extracting feature amounts of patterns from the learning data;
Creating a subspace that approximates the distribution of the learning data in each class, and creating a discriminant function parameter that calculates the parameters of the discriminant function using the similarity between the subspaces;
A pattern recognition learning processing method characterized by comprising: In the identification function creating step,
A subspace creation step for calculating a base vector of a subspace that approximates the class distribution, and a parameter of a discriminant function that is configured based on the similarity between the subspaces by the mutual subspace method is used in the SVM learning process. Discriminating function parameter calculation step calculated by
The pattern recognition learning processing method according to claim 5, further comprising: In the identification function creating step,
A subspace creation step for calculating a basis vector of a nuclear nonlinear subspace that approximates the class distribution, and SVM learning of discriminant function parameters configured based on the degree of similarity between subspaces by the nonlinear nuclear mutual subspace method A discriminant function parameter calculation step to calculate using a process;
The pattern recognition learning processing method according to claim 5, further comprising: A pattern recognition method for identifying an arbitrary pattern included in image data,
An image data input step for inputting digital image data to be identified;
A feature extraction step of extracting a feature amount of the pattern from the input image data;
An identification step for identifying a pattern based on an identification function and its parameters read from the identification function storage unit;
An output step for outputting the identification result;
A pattern recognition / identification processing method comprising: A pattern recognition program, characterized in that the pattern recognition learning device and the identification device or the pattern recognition learning processing method and the identification processing method according to any one of claims 1 to 8 are programs for causing a computer to execute. A pattern recognition learning device and an identification device according to claim 1 or a pattern recognition learning processing method and an identification processing method as a program for causing a computer to execute the program, and the computer can read the program A recording medium on which is recorded a pattern recognition program characterized by being recorded on the medium.

Images