JP3475886B2 - Pattern recognition apparatus and method, and recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition apparatus for performing character recognition, voice recognition, etc., and more particularly to a pattern recognition apparatus and method and a recording medium having a large classification section.

[0002]

2. Description of the Related Art In character recognition and voice recognition, pattern recognition using a statistical pattern recognition method generally comprises a learning phase and a recognition phase.

In the learning phase, tens to thousands of feature quantities are extracted from each learning pattern, and then the distribution of recognition target categories and the position of category boundaries in this feature space are learned. The learning result is saved as an identification dictionary.

In the recognition phase, the feature quantity extracted from the input pattern in the same procedure is collated with the discrimination dictionary by the discrimination method, and the discrimination result is output.

As an example of the identification method, conventionally, ・ Nearest neighbor identification method, ・ Simple similarity method, ・ Mixed similarity method, ・ Subspace method, ・ Pseudo Bayes identification method, Various methods are known.

Since such an identification method is relatively easy to construct and can realize excellent identification performance, it is frequently used in a recognition device. However, since the feature quantity to be handled is high-dimensional, there is a problem that the calculation cost is high (a large amount of calculation is required) even though there is a difference in degree.

Therefore, as a method for reducing the amount of calculation, a method of recognizing by reducing the dimension number of the feature by the feature selection, a method of performing a large classification before the detailed identification, and the like have been proposed.

Typical examples of a method for reducing the number of dimensions to identify by feature selection are: A method for reducing the number of feature dimensions by principal component analysis or canonical discriminant analysis (IEICE Transactions Vol.J78 -D-
II No.11 pp.1627-1638, hereinafter referred to as "reference a") ・ A method for dimension reduction and identification by the FKL method that uses both the variance ratio and the variance as criteria for feature selection (IEICE Transactions Vol. .J80-D-II No.1 pp.73-80, hereinafter referred to as "reference b"), and a method for simultaneously learning a dictionary for feature selection and a recognition dictionary (Japanese Patent Laid-Open No. 9-245125). (Hereinafter referred to as “reference c”) and the like are known.

In the methods described in the above references a and b, first, principal component analysis, canonical discriminant analysis, or "FKL method" in which these are combined with respect to the n-dimensional feature extracted from each learning pattern. Is performed to create a feature selection dictionary, and based on this, the n-dimensional original features are converted into m (m ≦ n) -dimensional features.

Then, the category distribution or category boundaries in the transformed m-dimensional feature space is learned to create an identification dictionary.

At the time of recognition, similar feature extraction and feature selection are performed on the input pattern, and the extracted m-dimensional feature is collated with the identification dictionary for recognition.

The method described in the above-mentioned document c creates not only an identification dictionary composed of a set of m-dimensional features but also a feature selection dictionary for selecting m-dimensional features from n-dimensional features by learning at the same time.

Although the m-dimensional feature space transformed by using the eigenvectors obtained by the principal component analysis or the canonical discriminant analysis is not always the optimal space for the identification, the method of the above-mentioned document c makes the identification more effective. There is a possibility that a feature selection dictionary and a detailed identification dictionary suitable for can be constructed.

However, according to the above-mentioned document c, the original feature amount 256
It is described that the recognition performance is roughly equivalent to that when recognition is performed by the subspace method by selecting a feature of m = 16 dimensions from the dimensions, and when recognition is performed in a subspace higher than 16 dimensions. It is considered that the recognition performance is lower than that of the above.
That is, it can be said that this method emphasizes the recognition speed rather than the recognition performance.

On the other hand, as an example of a method for performing large classification in a low-dimensional space, there are many cases where original features are averaged and aggregated for each number of dimensions, and the obtained low-dimensional features are used for large classification.

The dimensional compression method by the above feature selection is n
M × n to select m-dimensional features from dimensional original features
In this method, the dimension can be compressed by addition operation of n times and division of m times, whereas
There is an advantage that the m-dimensional feature can be extracted at high speed.

However, since this feature selection method is likely to cause deterioration in recognition performance, in order to further improve recognition performance, feature selection is performed by principal component analysis or canonical discriminant analysis, and classification is performed roughly (hereinafter referred to as "method d"). Is good).

The method d is generally slower than the methods represented by the documents a and b, but the recognition performance is equal to or higher than the recognition performance.

As a conventional technique for generating a recognition dictionary used for pattern recognition by the subspace method by learning, for example, Japanese Patent Laid-Open No. 10-301917 discloses a subspace group belonging to the same category as the input vector. From among the subspaces belonging to a category different from the input vector, the second subspace having the smallest distance to the input vector is obtained. The subspace is obtained, and the first dictionary correction means makes the first subspace closer to the input vector,
The position vector and base vector of the origin of the first subspace are corrected, and the second dictionary correction means sets the position vector and base vector of the origin of the second subspace so that the second subspace moves away from the input vector. Has been proposed, a recognition dictionary learning method and apparatus for reducing the erroneous recognition rate for a learning pattern have been proposed. However, the above-mentioned JP-A-10-
The method and apparatus described in Japanese Laid-Open Patent Publication No. 301917 does not aim to provide a means for reducing the dimensionality of a feature and recognizing it by feature selection.

[0020]

From the above, it is considered that the above method d is most effective for improving the recognition speed while maintaining the recognition performance that can be realized by using the original feature.

However, even when the above method d is used, the recognition performance tends to decrease as the number of selected dimensions decreases. One of the reasons is that the feature axis selected by the principal component analysis or the canonical discriminant analysis is not necessarily optimal for large classification.

First, the problem in using the principal component analysis will be described with reference to FIG. FIG. 10 is a diagram showing distribution of data belonging to three types of categories existing in the two-dimensional original feature space.

Data distribution range 1001, 1002, 10
Reference numeral 03 denotes a distribution range of data belonging to categories A, B, and C, respectively. The characteristic axis 1004 indicates the characteristic axis (first main axis) selected by the principal component analysis for this data.

Converting a two-dimensional feature into a one-dimensional feature by principal component analysis is equivalent to projecting each pattern in the two-dimensional feature space onto the feature axis 1004 in FIG.

Therefore, the one-dimensional reference patterns obtained by projecting the two-dimensional reference patterns of the categories A, B, and C in the two-dimensional feature space on the feature axis 1004 are referred to as one-dimensional reference patterns 1006, 1008, and 1009, respectively.

The pattern 100 belonging to category A
1 is a one-dimensional pattern obtained by projecting 5 on the feature axis 1004.
The dimension pattern 1007 is used.

Then, as is clear from FIG. 10, considering the case of performing large classification and extracting two candidate categories in this one-dimensional feature space, the one-dimensional reference pattern 1006 belonging to the correct category A is Since it is at the farthest position from the one-dimensional pattern 1007 among the three one-dimensional reference patterns, it is not selected as a candidate for large classification.

That is, it is understood that, although the three categories are well separated in the Y-axis direction, large classification cannot always be performed well on the feature axis obtained by the principal component analysis.

The problems in using the discriminant analysis are as follows: 1) It is assumed that the data distribution of each category is a normal distribution. However, the distribution pattern of voice patterns and character patterns in the feature space is generally Not a normal distribution, 2) The maximum number of features that are effective for identification (number of categories to be identified −
1) Only points can be obtained.

For example, when recognizing a handwritten numeral 10 category, only nine features can be selected at the maximum, so that it is not used as a feature selection method at present.

Therefore, the present invention has been made in view of the above problems, and its purpose is to compare the number of feature dimensions used for large classification with a recognition apparatus that performs large classification based on principal component analysis or discriminant analysis. Another object of the present invention is to provide a recognition device pattern recognition device and method, and a recording medium that can realize higher recognition accuracy when the number of candidates is the same. Other objects, features, advantages and the like of the present invention will be immediately apparent to those skilled in the art from the following description.

[0032]

In order to achieve the above object, the present invention provides a feature extraction unit for extracting an n-dimensional feature as a primary feature from an input pattern and an m-dimensional (m <m
n) a feature selection dictionary that stores a function for converting into features,
A feature selection unit that extracts an m-dimensional feature from the primary feature as a secondary feature based on a function stored in the feature selection dictionary, and a detailed identification created in advance for recognition based on the primary feature Creating a dictionary and a large classification dictionary that converts each detailed identification reference pattern included in the detailed identification dictionary into an m-dimensional large classification reference pattern based on a function stored in the feature selection dictionary and stores the reference pattern Part and the large-classification reference patterns, the upper K large-classification reference patterns that are close in distance to the secondary features extracted by the feature selection part are extracted, and the detailed identification reference patterns corresponding thereto are extracted. And a K-classification reference pattern specified by the output information of the large classification unit and the primary feature extracted by the feature extraction unit are collated to collate each other. The result Included in the general classification reference pattern at the time of learning the function of the detailed identification unit, the recognition result output unit that processes the collation result and outputs it to a memory or a display, and the function stored in the feature selection dictionary. Among the reference patterns of the correct answer category, the reference pattern for large classification, which is the closest to the secondary feature extracted from the learning pattern, is extracted as the reference pattern A, and the incorrect category included in the reference pattern for large classification is further extracted. Of the reference patterns, the large classification reference pattern closest to the secondary feature is extracted as the reference pattern B, and
A feature selection dictionary modifying unit that modifies the feature selection function so that the distance between the next feature and the reference pattern A becomes closer, and the distance between the secondary feature and the reference pattern B becomes farther. To do.

Further, according to the present invention, a feature extracting section for extracting n-dimensional features as primary features from an input pattern, and a feature selection storing a function for converting the primary features into m-dimensional (m <n) features. A feature selection unit that extracts an m-dimensional feature from the primary feature as a secondary feature based on a dictionary and a function stored in the feature selection dictionary, and is created in advance for recognition based on the primary feature. The detailed identification dictionary and each of the detailed identification reference patterns included in the detailed identification dictionary are converted into an m-dimensional large classification reference pattern based on the function stored in the feature selection dictionary, and stored. The top K large classification reference patterns having a high degree of similarity between the classification dictionary creation unit and the large classification reference patterns with the secondary features extracted by the feature selection unit are extracted, and the details corresponding to these are extracted. Identification reference A major classification section for outputting information specifying the over emissions, K specified by the output information of the major classification section
The detailed identification reference pattern for each of the detailed identification and the primary feature extracted by the feature extraction unit are collated to output a collation result, and the output of the detailed identification unit is processed into a memory or a display. At the time of learning the recognition result output unit to output and the feature selection function stored in the feature selection dictionary, the secondary feature extracted from the learning pattern among the reference patterns of the correct category included in the reference pattern for large classification. The reference pattern for large classification having the highest similarity with is extracted as a reference pattern A, and further, among the reference patterns of the incorrect category included in the reference pattern for large classification,
After extracting the reference pattern B for large classification having the highest degree of similarity with the secondary feature as the reference pattern B, the degree of similarity between the secondary feature and the reference pattern A increases, and further, the secondary feature and the reference A feature selection dictionary modifying unit that modifies the feature selection function so that the degree of similarity with the pattern B is further reduced is provided.

The method according to the present invention includes a feature selection dictionary for storing feature selection information for converting a primary feature of an input pattern into an m-dimensional (where m <n) feature, and a primary feature of the input pattern. The data processing device includes a feature extraction unit and a feature selection unit, and a storage unit that stores a detail identification reference dictionary that stores, for each category, a detail identification reference pattern created in advance for recognition based on A pattern recognition method of a pattern recognition device comprising: a large classification dictionary creation unit, a large classification unit, a detailed identification unit, a recognition result output unit, and a feature selection dictionary correction unit, wherein the learning phase is (A) In the feature extraction unit, n is used as a primary feature from the input pattern.
Extracting a dimensional feature, and (b) extracting an m-dimensional feature from the primary feature as a secondary feature on the basis of a function stored in the feature selection dictionary in the feature selecting unit, c) In the large classification dictionary creating section, each of the detailed identification reference patterns included in the detailed identification dictionary is converted into an m-dimensional large classification reference pattern based on a function stored in the feature selection dictionary. And (d) in the large classification unit, a predetermined number (K) of upper ranks having a close distance to the secondary feature extracted by the feature selection unit from the large classification reference pattern. A step of extracting reference patterns for large classification and outputting information specifying the reference patterns for detailed identification corresponding to these; (e) in the detailed identification section, according to output information of the large classification processing. Collating a predetermined number (K) of the reference patterns for detailed identification with the primary features extracted by the feature extraction processing, and outputting a collation result; (f) the collation A recognition result output process of processing a result and outputting the result to a storage device or an output device, and (g) in the feature selection dictionary correction unit, at the time of learning a function stored in the feature selection dictionary, in the reference pattern for large classification Of the reference patterns included in the above, which belong to the same category as the learning pattern and are extracted from the learning pattern.
The first is the reference pattern for large classification that is closest to the next feature.
Of the reference patterns included in the reference pattern for large classification, which belong to a category that is not the same as the learning pattern and are closest to the secondary feature, So that the distance between the secondary feature and the first reference pattern becomes closer and the distance between the secondary feature and the second reference pattern becomes longer.
Modifying the information in the feature selection dictionary.

[0035]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described. First, the principle and operation of the present invention will be described with reference to FIG. FIG. 9 is a diagram showing how data of three categories (categories A, B, and C) are distributed in a two-dimensional original feature space.

In the present invention, the reference patterns of the categories A, B, and C in the two-dimensional space are feature-selected by the current feature selection dictionary (106 in FIG. 1) and converted into a one-dimensional reference pattern. One-dimensional reference pattern 904,
Let 905 and 906 be one-dimensional features obtained by converting one learning pattern 901 belonging to category A by the same feature selection dictionary (106) as one-dimensional feature 910.

Then, at this point, the one-dimensional feature 910
Has the shortest distance from the one-dimensional reference pattern 905 of the category B, and the farthest from the one-dimensional reference pattern 904 of the category A which is the correct category.

Therefore, even if the feature selection dictionary (106) is used to perform a large classification for extracting two candidates, the category A is not selected as a candidate.

Therefore, the feature selection dictionary correction unit (10 in FIG. 1).
8) is the one-dimensional reference patterns 904 and 1 of the correct answer category.
The feature selection dictionary (106) is modified so that the distance to the one-dimensional feature 910 becomes closer, and further the distance between the one-dimensional feature and the reference pattern 905 in the incorrect category that is closest to the one-dimensional reference pattern 904 becomes farther.

In the example shown in FIG. 9, the correction of the feature selection dictionary (106) causes the feature axis to rotate to the right.

Based on the feature selection dictionary (106) thus obtained, the one-dimensional reference pattern and the one-dimensional feature are recalculated.

If the new feature axis is the feature axis 903, on the main axis, the one-dimensional feature 9 on which the learning pattern is projected is projected.
08 is closest to the one-dimensional reference pattern 907 of the correct category, and is correctly selected as a candidate for large classification.

With such an operation, the recognition rate can be improved as compared with a device having no feature selection dictionary correction section (108).

In a preferred embodiment of the method according to the present invention, referring to FIG. 1, feature selection information for converting primary features of an input pattern into m-dimensional features (where m <n) is provided. A feature selection dictionary (106) to be stored and a detail identification dictionary (104) that stores a reference pattern for detail identification, which is created in advance for recognition based on the primary feature of the input pattern, are stored for each category. From the n-dimensional reference pattern in the identification dictionary (104) and the n-dimensional features extracted from the learning pattern, the feature selection dictionary (10
Based on the content of 6), an m (m <n) -dimensional reference pattern and an m-dimensional feature are extracted, and among the m-dimensional reference patterns that belong to the same category as the learning pattern, a reference that is closest to the m-dimensional feature. The pattern is extracted as the first reference pattern and belongs to a category different from the learning pattern m
Of the dimensional reference patterns, the reference pattern closest to the m-dimensional feature is extracted as a second reference pattern, and the distance between the m-dimensional feature and the first reference pattern becomes shorter, and the m Dimensional feature and the second
The content of the feature selection dictionary (106) is updated so that the distance from the reference pattern of is further.

In a preferred embodiment of the pattern recognition apparatus of the present invention, a feature extraction unit (102) for extracting n-dimensional features as primary features from a pattern input from a data input unit (101), Based on a feature selection dictionary (106) storing a function (feature selection information) for converting a primary feature into an m-dimensional (m <n) feature, and information stored in the feature selection dictionary (106), 2 A feature selection unit (103) for extracting an m-dimensional feature from the primary feature as the next feature, and a detailed identification dictionary (104 for storing a detailed identification reference pattern created in advance for recognition based on the primary feature). ) And each of the detailed identification reference patterns included in the detailed identification dictionary (104) are assigned to the feature selection dictionary (1
Based on the function (feature selection information) stored in 06), a large classification dictionary creation unit (105) that converts and stores the m-dimensional large classification reference pattern and a feature selection from the large classification reference pattern. A predetermined large number (K) of large-classification reference patterns that are close in distance to the secondary features extracted by the unit (103) are extracted, and information that specifies the detailed identification reference patterns corresponding to these is output. The classification unit (107) and a predetermined number (K) of the detailed identification reference patterns specified by the output information of the large classification unit (107) are collated with the primary features extracted by the feature extraction unit. A detailed identification section (109) for outputting the collation result, and a recognition result output section (11) for processing the collation result and outputting it to a storage device or an output device.
0) and the function (feature selection information) stored in the feature selection dictionary (106), the same category (hereinafter, “correct answer”) as the learning pattern among the reference patterns included in the large classification reference pattern. Category)),
In addition, the reference pattern for large classification, which is the closest to the secondary feature extracted from the learning pattern, is the reference pattern (A).
Of the reference patterns included in the reference pattern for large classification, which belong to a category that is not the same as the learning pattern (hereinafter referred to as “incorrect answer category”),
Further, after the reference pattern (B) for large classification having the closest distance to the secondary feature is extracted as the reference pattern (B), the distance between the secondary feature and the reference pattern (A) becomes closer,
Furthermore, a feature selection dictionary correction unit (108) that corrects the previous feature selection dictionary (106) is provided so that the distance between the secondary feature and the reference pattern (B) becomes longer.

In one embodiment of the pattern recognition apparatus of the present invention, the above-mentioned feature extraction section (102), feature selection section (103), and large classification dictionary creation section (105).
A large classification section (107), a detailed identification section (109),
A recognition result output unit (110) and a feature selection dictionary correction unit (1
The processing in each unit of 08) is realized by a program executed on a computer. In this case, various media such as a recording medium (semiconductor memory, FD (floppy disk), CD-ROM, MT (magnetic tape), DVD (digital versatile disk)) on which the program is recorded or a communication medium are used. The present invention can be implemented by reading the program into the computer through the reading device and the interface of the medium, loading the executable program into the main memory of the computer, and executing the program.

More specifically, a recording medium having a program recorded thereon according to the present invention is, in one embodiment thereof, a primary feature (n) of an input pattern in a pattern recognition device.
Feature selection dictionary that stores a function for converting a dimension to an m-dimensional feature (where m is a positive integer smaller than n), and a reference pattern for detailed identification created in advance for recognition based on the primary feature And a storage device for storing the detailed identification dictionary for storing
A feature extraction process for extracting a dimensional feature, and (b) a feature selection process for extracting an m-dimensional feature from the primary feature as a secondary feature based on a function stored in the feature selection dictionary; A large classification dictionary creating process of converting each detailed identification reference pattern included in the detailed identification dictionary into an m-dimensional large classification reference pattern based on a function stored in the feature selection dictionary and storing the reference pattern. (D) 2 extracted by the feature selection unit from the large-classification reference pattern
A large classification process of extracting a predetermined number (K) of large classification reference patterns close to the next feature, and outputting information for specifying the detailed identification reference pattern corresponding to these; A predetermined number (K) of the detailed identification reference patterns specified by the output information of the large classification process;
A detailed identification process for collating the primary feature extracted by the feature extraction process and outputting a collation result, and (f) a recognition result output process for processing the collation result and outputting it to a storage device or an output device. (G) At the time of learning the function stored in the feature selection dictionary, it belongs to the same category as the learning pattern (hereinafter, referred to as “correct answer category”) among the reference patterns included in the large-classification reference pattern. Further, a large classification reference pattern that is closest to the secondary feature extracted from the learning pattern is extracted as a reference pattern (A), and further learning is performed from the reference patterns included in the large classification reference pattern. A large-classification reference pattern that belongs to a category that is not the same as the pattern (hereinafter referred to as “incorrect solution category”) and that has the closest distance to the secondary feature is referred to as a reference pattern (B). By extracting the function of the feature selection dictionary so that the distance between the secondary feature and the reference pattern (A) becomes closer and the distance between the secondary feature and the reference pattern (B) becomes further away. A program for causing the computer constituting the pattern recognition apparatus to execute the feature selection dictionary correction process to be corrected and each of the processes (a) to (g) is stored.

[0048]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In order to describe the embodiment of the present invention described above in more detail, an embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a first pattern recognition apparatus according to the present invention.
It is a figure which shows the structure of the Example of this. Referring to FIG. 1, the pattern recognition device includes a data input unit 101, a feature extraction unit 102, a feature selection unit 103, and a large classification unit 104.
A feature selection dictionary correction unit 105 and a feature selection dictionary 106
A detailed identification dictionary 107 and a large classification dictionary creation unit 108
A detailed identification section 109 and a recognition result output section 110.

The data input unit 101 takes in the data to be recognized and outputs it to the feature extraction unit 102. Examples of the data to be recognized include image data and audio data.

The feature extraction unit 102 is a data input unit 101.
The n-dimensional feature effective for recognition is extracted from the data input from the, and output to the feature selecting unit 103 and the detail identifying unit 109.

The feature selection unit 103 has a feature selection dictionary 106.
With reference to, the n-dimensional feature output by the feature extraction unit 102 is converted into an m-dimensional feature, and this is output to the large classification unit 104 and the feature selection dictionary correction unit 105.

The detailed identification dictionary 104 is a dictionary for detailed identification that has been constructed in advance by some learning method, and has n added with a category code and a reference pattern number.
It is a set of dimension reference patterns.

The large classification dictionary creating unit 105 is a processing unit that converts each n-dimensional reference pattern stored in the detailed identification dictionary 104 into an m-dimensional reference pattern for large classification based on the feature selection dictionary 106. The same reference pattern number is added to both the n-dimensional reference patterns and the m-dimensional reference patterns obtained by converting the n-dimensional reference patterns.

The feature selection dictionary 106 is a dictionary describing a function for feature selection. In the case of feature selection by linear conversion, each element value of the conversion matrix (m × n matrix) may be stored.

The large classification unit 107 is a large classification dictionary creating unit 10.
5 is a processing unit that extracts K patterns having a high degree of coincidence with the m-dimensional feature pattern output from the feature selecting unit 103 from the set of m-dimensional reference patterns output, and outputs the reference pattern number. The evaluation of the degree of coincidence may be based on the distance between the feature pattern and the reference pattern in the m-dimensional space, or may be based on the degree of similarity.

In this embodiment, the degree of coincidence is defined based on the square of the Euclidean distance.

The detailed identification section 109 extracts from the n-dimensional reference patterns corresponding to the reference pattern numbers output by the general classification section 107, those close to the n-dimensional feature pattern output by the feature extraction section 102, and the category code thereof. And the recognition evaluation value are output. As a specific identification algorithm,
There are various methods such as nearest neighbor identification, simple similarity identification, and pseudo Bayes identification, and any method is used. In the embodiment, an example using the nearest neighbor discriminator is shown.

The recognition result output unit 110 includes a detailed identification unit 10
The category code and recognition evaluation value output by 9 are processed and output to a memory, a display, or the like.

The feature selection dictionary correction unit 108 is a processing unit that executes only the learning phase, and a process of correcting the feature selection dictionary 106 so that the reference pattern of the correct answer category is more likely to remain as a candidate in the large classification unit 107. It is a department.

Next, the operation of the first embodiment of the present invention shown in FIG. 1 will be described.

Here, for easier understanding,
An example in which the present invention is applied to a handwritten numeral recognition device will be specifically described. Further, since this handwritten numeral recognition device is composed of a learning phase for learning the feature selection dictionary and a recognition phase for successively recognizing input data by using the completed feature selection dictionary, an operation is performed for each phase. explain.

First, an example of the operation of the learning phase will be described with reference to FIG. FIG. 2 is a flowchart showing the flow of a large process in the learning phase. This learning is
This is so-called sequential learning.

First, the number of learning cycles and the number of patterns to be learned are set (step S201).

Further, the feature selection dictionary 106 is initialized (step S202). The initial value is set by, for example, principal component analysis. That is, the covariance matrix is created from the n-dimensional features extracted from all the learning patterns, m of the eigenvalues are selected in descending order, and m sets of eigenvectors for them are used as initial values. The number of learning cycles is a variable that specifies how many times the learning work for all learning patterns is repeated when the learning work is counted as one cycle.

Next, a counter LC for storing the number of next cycles
Is initialized to 1 (step S203).

Thereafter, it is checked whether or not the learning cycle number exceeds the specified number (step S204), and if it exceeds the specified number, the learning is ended, and if it is less than the specified number, the following processing is performed.

First, a counter PN that stores the next pattern number
Is initialized to 1 (step S205), and it is checked whether the number of learning patterns exceeds a prescribed number (step S20).
6).

If it exceeds, the counter LC for counting the number of next cycles is incremented by 1, and the process returns to step S202. If the number is less than the specified number, one learning pattern is read, the correction amount is calculated, and the feature selection dictionary is corrected (step S20).
7).

When the correction is completed, the counter PN for storing the next pattern is incremented by 1, and the process returns to step S206.

As a result, the learning cycle correction procedure is executed for each learning pattern, and the learning phase ends.

Next, the correction of the feature selection dictionary in the learning phase will be described in detail. This process is a process that is the main feature of the present invention.

First, the learning pattern is input to the data input unit 101 and output to the feature extraction unit 102. The learning pattern is a character image cut out in advance, and may be a binary image or a multivalued image.

The feature extraction unit 102 extracts an n-dimensional feature effective for character recognition from the learning pattern output by the data input unit 101. As the features effective for character recognition, for example, a local direction index histogram, an expanded cell feature, an outer direction contribution degree feature, etc. are known, and the value of n when using each feature is about 64 to 1176 dimensions. It is considered good.

The feature selection unit 103 converts the n-dimensional features extracted by the feature extraction unit 102 into m-dimensional features by referring to the feature selection dictionary 106.

The n-dimensional feature values are x1, x2, ..., xn,
Assuming that m-dimensional feature values are y1, y2, ..., ym and m functions for feature selection are f1, f2, ..., fm, the calculation performed by the feature selection unit 103 is as follows. Can be described as

[0077] y1 = f1 (x1, x2, ...., xn) y2 = f2 (x1, x2, ...., xn) ..................... ym = fm (x1, x2, ...., xn)

If the feature selection method of the feature selection unit 103 is a linear transformation, the n-dimensional feature and the m-dimensional feature are vector X = (x1, x2, x3, ..., xn) ^T , vector Y =, respectively. (y1, y2, y3, ...., yn) ^T , the feature selection is an m * n matrix,

Expressed as, the operation performed by the feature selection unit 103 can be simply expressed as Y = ZX. However, here, the symbol T represents transposition.

The detailed identification dictionary 104 is a dictionary for the detailed identification section that has been learned in advance, and does not change in the present invention.

As a concrete example, it can be expressed by a set of n-dimensional reference patterns to which category code information and reference pattern number information as shown in FIG. 3 are added.

In FIG. 3, the category code 301 is
This is a numerical value that indicates the category of characters, and here is an example using an ASCII code. Besides this, JIS code and EUC
It is possible to use codes, SJIS codes, etc.,
When using JIS code, EUC code, SJIS code, etc., a 2-byte area is required.

The reference pattern number 302 is a number for associating the detailed identification reference pattern with the large classification reference pattern obtained by feature selection of the detailed identification reference pattern. Just number them. As for the size of the area for storing the reference pattern number, 1 byte is sufficient if the number of reference patterns is 255 or less, but if the number of reference patterns is large, a storage area larger than that is prepared.

The n-dimensional reference pattern 303 stores characteristic values in order, and if each element is represented by 1 byte, an area of n bytes is required.

Further, in FIG. 3, the category code 3 immediately after the final reference pattern is shown to show the final position of the reference pattern.
04 and the reference pattern number 305 are set to the value -1, but as another method, another storage area for holding the number of reference patterns may be provided in the detailed identification dictionary 104.

The large classification dictionary creation unit 105 refers to the feature selection dictionary 106 and calculates each n-dimensional reference pattern in the detailed identification dictionary 104 by the same calculation as the feature selection unit 103.
It is converted into a dimensional reference pattern and stored in association with it.

There is a one-to-one correspondence between the n-dimensional reference pattern in the detailed identification dictionary 104 and the m-dimensional reference pattern stored in the large classification dictionary creating section 105.

When the m-dimensional feature output by the feature selection unit 103 and the m-dimensional reference pattern stored in the large classification dictionary creation unit 105 are calculated, the feature selection dictionary correction unit 108 is next.
Is started.

The operation of the feature selection dictionary correction unit 108 will be described with reference to FIG. For simplification, FIG. 13 shows a case where the original feature is two-dimensional and the number of dimensions after feature selection is one-dimensional.

Three types of categories 1307 in the two-dimensional space
There are patterns belonging to ˜1309. In addition, two n-dimensional reference patterns 1301 to 1306 are provided for each category.
Is set, and each reference pattern corresponds to the current feature selection unit 103.
Is converted into m-dimensional reference patterns 1311 to 1316.

The learning pattern 1310 belonging to the category 1307 is the one-dimensional learning pattern 13 by the feature selection.
Converted to 17.

As a first step, two patterns having a distance close to the m-dimensional characteristic pattern are extracted from the m-dimensional reference patterns. However, one of the two reference patterns belongs to the same category as the input learning pattern, and if there are a plurality of m-dimensional reference patterns having the same category code as the learning pattern, the reference pattern is regarded as an m-dimensional feature. Select the one with the closest distance.

In FIG. 13, the one-dimensional reference pattern 1312
Corresponds to this. The other m-dimensional reference pattern has the closest distance to the m-dimensional feature among the m-dimensional reference patterns belonging to a category different from the input learning pattern. In FIG. 13, the one-dimensional reference pattern 131
3 corresponds to this.

The distance is, for example, the square of the Euclidean distance,
That is, d (Y, Pi) = ‖ (Y − Pi) ‖ ² = ‖ (ZX − ZRi) ‖ ²

It is defined by However, Pi is the reference pattern number
i is the m-dimensional reference pattern, Ri is the n of the reference pattern number i
A dimensional reference pattern is shown.

Of the two extracted reference patterns, the number of the m-dimensional reference pattern belonging to the same category as the learning pattern is c1, and the number of the m-dimensional reference pattern belonging to a different category is
If c2, the distance between the m-dimensional feature pattern and each reference pattern is d1 = d (Y, Rc1), d2 = d (Y, Rc2), respectively.

In the modification of the feature selection dictionary, the distance between the m-dimensional pattern extracted from the learning pattern and the reference pattern Rc1 becomes shorter, and the distance between the m-dimensional pattern extracted from the learning pattern and the reference pattern Rc2 becomes longer. To be executed. The maximum change direction of the distance d due to the change of the feature selection dictionary Z is given by the following expression (2).

[0098]

Therefore, in order to make the distance 1318 between the m-dimensional pattern extracted from the learning pattern and the reference pattern Rc1 closer,

[0100]

It suffices to subtract the value from the original feature selection dictionary Z.

On the other hand, in order to further increase the distance 1319 between the m-dimensional pattern extracted from the learning pattern and the reference pattern Rc2,

[0103]

It suffices to add such a value to the original feature selection dictionary Z. Where t is the learning cycle, ε1
(t) and ε2 (t) are functions for adjusting the correction amount. Therefore, both correction amounts are added to finally obtain the following correction formula (5).

[0105]

With this correction, if the feature selection dictionary is sequentially revised, the feature selection axis gradually changes to a direction suitable for identification as already described, and the reference pattern of the correct category becomes a high-ranking candidate in the large classification unit. It is easy to go up, and the recognition performance can be improved even if the number of dimensions m is relatively small.

The feature selection dictionary correction unit 108 corrects the feature selection dictionary 106 based on the correction amount.

In the above modified equation, ε1 (t), ε2
It is theoretically desirable that the value of (t) be proportional to 1 / t, but in actual application, it may be a constant.

Next, the operation of the recognition phase in the embodiment of the present invention will be described with reference to FIG. Figure 4
FIG. 6 is a diagram showing a flowchart of a recognition phase in one embodiment of the present invention.

In the recognition phase, initialization (step S4
01), input for each input data (step S40
2), feature extraction (step S403), feature selection (step S404), large classification (step S405), detailed classification (step S406), result output (step S40).
7) is performed in order. Each processing will be specifically described below.

First, initialization is performed prior to recognition (step S4).
01) is performed. In the initialization, the large classification dictionary creation unit 10
5 is activated, and the m-dimensional reference pattern is created and stored by referring to the feature selection dictionary 106 and the detailed identification dictionary 104 created in the learning phase.

In the learning phase, the large classification dictionary creating unit 105 is activated for each learning pattern, but in the recognition phase it is activated only once during initialization. Therefore, the reference pattern for recognition and large classification does not change.

At the input step (step S402),
The patterns to be recognized are read one by one.

The feature extracting step (step S403) in the feature extracting section 102 and the feature selecting step (step S404) in the feature selecting section 103 are executed. These processing contents are completely the same as the learning phase except that the input data is not the learning pattern but the identification target pattern.

Then, a large classification step (step 405)
Is executed by the general classification unit 107. Major classification 10
Reference numeral 7 extracts only the upper K pieces of the m-dimensional reference patterns that are close to the m-dimensional feature pattern from the m-dimensional reference patterns extracted by the large classification dictionary creating section 105, and outputs the reference pattern numbers to the detailed identifying section 109. At this time, the categories to which the extracted K reference patterns belong may be the same or different.

The detailed identification section 109 extracts the n-dimensional reference pattern corresponding to the reference pattern number output from the large classification section 107 from the detailed identification dictionary 104, and the feature extraction section 102 outputs n out of the n-dimensional reference patterns. Outputs a distance close to the dimensional feature pattern. The distance may be defined by Euclidean distance.

The recognition result output unit 110 includes a detailed identification unit 10
The recognition result output from 9 is processed and output to a memory or a screen.

The recognition experiment results of the handwritten alphanumeric kana recognition device actually constructed by the above configuration will be shown.

The number of categories to be recognized used in the experiment is 82.
Seeds, the number of input patterns (the number of evaluation patterns) is about 85,000 patterns, the features are the features (400 dimensions) in which the inclination of the contour line is aggregated for each local region / direction, and the detailed identification dictionary is a generalized learning vector quantum. 10 reference vectors were created for each category by the optimization method, and a multi-template dictionary was constructed.

The classification algorithms used in the large classification section and the detailed classification section are both the nearest neighbor classification algorithm based on the Euclidean distance.

FIG. 5 shows an example of the experimental results. FIG. 5 shows recognition rates and processing times according to four different recognition procedures.

The first stage shows the recognition performance when the 400-dimensional feature pattern is directly identified without performing the major classification, and the second stage shows the 400-dimensional feature pattern for every four dimensions. The recognition performance in the case where the 100-dimensional feature patterns are aggregated to perform the large classification and the 400-dimensional features are used for the detailed classification is shown. The recognition performance is shown when 15-dimensional feature patterns are selected for major classification and 400-dimensional features are used for detailed classification.

The fourth row shows the performance when the feature selection dictionary is learned and the large classification is performed in the 15-dimensional feature space and then the detailed identification is performed in the 400-dimensional feature space based on the present invention. .

It can be seen that in the recognition device, the recognition accuracy is almost the same as that in the recognition using the 400-dimensional feature pattern directly, and the recognition time is reduced to about 1/9.

On the other hand, it can be seen that the performance is lower than that of the method in the first step when dimension compression is performed using the principal component analysis shown in the third step.

As a result of the additional experiment, in the case of using the principal component analysis, the equivalent recognition performance could not be realized unless the features of at least 25 dimensions were selected.

It can also be seen from FIG. 5 that the present invention is superior to both the recognition rate and the recognition time in comparison with the method of roughly classifying 100-dimensional features shown in the second row.

FIG. 6 is a diagram showing the relationship between the number of learning cycles and the recognition performance in the learning phase of the feature selection dictionary in the recognition apparatus according to the embodiment of the present invention.

The initial value of the feature selection dictionary was created by principal component analysis.

It can be seen from FIG. 6 that the recognition performance is gradually improved by learning, and the recognition performance is achieved at about 12 learning cycles, which is almost the same as that in the case of no major classification.

In the above embodiments, the large classifying section, the large class dictionary creating section, and the detailed classifying section have shown an example in which the classification / identification processing is performed by the distance measure, but these may use the degree of similarity as a measure.

When the large classification unit and the large classification dictionary creating unit are based on the similarity, similarly, the value obtained by partially differentiating the similarity with the feature selection dictionary Z is multiplied by ε1 (t) and ε2 (t). After that, by adding / subtracting this to / from the original feature selection dictionary Z, the similarity can be further increased / decreased.

Next, the second embodiment of the present invention will be described with reference to FIG.
Will be described with reference to. In the second embodiment of the present invention, unlike the above embodiment, the feature selection dictionary correction unit 708 further refers to the output of the detailed identification unit 709 to calculate the correction amount.

The second embodiment of the present invention will be described with reference to FIG. n-dimensional reference patterns 1201, 1
Reference numeral 202 denotes a reference pattern belonging to category D, n
The dimensional reference patterns 1203 and 1204 indicate reference patterns belonging to the category E, and the n-dimensional reference pattern 120
Reference numerals 5 and 1206 represent reference patterns belonging to the category F.

Further, the m-dimensional reference patterns projected on the feature axis 1208 by the current feature selection dictionary are set as m-dimensional reference patterns 1211 to 1216, respectively.

Assuming that the n-dimensional feature extracted from the learning pattern is the n-dimensional feature 1207, in the first embodiment, the one-dimensional reference pattern 1212 and the one-dimensional reference pattern 1213 are extracted and the correction amount of the dictionary is calculated. I have decided.

However, the detailed identification dictionary 104 is
In the case of a multi-template dictionary including a plurality of reference patterns for one category, the reference pattern of the correct category closest to the n-dimensional feature extracted from the input pattern in the n-dimensional feature space is the input pattern in the m-dimensional feature space. It does not always correspond to the reference pattern of the correct answer category that is closest to the m-dimensional feature obtained from.

Actually, in the case of the example shown in FIG. 12, n which is the closest to the n-dimensional feature 1207 in the two-dimensional feature space.
The one-dimensional reference pattern 1201 does not correspond to the one-dimensional reference pattern 1212 that is closest to the one-dimensional feature 1217.

Therefore, in order to select the reference pattern closest to the input pattern at the time of the detailed identification as a candidate for the larger classification, the one-dimensional reference pattern 12
12 does not correct the feature selection dictionary 706 so that the m-dimensional feature 1217 and the 12-dimensional feature 1217 come close to each other.
The feature selection dictionary 706 may be modified so that 1 and the m-dimensional feature 1217 come close to each other.

Therefore, if the number of the n-dimensional reference pattern that belongs to the same category as the input pattern in the n-dimensional feature space and is closest to the n-dimensional feature vector obtained from the input pattern is c1 ', the feature selection dictionary Correction unit 707
The correction procedure in can be defined by the following equation (6).

[0141]

FIG. 11 is a flow chart showing the processing procedure of the second embodiment of the present invention. Referring to FIG.
In the second embodiment of the present invention, first, the reference pattern Rc1 ′ is extracted by the detailed identification section 7-9 for all learning patterns (step S1102). At this time,
The detailed identification section 709 operates independently of the output of the large classification section, and extracts and stores the closest one of all reference patterns in the detailed identification dictionary 704 belonging to the same category as the learning pattern.

As a result, the reference pattern number c1 'of the correct answer category can be extracted for all the learning patterns. Therefore, in the subsequent correction step (step S1107), if the reference pattern number c2 of a different category is determined, the correction described above is performed. It becomes possible to do.

FIG. 8 is a diagram showing the recognition performance and the recognition time by the recognition device constructed according to the second embodiment of the present invention. The experimental conditions are the same as those shown in FIG.

It can be seen from FIG. 8 that the second embodiment of the present invention achieves the recognition performance equivalent to that of the first embodiment.

[0146]

As described above, according to the present invention, it is possible to realize higher recognition performance as compared with the recognition method in which large classification is performed using a feature selection dictionary generated by principal component analysis or the like. Play.

The reason for this is that the present invention is provided with means for sequentially learning the feature selection dictionary so that the reference vector corresponding to the correct answer category can be easily selected as a candidate for large classification.

In addition, according to the present invention, from a different point of view, the dimension number of the features used for the large classification is larger than that of the recognition method for performing the large classification using the feature selection dictionary generated by the principal component analysis or the like. Even if the number is smaller, the same recognition performance can be realized, and the recognition speed can be improved.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is a flowchart showing a processing procedure of a learning phase in the first embodiment of the present invention.

FIG. 3 is a diagram showing a specific example of a detailed identification dictionary in the first exemplary embodiment of the present invention.

FIG. 4 is a flow chart showing a processing procedure of a recognition phase in the first embodiment of the present invention.

FIG. 5 is a diagram showing a recognition experiment result of handwritten alphanumeric kana by the recognition device to which the first embodiment of the present invention is applied.

FIG. 6 is a diagram showing a relationship between the number of times of learning and recognition performance in the recognition device to which the first embodiment of the present invention is applied.

FIG. 7 is a diagram showing a configuration of a second exemplary embodiment of the present invention.

FIG. 8 is a diagram showing a recognition experiment result of handwritten alphanumeric kana by a recognition device to which the second embodiment of the present invention is applied.

FIG. 9 is a diagram for explaining the present invention, and is a diagram for explaining the manner in which candidates of a large classification are changed by modifying the feature selection dictionary.

FIG. 10 is a diagram showing an example of distribution of data belonging to three categories in a two-dimensional space and an example of a first principal axis obtained by principal component analysis for the data.

FIG. 11 is a flow chart showing a processing procedure of a learning phase in the second embodiment of the present invention.

FIG. 12 is a diagram for explaining the second embodiment of the present invention and is a diagram showing a reference pattern used for correcting a feature selection dictionary.

FIG. 13 is a diagram for explaining correction of the feature selection dictionary in the embodiment of the present invention.

[Explanation of symbols]

101 data input unit 102 feature extraction unit 103 feature selection unit 104 detailed identification dictionary 105 large classification dictionary creation unit 106 feature selection dictionary 107 large classification unit 108 feature selection dictionary correction unit 109 detailed identification unit 110 recognition result output unit 301 reference vector number 302 Category code 303 n-dimensional feature 601 Relationship between recognition performance and learning count according to the first embodiment 602 Recognition performance without major classification 701 Data input unit 702 Feature extraction unit 703 Feature selection unit 704 Detailed identification dictionary 705 Large Classification dictionary creation unit 706 Feature selection dictionary 707 Large classification unit 708 Feature selection dictionary correction unit 709 Detail identification unit 710 Recognition result output unit 901 Two-dimensional feature 902 extracted from learning pattern Feature axis 903 Determined by current feature selection dictionary 903 Feature selection dictionary Modified feature axis 904 Current feature selector One-dimensional reference pattern 905 of category A obtained by the one-dimensional reference pattern 906 of category B obtained by the current feature selection dictionary One-dimensional reference pattern 907 of category C obtained by the current feature selection dictionary 907 1-dimensional reference pattern 908 of A. 1-dimensional feature 909 after correction of the feature selection dictionary 1-dimensional reference pattern 910 of category B after correction of the feature selection dictionary 1-dimensional feature 1001 obtained by the current feature selection dictionary 1001 Category A in the 2-dimensional feature space Distribution range 1002 of category B in the two-dimensional feature space 1003 distribution range of category C in the two-dimensional feature space 1004 feature axis 1005 determined by the current feature selection dictionary two-dimensional feature 1006 extracted from the learning pattern based on the feature selection dictionary Obtained Category A
One-dimensional reference pattern 1007 of 1) One-dimensional feature 1008 obtained by converting the two-dimensional features extracted from the learning pattern based on the feature selection dictionary Category B obtained based on the feature selection dictionary
One-dimensional reference pattern 1009 of category C obtained based on the feature selection dictionary
1D reference patterns 1201 and 1202 Category D 2D reference patterns 1203 and 1204 Category E 2D reference patterns 1205 and 1206 Category F 2D reference patterns 1207 2D features 1208 extracted from learning patterns Defined by a feature selection dictionary 1D reference patterns 1213, 1214 obtained by projecting two-dimensional reference patterns of feature axes 1211, 1212 category D onto the feature axis based on the feature selection dictionary Two n belonging to the one-dimensional reference patterns 1301, 1302 category 1307 obtained by projecting the two-dimensional reference patterns of the one-dimensional reference patterns 1215, 1216 category F obtained by projecting onto the axis on the feature axis based on the feature selection dictionary. Dimensional reference patterns 1303, 13 4 two n two n-dimensional reference patterns 1307,1308,1309 Category 1310 learning pattern (two-dimensional) belonging to the dimension reference pattern 1305 and 1306 Category 1309 1311 and 1312 n-dimensional reference patterns belonging to the category 1308 1301,1
M-dimensional reference patterns 1313 and 1314 obtained by feature selection of 302 n-dimensional reference patterns 1303 and 1
M-dimensional reference patterns 1315 and 1316 obtained by selecting 304 as a feature n-dimensional reference patterns 1305 and 1316
M-dimensional reference pattern 1317 obtained by selecting features of 306 1 obtained by selecting learning patterns 1310
Dimensional learning pattern 1318 Distance between one-dimensional learning pattern 1317 and reference pattern R _C1 1319 Distance between one-dimensional learning pattern 1317 and reference pattern R _C2

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06T 7/00 G06K 9/62-9/72 G10L 15/06

Claims (16)

(57) [Claims] 1. A feature extraction unit for extracting n-dimensional features as primary features from an input pattern, and a feature selection dictionary storing a function for converting the primary features into m-dimensional (where m <n) features. A feature selection unit that extracts m-dimensional features from the primary features as secondary features based on a function stored in the feature selection dictionary; and details created in advance for recognition based on the primary features. A detailed identification dictionary storing identification reference patterns, and each detailed identification reference pattern included in the detailed identification dictionary are classified into m-dimensional large categories based on a function stored in the feature selection dictionary. Classification dictionary creating unit for converting and storing the reference patterns for use, and a predetermined number of upper ranks (K) that are close to the secondary features extracted from the feature selecting unit from the reference patterns for large classification
Of the large-classification reference patterns and outputs information for specifying the detailed identification reference patterns corresponding to the large-classification reference patterns; and a predetermined number (K) specified by the output information of the large-classification unit. A detailed identification unit that collates the detailed identification reference pattern and the primary feature extracted by the feature extraction unit and outputs a collation result; and a detailed identification unit that processes the collation result and outputs the result to a storage device or an output device. At the time of learning the recognition result output unit and the function stored in the feature selection dictionary, among the reference patterns included in the large classification reference pattern, the same category as the learning pattern (hereinafter, referred to as “correct category”) The large classification reference pattern that belongs to the learning pattern and has the shortest distance to the secondary feature is extracted as a reference pattern (A). Of the reference patterns included in the reference pattern, a reference pattern for large classification that belongs to a category that is not the same as the learning pattern (hereinafter, referred to as “incorrect solution category”) and has the closest distance to the secondary feature is set as the reference pattern (B). After the extraction, the feature selection function is modified so that the secondary feature and the reference pattern (A) are closer to each other, and the secondary feature and the reference pattern (B) are further away from each other. A pattern recognition device comprising: a feature selection dictionary correction unit; 2. A feature extraction unit that extracts n-dimensional features as primary features from an input pattern, and a feature selection dictionary that stores a function that converts the primary features into m-dimensional (where m <n) features. A feature selection unit that extracts m-dimensional features from the primary features as secondary features based on a function stored in the feature selection dictionary; and details created in advance for recognition based on the primary features. A detailed identification dictionary storing identification reference patterns, and an m-dimensional large classification reference for each detailed identification reference pattern included in the detailed identification dictionary based on a function stored in the feature selection dictionary. A large predetermined number (K) having a large degree of similarity between the large classification dictionary creating unit which converts the pattern into a pattern and stores the pattern and the secondary feature extracted by the feature selecting unit from the large classification reference pattern.
A) a large classification reference pattern for extracting the large classification reference patterns and outputting information for specifying the detailed identification reference patterns corresponding thereto; and a predetermined number (K) specified by the output information of the large classification part. ) The detailed identification reference pattern for collating the detailed identification reference pattern with the primary feature extracted by the feature extraction unit, and a detailed identification unit for outputting a collation result, and a recognition result output for processing and outputting the output of the detailed identification unit And a feature selection function stored in the feature selection dictionary, the similarity between the secondary feature extracted from the learning pattern among the reference patterns of the correct category included in the large classification reference pattern is The highest reference pattern for large classification is extracted as a reference pattern (A), and the reference pattern of the incorrect category included in the reference pattern for large classification is classified as the secondary feature. The reference pattern for the large classification with the highest similarity is the reference pattern (B)
After the extraction, the similarity between the secondary feature and the reference pattern (A) is increased, and the similarity between the secondary feature and the reference pattern (B) is further decreased.
And a feature selection dictionary modifying unit that modifies the feature selecting function. 3. A feature extraction unit that extracts an n-dimensional feature vector from an input pattern, and a feature selection dictionary that stores information for selecting an m-dimensional feature vector (where m <n) from the n-dimensional feature vector. A feature selection unit for extracting an m-dimensional feature vector for large classification from the n-dimensional feature vector extracted by the feature extraction unit with reference to the feature selection dictionary; and a detailed identification dictionary including a set of n-dimensional reference vectors. A large classification dictionary creating unit that extracts a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selecting dictionary; an m-dimensional feature vector extracted by the feature selecting unit and the large classification dictionary creating unit. The extracted m-dimensional reference vector is collated, and the higher order K is ranked in order of decreasing distance from the m-dimensional feature vector.
A large classification unit that extracts the numbers of the m-dimensional reference vectors as candidate reference vector numbers, and a predetermined number (K) in the detailed identification dictionary corresponding to the candidate reference vector numbers extracted by the large classification unit A detailed identification unit for collating the n-dimensional reference vector of n with the n-dimensional feature vector extracted by the feature extraction unit and outputting a collation result; and a recognition result for outputting a recognition result based on the output result of the detailed discrimination unit. At the time of learning the output unit and the feature selection dictionary, refer to the m-dimensional reference vector that is closest to the m-dimensional feature vector extracted from the learning pattern among the reference vectors of the correct category included in the m-dimensional reference vector. Of the reference vectors of the incorrect answer category that are extracted as the vector (A) and are included in the m-dimensional reference vector, After extracting the nearest m-dimensional reference vector away as a reference vector (B), the distance between the m-dimensional feature vector and the reference vector (A) approaches more,
Further, the m-dimensional feature vector and the reference vector (B)
And a feature selection dictionary correction unit that corrects the feature selection dictionary so that the distance from the feature selection dictionary is further increased. 4. A feature extraction unit for extracting an n-dimensional feature vector from an input pattern, and an m-dimensional feature vector from the n-dimensional feature vector (where m
Feature selection dictionary that stores information for selecting <n), and feature selection that extracts an m-dimensional feature vector for large classification from the n-dimensional feature vector extracted by the feature extraction unit with reference to the feature selection dictionary. Section, a detailed identification dictionary composed of a set of n-dimensional reference vectors, a general classification dictionary creating section for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary, and the feature selection section The m-dimensional feature vector extracted in step 1 is collated with the m-dimensional reference vector extracted by the large classification dictionary creating unit, and a predetermined upper number (K) of the m-dimensional reference vectors in descending order of similarity with the m-dimensional feature vector. A large classification unit that extracts a vector number as a candidate reference vector number, and a predetermined number (K pieces) in the detailed identification dictionary corresponding to the candidate reference vector numbers extracted by the large classification unit. A detailed identification unit for collating the n-dimensional reference vector of n with the n-dimensional feature vector extracted by the feature extraction unit and outputting a collation result; and a recognition result for outputting a recognition result based on the output result of the detailed discrimination unit. During learning of the output unit and the feature selection dictionary, among the reference vectors of the correct answer category included in the m-dimensional reference vector, the m-dimensional reference vector having the highest similarity to the m-dimensional feature vector extracted from the learning pattern is selected. Reference vector (A)
And extracting the m-dimensional reference vector having the highest similarity to the m-dimensional feature vector among the reference vectors of the incorrect category included in the m-dimensional reference vector as the reference vector (B). , The feature selection dictionary such that the degree of similarity between the m-dimensional feature vector and the reference vector (A) increases and the degree of similarity between the m-dimensional feature vector and the reference vector (B) further decreases. A pattern recognition device comprising: a feature selection dictionary correction unit that corrects. 5. A feature extraction unit for extracting an n-dimensional feature vector from an input pattern, and an m-dimensional feature vector from the n-dimensional feature vector (where m
Feature selection dictionary that stores information for selecting <n), and feature selection that extracts an m-dimensional feature vector for large classification from the n-dimensional feature vector extracted by the feature extraction unit with reference to the feature selection dictionary. Section, a detailed identification dictionary composed of a set of n-dimensional reference vectors, a general classification dictionary creating section for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary, and the feature selection section The m-dimensional feature vector extracted in step 1 is collated with the m-dimensional reference vector extracted by the large classification dictionary creating unit, and a predetermined number (K) of upper-rank m-dimensional reference vectors are arranged in the order of decreasing distance from the m-dimensional feature vector. Number as a candidate reference vector number and a predetermined number (K) in the detailed identification dictionary corresponding to the candidate reference vector number extracted by the large classification unit. A detailed identification unit that collates the dimensional reference vector with the n-dimensional feature vector extracted by the feature extraction unit and outputs a collation result, and a recognition result output unit that outputs the recognition result based on the output result of the detailed discrimination unit. During learning of the feature selection dictionary, an n-dimensional reference vector closest in distance to the n-dimensional feature vector extracted from the learning pattern is selected from n-dimensional reference vectors of correct categories included in the detailed identification dictionary. , An m-dimensional reference vector corresponding to this is extracted as a reference vector (C), and further, of the incorrect m-dimensional reference vectors in the m-dimensional reference vector, the m-dimensional reference vector closest to the m-dimensional feature vector. Is extracted as a reference vector (B), the distance between the m-dimensional feature vector and the m-dimensional reference vector (C) becomes closer, and Wherein the m-dimensional feature vector distance between the m-dimensional reference vector (B) is to be more away, and a feature selection dictionary modification unit for modifying the feature selection function, the pattern recognition apparatus characterized by. 6. A feature extraction unit for extracting an n-dimensional feature vector from an input pattern, and an m-dimensional feature vector from the n-dimensional feature vector (where m
Feature selection dictionary that stores information for selecting <n), and feature selection that extracts an m-dimensional feature vector for large classification from the n-dimensional feature vector extracted by the feature extraction unit with reference to the feature selection dictionary. Section, a detailed identification dictionary composed of a set of n-dimensional reference vectors, a general classification dictionary creating section for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary, and the feature selection section The m-dimensional feature vector extracted in step 1 is collated with the m-dimensional reference vector extracted by the large classification dictionary creating unit, and a predetermined upper number (K) of the m-dimensional reference vectors in descending order of similarity with the m-dimensional feature vector. A large classification unit that extracts a vector number as a candidate reference vector number, and a predetermined number (K pieces) in the detailed identification dictionary corresponding to the candidate reference vector numbers extracted by the large classification unit. A detailed identification unit for collating the n-dimensional reference vector of n with the n-dimensional feature vector extracted by the feature extraction unit and outputting a collation result; and a recognition result for outputting a recognition result based on the output result of the detailed discrimination unit. At the time of learning the output unit and the feature selection dictionary, an n-dimensional reference vector having the highest similarity with the n-dimensional feature vector extracted from the learning pattern among the n-dimensional reference vectors of the correct category included in the detailed identification dictionary A vector is selected, an m-dimensional reference vector corresponding to the vector is extracted as a reference vector (C), and further, of the incorrect m-dimensional reference vectors in the m-dimensional reference vector, the similarity with the m-dimensional feature vector is the largest. By extracting the m-dimensional reference vector as the reference vector (B), the similarity between the m-dimensional feature vector and the reference vector (C) is further increased. Further, the m
A feature selection dictionary correction unit that corrects the feature selection function so that the degree of similarity between the dimensional feature vector and the reference vector (B) is further reduced. 7. The primary features of the input pattern are m-dimensional (where
For each category, a feature selection dictionary that stores feature selection information for converting into features of m <n) and a reference pattern for detailed identification created by learning in advance for recognition based on the primary feature of the input pattern A detailed identification dictionary to be stored; and storage means for storing the detailed identification dictionary, based on the contents of the feature selection dictionary from the n-dimensional reference pattern in the detailed identification dictionary and the n-dimensional features extracted from the learning pattern. , A means for extracting an m-dimensional reference pattern and an m-dimensional feature, and a reference pattern having the closest distance to the m-dimensional feature among the m-dimensional reference patterns belonging to the same category as the learning pattern, the first reference pattern. And a reference pattern closest in distance to the m-dimensional feature among the m-dimensional reference patterns belonging to a category different from the learning pattern. And a means for extracting as the second reference pattern, so that the distance between the m-dimensional feature and the first reference pattern becomes closer, and the distance between the m-dimensional feature and the second reference pattern becomes further away. A pattern recognition device comprising: a means for modifying the contents of a feature selection dictionary. 8. A pattern recognition apparatus, wherein the primary features (n dimensions) of an input pattern are m dimensions (where m
Is a positive integer smaller than n), a feature selection dictionary that stores a function for converting into features, and a detailed identification dictionary that stores a detailed identification reference pattern that is created in advance for recognition based on the primary features, (A) feature extraction processing for extracting n-dimensional features as primary features from the input pattern, and (b) 2) based on the function stored in the feature selection dictionary.
Feature selection processing for extracting an m-dimensional feature from the primary feature as the next feature, and (c) each detail identification reference pattern included in the detail identification dictionary is stored in the feature selection dictionary. A large classification dictionary creation process of converting to and storing an m-dimensional large classification reference pattern based on a function; (d) distance from the large classification reference pattern to the secondary feature extracted by the feature selecting unit A large-classification process of extracting a predetermined number (K) of large-classification reference patterns that are close to each other, and outputting information for specifying the detailed-identification reference pattern corresponding thereto; and (e) the large-classification process. A detailed identification process of collating a predetermined number (K) of the reference patterns for detailed identification specified by the output information with the primary features extracted by the feature extraction process, and outputting a collation result; f) The verification result A recognition result output process of processing the fruit and outputting the result to a storage device or an output device; and (g) at the time of learning the function stored in the feature selection dictionary, among the reference patterns included in the large classification reference pattern, A large classification reference pattern that belongs to the same category as the learning pattern (hereinafter referred to as “correct answer category”) and that is closest to the secondary feature extracted from the learning pattern is extracted as a reference pattern (A), Further, among the reference patterns included in the large-classification reference pattern, the large-classification reference that belongs to a category that is not the same as the learning pattern (hereinafter, referred to as “incorrect solution category”) and has the shortest distance from the secondary feature. The pattern is extracted as a reference pattern (B), the distance between the secondary feature and the reference pattern (A) becomes closer, and the secondary feature and the reference pattern are further reduced. A computer that configures the pattern recognition device, and a feature selection dictionary correction process that corrects a function of the feature selection dictionary so that the distance from the pattern recognition device is further increased. A recording medium that records a program to be executed in. 9. A pattern recognition apparatus, wherein the primary features (n dimensions) of an input pattern are m dimensions (where m
Is a positive integer smaller than n), a feature selection dictionary that stores a function for converting into features, and a detailed identification dictionary that stores a detailed identification reference pattern that is created in advance for recognition based on the primary features, (A) feature extraction processing for extracting n-dimensional features as primary features from the input pattern; and (b) secondary features based on a function stored in the feature selection dictionary. A feature selection process for extracting an m-dimensional feature from the next feature, and (c) an m-dimensional feature of each detailed identification reference pattern included in the detailed identification dictionary, based on a function stored in the feature selection dictionary. A large-classification dictionary creating process of converting the large-classification reference pattern into a large-classification reference pattern and storing the large-classification reference pattern; (K (B) a large-classification process of extracting the large-classification reference patterns and outputting information for specifying the detailed identification reference patterns corresponding to these; (e) a predetermined number (a) specified by the output information of the large-classification unit ( (K) the detailed identification reference patterns and the primary features extracted by the feature extraction unit are collated, and the detailed identification processing for outputting the collation result; (f) The output of the detailed identification unit is processed. In the recognition result output process of outputting to the storage device or the output device, and (g) during learning of the feature selection function stored in the feature selection dictionary, among the reference patterns of the correct category included in the reference pattern for large classification, The large classification reference pattern extracted from the learning pattern and having the highest degree of similarity to the secondary feature is extracted as a reference pattern (A), and further, the incorrect answer pattern included in the large classification reference pattern is extracted. Among reference pattern Gori, reference pattern rough classification reference pattern highest similarity with the secondary features (B)
After the extraction, the similarity between the secondary feature and the reference pattern (A) is increased, and the similarity between the secondary feature and the reference pattern (B) is further decreased.
A recording medium which records a program for causing a computer constituting the pattern recognition device to execute the feature selection dictionary correction process for correcting the function of the feature selection dictionary and the processes (a) to (g). 10. In a pattern recognition device, m from a primary feature (n-dimensional feature vector) of an input pattern
A feature selection dictionary that stores information for selecting a dimensional feature vector (where m is a positive integer smaller than n), and a detailed identification dictionary that includes a set of n-dimensional reference vectors,
(A) a feature extraction process for extracting an n-dimensional feature vector from an input pattern; and (b) an n-dimensional feature vector extracted by the feature extraction process with reference to the feature selection dictionary. A feature selection process for extracting an m-dimensional feature vector for major classification from (c) a major dictionary creation process for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary; d) The m-dimensional feature vector extracted by the feature selection process and the m-dimensional reference vector extracted by the large classification dictionary creation process are collated, and a predetermined number of upper ranks (K) are arranged in the order of close distance to the m-dimensional feature vector. (B) major classification processing for extracting the number of the m-dimensional reference vector as a candidate reference vector number; and (e) before corresponding to the candidate reference vector number extracted in the major classification processing. Predetermined number in detail identification dictionary (K
Individual) n-dimensional reference vector and the n-dimensional feature vector extracted by the feature extraction unit are collated and a collation result is output, and (f) a recognition result based on the output result of the detail discrimination process. Recognition result output processing for outputting to the storage device or the output device, and (g) during learning of the feature selection dictionary, among the reference vectors of the correct category included in the m-dimensional reference vector,
An m-dimensional reference vector that is closest to the m-dimensional feature vector extracted from the learning pattern is extracted as a reference vector (A), and further, of the reference vectors of the incorrect solution category included in the m-dimensional reference vector, After the m-dimensional reference vector having the closest distance to the m-dimensional feature vector is extracted as the reference vector (B), the distance between the m-dimensional feature vector and the reference vector (A) becomes closer, and the m-dimensional feature vector is further reduced. And the reference vector (B) so that the distance between the reference vector (B) and the feature selection dictionary is corrected so that the distance between the reference vector (B) and the reference vector (B) is further increased. A recording medium recording a program to be executed by a computer. 11. In a pattern recognition device, m from a primary feature (n-dimensional feature vector) of an input pattern
A feature selection dictionary that stores information for selecting a dimensional feature vector (where m is a positive integer smaller than n), and a detailed identification dictionary that includes a set of n-dimensional reference vectors,
(A) a feature extraction process for extracting an n-dimensional feature vector from an input pattern; and (b) an n-dimensional feature vector extracted by the feature extraction process with reference to the feature selection dictionary. A feature selection process for extracting an m-dimensional feature vector for large classification from (c) a large classification dictionary creation process for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary; d) The m-dimensional feature vector extracted by the feature selection process and the m-dimensional reference vector extracted by the large classification dictionary creation process are collated, and a predetermined number of higher ranks in descending order of similarity with the m-dimensional feature vector ( Corresponding to the candidate reference vector numbers extracted in the large classification processing; and (e) the large classification processing for extracting the numbers of the K) m-dimensional reference vectors as candidate reference vector numbers. A predetermined number in the serial detailed identification dictionary (K number)
And a detailed identification process for collating the n-dimensional reference vector of n with the n-dimensional feature vector extracted by the feature extraction process and outputting a collation result; (f) a recognition result based on the output result of the detailed discrimination process. During the recognition result output process of outputting and (g) learning of the feature selection dictionary, among the reference vectors of the correct category included in the m-dimensional reference vector,
An m-dimensional reference vector having the highest similarity to the m-dimensional feature vector extracted from the learning pattern is extracted as a reference vector (A), and further, of the reference vectors of the incorrect category included in the m-dimensional reference vector, After extracting the m-dimensional reference vector having the highest similarity to the m-dimensional feature vector as the reference vector (B), the similarity between the m-dimensional feature vector and the reference vector (A) is further increased, and In order to reduce the similarity between the m-dimensional feature vector and the reference vector (B),
A recording medium which records a program for causing a computer constituting the pattern recognition device to execute the feature selection dictionary correction process for correcting the feature selection dictionary and each of the processes (a) to (g). 12. A pattern recognition apparatus, wherein m is calculated from a primary feature (n-dimensional feature vector) of an input pattern.
A feature selection dictionary that stores information for selecting a dimensional feature vector (where m is a positive integer smaller than n), and a detailed identification dictionary that includes a set of n-dimensional reference vectors,
(A) a feature extraction process for extracting an n-dimensional feature vector from an input pattern; and (b) an n-dimensional feature vector extracted by the feature extraction process with reference to the feature selection dictionary. A feature selection process for extracting an m-dimensional feature vector for major classification from (c) a major dictionary creation process for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary; d) The m-dimensional feature vector extracted by the feature selection processing is collated with the m-dimensional reference vector extracted by the large classification dictionary creating unit, and a predetermined number of higher ranks (K Number) of the m-dimensional reference vector as a candidate reference vector number, and (e) before corresponding to the candidate reference vector number extracted in the large classification process. Predetermined number in detail identification dictionary (K
Individual) n-dimensional reference vector and the n-dimensional feature vector extracted by the feature extraction process, and a detailed identification process for outputting a comparison result; (f) a recognition result based on the output result of the detailed identification process. And (g) at the time of learning of the feature selection dictionary, the n-dimensional feature vector extracted from the learning pattern among the n-dimensional reference vectors of correct categories included in the detailed identification dictionary. Select the n-dimensional reference vector with the closest distance
An m-dimensional reference vector corresponding to this is extracted as a reference vector (C), and further, of the incorrect m-dimensional reference vectors in the m-dimensional reference vector, the m-dimensional reference vector closest to the m-dimensional feature vector is selected. As the reference vector (B), the distance between the m-dimensional feature vector and the m-dimensional reference vector (C) becomes closer, and the distance between the m-dimensional feature vector and the m-dimensional reference vector (B) becomes larger. A program for causing the computer constituting the pattern recognition device to execute the feature selection dictionary correction process for correcting the function of the feature selection dictionary and the respective processes of (a) to (g) described above is recorded so as to move away. recoding media. 13. In a pattern recognition device, m from a primary feature (n-dimensional feature vector) of an input pattern
A storage device for storing a feature selection dictionary that stores information for selecting a dimensional feature vector (where m is a positive integer smaller than n) and a detailed identification dictionary that is a set of n-dimensional reference vectors (A) a feature extraction process for extracting an n-dimensional feature vector from an input pattern, and (b) an n-dimensional feature vector for large classification from the n-dimensional feature vector extracted by the feature extraction process with reference to the feature selection dictionary. A feature selection process for extracting a feature vector; (c) a large classification dictionary creation process for extracting a set of m-dimensional reference vectors from the detailed identification dictionary based on the feature selection dictionary; and (d) a feature selection process. The extracted m-dimensional feature vector and the m-dimensional reference vector extracted in the large classification dictionary creation process are collated, and a predetermined number (K) of higher ranks are ranked in descending order of similarity with the m-dimensional feature vector. A large classification process for extracting the number of the m-dimensional reference vector as a candidate reference vector number; and (e) a predetermined number (K pieces) in the detailed identification dictionary corresponding to the candidate reference vector number extracted in the large classification process. )
Detailed identification processing for collating the n-dimensional reference vector of n with the n-dimensional characteristic vector extracted by the feature extraction processing and outputting the collation result; (f) outputting the recognition result based on the output result of the detailed discrimination processing. (G) Similarity to the n-dimensional feature vector extracted from the learning pattern among the n-dimensional reference vectors of the correct category included in the detailed identification dictionary when learning the feature selection dictionary. The n-dimensional reference vector with the highest degree is selected, the m-dimensional reference vector corresponding to this is extracted as a reference vector (C), and the m-dimensional feature vector among the incorrect m-dimensional reference vectors in the m-dimensional reference vector is extracted. The m-dimensional reference vector having the highest similarity to is extracted as a reference vector (B), and the m-dimensional feature vector and the reference vector (C And a feature selection dictionary modification process for modifying the function of the feature selection dictionary so that the similarity between the feature selection dictionary and the reference vector (B) further decreases. A recording medium on which a program for causing the computer constituting the pattern recognition apparatus to execute the processes (a) to (g) is recorded. 14. A feature selection dictionary storing feature selection information for converting a primary feature of an input pattern into an m-dimensional (where m <n) feature, and for recognition based on the primary feature of the input pattern. A detailed identification dictionary that stores in advance a detailed identification reference pattern for each category; and a storage unit that stores the detailed identification dictionary, and is extracted from the n-dimensional reference pattern and the learning pattern in the detailed identification dictionary. Based on the content of the feature selection dictionary, the m-dimensional reference pattern and the m-dimensional feature are extracted from the n-dimensional feature, and the m-dimensional feature among the m-dimensional reference patterns that belong to the same category as the learning pattern. The reference pattern having the closest distance is extracted as a first reference pattern, and the distance to the m-dimensional feature among the m-dimensional reference patterns belonging to a category different from the learning pattern is extracted. Is extracted as the second reference pattern, and the distance between the m-dimensional feature and the first reference pattern becomes closer, and the distance between the m-dimensional feature and the second reference pattern is further reduced. A pattern recognition method, characterized in that the contents of the feature selection dictionary are updated so as to be farther away. 15. A feature selection dictionary storing feature selection information for converting a primary feature of an input pattern into an m-dimensional (where m <n) feature, and recognition based on the primary feature of the input pattern. The data processing device includes a feature extraction unit, a feature selection unit, and a large classification dictionary, and a storage unit that stores a detail identification dictionary that stores in advance a detail identification reference pattern for each category. A pattern recognition method of a pattern recognition device, comprising: a section, a large classification section, a detailed identification section, a recognition result output section, and a feature selection dictionary correction section, wherein a learning phase includes (a) the feature extraction. A step of extracting an n-dimensional feature from the input pattern as a primary feature in the unit, and (b) the feature selecting unit, based on the function stored in the feature selection dictionary, from the primary feature as the secondary feature. extracting an m-dimensional feature, and (c) in the large classification dictionary creation unit, each detailed identification reference pattern included in the detailed identification dictionary is converted into a function stored in the feature selection dictionary. A step of converting to and storing the m-dimensional reference pattern for large classification based on the following: (d) In the large classification unit, the distance to the secondary feature extracted by the feature selection unit from the large classification reference pattern. A step of extracting a predetermined number (K) of large-classification reference patterns close to each other and outputting information for specifying the detailed identification reference patterns corresponding to these; (e) in the detailed identification section, A predetermined number (K) of the detailed identification reference patterns specified by the output information of the large classification process and 1 extracted by the feature extraction process.
Collating the following features and outputting a collation result; (f) a recognition result output process of processing the collation result and outputting the result to a storage device or an output device; (g) the feature selection dictionary correction unit In learning the function stored in the feature selection dictionary, among the reference patterns included in the large-classification reference pattern, the secondary features belonging to the same category as the learning pattern and extracted from the learning pattern. The large-classification reference pattern having the closest distance to is extracted as a first reference pattern, and further belongs to a category that is not the same as the learning pattern among the reference patterns included in the large-classification reference pattern, and the secondary pattern The large-classification reference pattern closest to the feature is extracted as a second reference pattern, and the secondary feature and the first feature are extracted.
Of the feature selection dictionary such that the distance between the secondary feature and the second reference pattern is further reduced as the distance between the second reference pattern and the second reference pattern is further reduced. Pattern recognition method. 16. In the recognition phase, (h) as initialization processing, the large classification dictionary creating section refers to the feature selection dictionary and the detailed identification dictionary created in advance in the learning phase, and refers to an m-dimensional reference. Creating and storing a pattern, (i) reading a pattern to be recognized, extracting n-dimensional features as primary features from the input pattern by the feature extraction unit, (j) the feature selection dictionary in the feature selection unit M-dimensional feature patterns are extracted from the primary features as secondary features based on the function stored in (4), (k) in the large classification unit, each of the detailed identification references included in the detailed identification dictionary. An m-dimensional reference pattern extracted by a large classification dictionary creating unit that extracts and stores an m-dimensional large classification reference pattern based on a function stored in the feature selection dictionary. A predetermined number of patterns having a distance close to the m-dimensional characteristic pattern are extracted from among the reference pattern information, and the reference pattern information is output to the detailed identification unit, and (l) the reference output by the large classification unit in the detailed identification unit. The n-dimensional reference pattern corresponding to the pattern number is extracted from the detailed identification dictionary, and the one having a distance close to the n-dimensional feature pattern output by the feature extraction unit is output from the dictionary, (m) the recognition result The pattern recognition method according to claim 15, wherein the output unit processes the recognition result output by the detailed identification unit and outputs the processed result to an output device.