JPH11345337A - Pattern dictionary generation device and method therefor, and machine-readable recording medium has recorded pattern dictionary generation method in computer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a pattern recognition dictionary whose recognition precision is high and which can recognize a pattern at a high speed. SOLUTION: The pattern dictionary generation device generating a dictionary for pattern recognition is provided with an input part 101 inputting a learning sample in the form of a feature vector, an average template deciding part 102 obtaining average templates representing the learning samples inputted from the input part 101 for every category, a main component analysis part 103 obtaining intrinsic value/intrinsic vectors (main components) for every category from the learning samples inputted from the input part 101, a displacement deciding part 104 deciding template displacement based on the intrinsic value/ intrinsic vector (main component) obtained in the main component analysis part 103 and a template registration part 105 registering a displacement position obtained in the displacement deciding part 104 in a dictionary 106 as the template.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern dictionary creating apparatus and a pattern dictionary creating method for creating a pattern dictionary generally used for pattern recognition including character recognition, and a machine-readable recording in which the pattern dictionary creating method is recorded in a computer. Regarding the medium.

[0002]

2. Description of the Related Art Heretofore, as a reference technical document relating to character recognition provided with a multi-template type dictionary, as a first conventional example, instead of using a new cluster center as a template as it is, all samples in the original category are used. The recognition accuracy can be improved by performing the correction using the average template of the IEICE Technical Report (PRU86
-36, 1986, July) in "Basic Study on Identification of Handwritten Characters".

Further, as a second conventional example, a dictionary of a multi-template type is provided, a feature value vector is extracted from an input character, and compared with an average feature value vector of each class. Japanese Patent Application Laid-Open No. 8-180139 discloses a "method and apparatus for unrestricted handwritten alphanumeric online recognition" which is recognized as a model character corresponding to an average feature value vector closest to a vector.

Further, as a third conventional example, a method with high recognition accuracy even for a single template is described in the IEICE Technical Report (PRU92-115, pp99-106, 1993,
(Jan.), “Consideration on Dimensionality and Variable Conversion of Feature in Character Recognition”. In this method, not only the average template is used to represent a category, but also eigenvalues / eigenvectors are used, thereby trying to represent the sample distribution within the category.

[0005]

However, in the second conventional example as described above, as pointed out in the first conventional example, a multi-template dictionary based on simple clustering is used. However, there is a problem that the recognition accuracy often decreases rather than increases the dictionary capacity.

Further, in the third conventional example, since it is necessary to use a higher-order discriminant function, there is a problem that the amount of calculation required for the recognition process is increased.

The present invention has been made in view of the above, and an object of the present invention is to provide a pattern recognition dictionary which has high recognition accuracy and can realize high-speed pattern recognition.

[0008]

According to a first aspect of the present invention, there is provided a pattern dictionary creating apparatus for creating a dictionary for pattern recognition. Input means for inputting in the form of a feature vector, average template determining means for obtaining, for each category, an average template representing the learning sample input from the input means, and eigenvalues for each category from the learning sample input from the input means. Principal component analysis means for obtaining an eigenvector (principal component); displacement determination means for determining a template displacement based on the eigenvalue and eigenvector (principal component) obtained by the principal component analysis means; Registering means for registering the displaced position in the dictionary as a template,
It is provided with.

According to a second aspect of the present invention, the template registering means combines the displacement templates and adds the combined displacement templates.

According to a third aspect of the present invention, the displacement determining means determines the number of templates based on the eigenvalues obtained by the principal component analyzing means.

According to a fourth aspect of the present invention, in the pattern dictionary creating apparatus, the template registering means deletes unnecessary templates which do not contribute to recognition accuracy among the created templates.

According to a fifth aspect of the present invention, in the pattern dictionary creating method for creating a dictionary for pattern recognition, a first step of inputting a learning sample in the form of a feature vector; A second step of obtaining, for each category, an average template representing the learning sample input in the first step, and eigenvalues / eigenvectors (principal components) for each category from the learning sample input in the second step. A third step of determining, a fourth step of determining a template displacement based on the eigenvalues / eigenvectors (principal components) determined in the third step, and a fourth step of determining the template displacement.
And a fifth step of registering the displacement position obtained in the step as a template in a dictionary.

According to a sixth aspect of the present invention, there is provided a machine-readable recording medium on which a program for causing a computer to execute the pattern dictionary creating method according to the fifth aspect is recorded.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A pattern dictionary creating apparatus, a pattern dictionary creating method, and a machine readable recording medium having a pattern dictionary creating method recorded in a computer according to the present invention will be described below in detail with reference to the accompanying drawings.

(Principle of the Invention) The principle of the present invention is that distance calculation in consideration of the distribution of categories is complicated and takes a long processing time, so that a good approximation is obtained by adopting a multi-template configuration. is there. This good approximation means that the template is arranged at a position that closely approximates the distribution using the eigenvalues / eigenvectors (principal components) of the category.

On the other hand, the difference from the conventional multi-template type is that a template is determined from eigenvalues / eigenvectors of a category instead of clustering. For this reason, the recognition accuracy of the pattern dictionary of the present invention is higher than that of the conventional multi-template dictionary.

(System Configuration) FIG. 1 is a block diagram showing a system configuration of a pattern dictionary creating apparatus according to an embodiment of the present invention. Reference numeral 101 denotes an input unit serving as an input unit for inputting a learning sample in the form of a feature vector. Reference numeral 102 denotes an average template determining unit serving as an average template determining unit for obtaining an average template representing the learning sample input from the input unit 101 for each category. , 103 are principal component analysis units as principal component analysis means for finding eigenvalues / eigenvectors (principal components) for each category from the learning sample input from the input unit 101, and 104 is a principal component analysis unit 10.
3. A displacement determination unit as a displacement determination unit that determines a template displacement based on the eigenvalues / eigenvectors (principal components) obtained in step 3; a template registration unit 105 that registers the displacement position obtained by the displacement determination unit 104 as a template in a dictionary The template registration unit 106 as a means is a dictionary.

(Processing Operation of System) Next, the processing operation of the system in the pattern dictionary creating apparatus configured as described above will be described. FIG. 2 is a flowchart showing an example of a processing operation for creating a pattern dictionary according to the embodiment of the present invention.

In FIG. 2, when the process is started, first, a learning sample is input from the input unit 101 (S20).
1). At this time, the learning sample is given in the form of a feature vector. Next, the average template determination unit 102
, An average vector is calculated for each category, and this is registered in the dictionary 106 as an average template (S2).
02).

Further, the principal component analysis unit 103 performs a principal component analysis for each category using the average vector (S203), and calculates eigenvalues and eigenvectors representing the distribution of each category. These values are the displacement determination unit 1
Then, the displacement is determined by the displacement determination unit 104 (S204). Then, a displacement template is registered (S205), and when the processing for all the categories is completed, the processing ends.

Here, the number of dimensions of the feature vector is n, and the average vector of the target category is M. The eigenvalues calculated for the target category are assigned as λ _i
(I = 1..n), and the corresponding eigenvector is Φ _i (i = 1..n).

For the eigenvalues and eigenvectors, m (m ≦ n) eigenvalues and the corresponding eigenvectors are used in descending order of value. That the eigenvalues lambda _i is large, it means that the variance of feature vector distribution with respect to the direction of the eigenvector [Phi _i corresponding about the M is large. That is, regarding the direction, there are some which are far from M even though they belong to the same category.

To compensate for this, consider adding a displacement template. For example, as the displacement calculation example shown in FIG. 3, when the predetermined constant as alpha, position M ⁺ _i of the displacement template, M ^- _{i, ^{respectively, M + i = M + α}} × λ i 1/2 × Φ _i M ⁻ _i = M−α × λ _i ^1/2 × Φ _i

As a result, even if the distance from M is long, the distance is closer to the added displacement template, so that they can be correctly identified.
At this time, the number of templates is increased because the calculation can be performed by a simple method such as Euclidean distance or city block distance instead of a method with a large amount of calculation like the Bayesian discriminant function. This contributes to speeding up the recognition processing.

Now, an example in which the template registration unit 105 has a function of adding a combined displacement template obtained by combining displacement templates will be described. That is, as shown in FIG. 4, the template registration unit 105 combines displacement templates for different i, and adds a combined displacement template. Note that a black circle in FIG. 4 is a combined displacement template.

The displacement determining unit 104 determines the number of templates based on the eigenvalues obtained by the principal component analyzing unit 103. That is, as shown in FIG. 5, the number of templates is determined according to the displacement eigenvalue. Thereby, the number of templates can be increased for an axis having a larger eigenvalue. This can be used together with the function of adding the combined displacement template by the template registration unit 105 described above.

Further, the template registration section 105 has a function of deleting unnecessary templates. The fact that it is unnecessary means that the contribution of the recognition accuracy improvement flat is small.

The present invention can be realized by software other than the above-described embodiment. FIG. 6 shows a configuration example of a computer system when the present invention is realized by software.

In FIG. 6, 601 is an input interface, 602 is a CPU for controlling the whole system based on a control program, 603 is a storage device, 604 is an output interface, and 605 is a drive for driving a recording medium 606 such as a CD-ROM. It is.

In the computer system configured as described above, the program of the above-described pattern dictionary creating method is recorded in advance on a recording medium 606 such as a CD-ROM. Then, the data is read through the drive 605. The CPU 602 executes each step of the program while exchanging information with the storage device 603.

[0031]

As described above, according to the pattern dictionary creating apparatus of the present invention (claim 1), input means for inputting a learning sample in the form of a feature vector, and representing the input learning sample. Mean template determination means for obtaining an average template for each category, principal component analysis means for obtaining an eigenvalue / eigenvector (principal component) for each category from an input learning sample, and template displacement based on eigenvalue / eigenvector (principal component) A pattern recognition dictionary that has high recognition accuracy and can realize high-speed pattern recognition is created because it is composed of a displacement determining means for determining and a template registering means for registering the obtained displacement position as a template in a dictionary. An apparatus can be provided.

According to the pattern dictionary creating apparatus of the present invention (claim 2), in addition to the effect of claim 1, it is possible to further improve recognition accuracy.

According to the pattern dictionary creating apparatus of the present invention (claim 3), in addition to the effect of claim 1, the number of templates can be increased for an axis having a larger eigenvalue.
It is possible to further improve recognition accuracy.

Further, according to the pattern dictionary creating apparatus of the present invention (claim 4), unnecessary templates that contribute little to improvement of recognition accuracy are deleted at the time of template registration, so that the capacity of the dictionary can be effectively utilized. , And the size can be reduced.

According to the pattern dictionary creating method of the present invention (claim 5), a learning sample is input in the form of a feature vector, and an average template representing the input learning sample is obtained for each category. Further, eigenvalues / eigenvectors (principal components) are obtained for each category from the input learning samples, template displacements are determined based on the obtained eigenvalues / eigenvectors (principal components), and the displacement positions are registered in the dictionary as templates. Therefore, it is possible to provide an apparatus for creating a pattern recognition dictionary that has high recognition accuracy and can realize high-speed pattern recognition.

Further, according to the machine-readable recording medium of the present invention (claim 6), a program for causing a computer to execute the pattern dictionary creation method described in claim 2 is recorded, so that Can be realized by a computer.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a system configuration of a pattern dictionary creation device according to an embodiment of the present invention.

FIG. 2 is a flowchart illustrating an example of a processing operation of creating a pattern dictionary according to the embodiment of the present invention.

FIG. 3 is an explanatory diagram showing a displacement calculation example according to the embodiment of the present invention.

FIG. 4 is an explanatory diagram showing an example of combining displacement templates according to the embodiment of the present invention.

FIG. 5 is an explanatory diagram showing an example of determining the number of templates according to an eigenvalue according to the embodiment of the present invention.

FIG. 6 is a block diagram illustrating a configuration example of a computer system according to an embodiment of the present invention.

[Explanation of symbols]

 Reference Signs List 101 input unit 102 average template determination unit 103 principal component analysis unit 104 displacement determination unit 105 template registration unit 106 dictionary

Claims (6)

[Claims] In a pattern dictionary creating apparatus for creating a dictionary for pattern recognition, an input means for inputting a learning sample in the form of a feature vector, and an average template representative of the learning sample input from the input means are classified into categories. Means for determining an average template for each category, principal component analysis means for obtaining eigenvalues / eigenvectors (principal components) for each category from the learning samples input from the input means, and eigenvalues / eigenvectors ( A pattern dictionary creating apparatus, comprising: displacement determining means for determining a template displacement based on the principal component); and template registering means for registering a displacement position obtained by the displacement determining means in a dictionary as a template. . 2. The pattern dictionary creating apparatus according to claim 1, wherein said template registering means combines the displacement templates and adds the combined displacement templates. 3. The pattern dictionary creating apparatus according to claim 1, wherein said displacement determining means determines the number of templates based on the eigenvalues obtained by said principal component analyzing means. 4. The pattern dictionary creation device according to claim 1, wherein the template registration unit deletes unnecessary templates that do not contribute to recognition accuracy from the created templates. apparatus. 5. In a pattern dictionary creating method for creating a dictionary for pattern recognition, a first step of inputting a learning sample in the form of a feature vector, and representing the learning sample input in the first step. A second step of obtaining an average template for each category; a third step of obtaining eigenvalues / eigenvectors (principal components) for each category from the learning samples input in the second step; A fourth step of determining a template displacement based on the obtained eigenvalues / eigenvectors (principal components);
And a fifth step of registering the displacement position obtained in the step as a template in a dictionary. 6. A machine-readable recording medium on which a program for causing a computer to execute the pattern dictionary creating method according to claim 5 is recorded.