JPH11203414A - Broadly classified dictionary preparing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a broadly classified dictionary preparing device capable of preparing a broadly classified dictionary having high recognition ratios. SOLUTION: A feature value extraction part 12 extracts the feature values of all sample images in a character image data file 11, calculates the mean value of these feature values and stores the mean value in a feature value storing part 13 in each character sort. A 1st broadly classified dictionary preparing part 14 prepares a primary broadly classfied dictionary by classifying the feature values of respective character sorts stored in the storing part 13 into categories by a 1st distance calculation method. A 2nd broadly classified dictionary preparing part 15 prepares a secondary broadly classified dictionary by calculating the sum of distances between the mean of categories at the time of changing the categories of character species and respective character species in the categories by a total distance calculation part 16 by using a 2nd distance calculation method, moving the character species to the minimum category and reconstituting the primary broadly classified dictionary.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a large classification dictionary creating apparatus for creating a large classification dictionary used in a large classification process for classifying into similar character categories as preprocessing of a detailed recognition process in a character recognition device or the like. It is.

[0002]

2. Description of the Related Art In a character recognition device or the like, in order to reduce the time required for recognition, a multidimensional feature is extracted from a plurality of input characters, and a category of characters similar in the feature is extracted. There is a method that uses a large classification process that classifies the data into two types. Of course, after the input characters are roughly classified, they are recognized in detail using more detailed feature amounts, and are determined as one character type.

As a technique using such a conventional technique, there is a technique described in, for example, Japanese Patent Application Laid-Open No. 1-161592. That is, a group of characters having similar feature amounts in the recognition dictionary is categorized, a standard feature amount is obtained from these average feature amounts, and a large classification dictionary is created as a category feature amount. Then, the feature amount of the input character image data is extracted and matched with the large classification dictionary,
Select the closest category. Further, a final character candidate is determined by matching the feature amount with each character in the selected category.

Japanese Patent Application Laid-Open No. 5-128312 discloses a technique for performing a large classification process accurately by using a more global character feature amount as a feature amount used when creating a large classification dictionary. ing.

Further, Japanese Patent Laid-Open No. 4-266188 discloses that an initial seed point is determined by using an average of all sample characters for creating a dictionary, and a sample farthest from all the seed points is successively determined. By setting the seed point of
A dictionary creation device that creates a large classification dictionary efficiently has been disclosed.

In each of the above-mentioned prior arts, a large classification dictionary is created by categorizing using the average between samples within a character type and the distance between the averages of each character type using one distance calculation method. In the created large classification dictionary, a square error between an average of the categories and a sample in the category is generally used as a scale indicating the goodness of unity of the categories. The smaller the square error is, the better the sample in the category is. The category is reconfigured so as to minimize the square error.

As a method of calculating the square error at this time, the Euclidean distance is generally used. However, in actual recognition processing, a distance calculation method that uses a distribution of character types within a category and has a higher calculation cost and a higher recognition rate is often used. in this case,
If a large classification dictionary that minimizes the square error with the Euclidean distance is used, there is a problem that the recognition rate is deteriorated because the calculation scale of the distance is different.

From the categorization stage, it is conceivable to use a distance calculation method having a high calculation cost such as principal component analysis. However, in the initial stage of categorization, the number of samples in a category is small, and the accuracy of calculation becomes poor. However, a large classification dictionary with a high identification rate cannot be created.

In the method described in Japanese Patent Application Laid-Open No. Hei 4-266188, an average is calculated using all samples and a large classification dictionary is created. There is no assurance that they will be classified into the same category, and there is also a problem that one character type is included in a plurality of categories, which may increase the size of the large classification dictionary.

[0010]

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and suppresses an adverse effect on the recognition rate due to the above-described distance calculation method or a difference in calculation accuracy, and has a high recognition rate. It is an object of the present invention to provide a large classification dictionary creation device capable of creating a large classification dictionary.

[0011]

According to the present invention, a primary large classification dictionary is created by a first distance calculation method using an average of each character type, and then a primary large classification dictionary is created by a second distance calculation method. Calculates the sum of the distances between the average of the categories within and the characteristics of each character type, and moves the character type between the categories so that the total sum of the distances in each category is minimized to create a secondary large classification dictionary . When creating a secondary large classification dictionary,
If the characteristics of the character type to be moved and the distance between the destination category are
In the first distance calculation method, the secondary large classification dictionary can be created by restricting the movement of the category only for the character type within the threshold value.

For example, an initial seed point is determined using a distance calculation method having a low calculation cost as a first distance calculation method, and a first-order large classification dictionary is constructed. The primary large classification dictionary can be reconfigured by using a distance calculation method having a high calculation cost actually used in the identification processing, and a secondary large classification dictionary actually used in the large classification processing can be created. Since the created secondary large classification dictionary is reconfigured by the distance calculation method actually used in the identification processing, the calculation scale of the distance can be matched in both the large classification processing and the detailed recognition processing. Rate can be improved. Also, by determining the initial seed point using, for example, a conventional Euclidean distance, the calculation cost can be reduced, and the accuracy can be prevented from being reduced when a distance calculation method having a high calculation cost is used from the beginning. It becomes possible.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a basic configuration diagram showing an embodiment of a large classification dictionary creating apparatus according to the present invention. In the figure, 1
1 is a character image data file, 12 is a feature amount extraction unit, 1
3 is a feature amount storage unit, 14 is a first large classification dictionary creation unit, 1
5 is a second large classification dictionary creation unit, 16 is a total distance calculation unit, 1
7 is a central control unit, and 18 is a storage device.

The character image data file 11 stores sample images classified for each character type as binary digital images. The feature amount extraction unit 12 extracts feature amounts from all sample images in the character image data file 11 and calculates an average thereof. The feature amount storage unit 13
The feature amount for each character type extracted by the feature amount extraction unit 12 is stored. The first large classification dictionary creating unit 14 creates a first large classification dictionary by a first distance calculation method. The second large classification dictionary creating unit 15 reconstructs the first large classification dictionary created by the first large classification dictionary creating unit 14 using the second distance calculation method, and creates a second large classification dictionary I do. The total distance calculation unit 16
Calculate the sum of the distance between the average of the category and the character type in the category. The central control device 17 controls the entire device.
The storage device 18 stores a program or the like for the central control device 17 to control the entire device.

Next, a first operation example of the embodiment of the large classification dictionary creating apparatus according to the present invention will be described using a specific example. The character image data file 11 stores a sample image of each character collected in advance as a binary image having a certain size. Here, as an example, it is assumed that each character type is set to a size of 64 × 64. A plurality of sample images are stored for each character type, and the number of sample images for each character type is stored in the storage device 17.
Is stored in At this time, the number of sample images in each character type does not need to be constant, but there are several tens of sample images composed of a plurality of fonts for one character type in consideration of improving the accuracy of recognition processing. It is desirable.

Next, the feature extracting unit 12 extracts features from all the character images in the character image data file 11, calculates an average value for each character type, and stores it in the feature storing unit 13 in correspondence with the character type code. Remember. Various features have been proposed so far. Here, as an example, description will be made using a primary peripheral feature.
The primary peripheral feature is characterized by scanning from each side of a circumscribed rectangle of a character image and measuring a distance from a white pixel (background) to a black pixel (character portion).
In addition, when the second point is also characterized by a point that changes from a white pixel to a black pixel, it is called a secondary peripheral characteristic.

FIG. 2 is an explanatory diagram of the primary peripheral features. In FIG. 2, the character image is represented by 8 stripes in the horizontal direction.
The area is divided into two areas, and the distance indicated by the broken arrow from the left side of the character to the character portion is obtained for each area. Similarly, the distance from the right side to the character portion can be obtained. Furthermore, it is divided into vertical stripes,
The distance from the upper side and the lower side to the character portion is obtained for each area. Although the number of divisions of the region is arbitrary, here, it is assumed that the region is divided into eight regions in both the horizontal and vertical directions, and 8 ×
Thus, a total of 32 feature vectors can be obtained.

The feature vector is extracted from each sample image of each character type, and the average vector is stored in the feature amount storage unit 13 as the characteristic amount of each character type in association with the character type code. FIG. 3 is an explanatory diagram of an example of a feature amount stored in the feature amount storage unit. In the example shown in FIG.
For example, for the character type “A”, the average vector (0.49, 0.88,...) Of the 32-dimensional feature vectors extracted from the plurality of sample images is stored in association with the characteristic amount of the character type. You.

Next, using the average feature amount of each character type stored in the feature amount storage unit 13, a first large classification dictionary creation unit 14 creates a primary large classification dictionary. In the first large classification dictionary creating unit 14, for example, by Duda, Harr,
“PATTERN CLASSIFICATION A
ND SCENE ANALYSIS ", 1973, W
Hierarchical clustering can be performed using the average feature amount of the character type using a method described in iley-Interscience.

FIG. 4 is a flowchart showing an example of the hierarchical clustering process. First, in S21,
Let m be the desired number of clusters, n be the total number of character types, and X = {c _i | i = 1,. . . , And n}, c _i is the representative feature vectors of character types that are similar is held. c
_As the initial value of _i , one representative feature vector of each character type is entered. In S22, the current number of clusters is compared with the desired number of clusters m, and if the current number of clusters is equal to m, the process ends with X at that time as a result of clustering. If not, the process proceeds to S23.
In S23, a set of two clusters in which the distance d of the cluster in the feature space is the smallest is found, and this is integrated into one cluster. Then, the process returns to S22.

The desired number m of clusters can be arbitrarily given. In this process, various methods of calculating the distance d between the clusters are conceivable. Here, the Euclidean distance is calculated by taking out the feature vectors in the two clusters one by one and forming a set. The distance between the closest vector set among them is defined as the distance between two clusters. In this example, a hierarchical clustering method is used, but a well-known K-mean method or the like is used.
Various other methods may be used to create the primary large classification dictionary.

As described above, the square error between the average of the categories and the samples in the categories is generally used as a scale indicating the goodness of unity of the categories created as described above. The smaller the square error is, the better the unit of samples in the category is. However, in such a hierarchical clustering process, it cannot be said that the square error is the minimum for each category. Therefore, the above-mentioned document (Duda, Harr, “PATTERN CL”
ASSIFICATION AND SCENE AN
ALYSIS ", 1973, Wiley-Inters
(Science Co., Ltd.) examines the amount of change in the square error when a sample in a category is moved to another category, and reconfigures the category so as to minimize the square error. Shows how to get categories.

In this method, since the square error within a category is generally calculated using the Euclidean distance, the recognition rate is reduced unlike the distance calculation method used in the actual recognition processing as described above. In order to avoid such a problem and create a large classification dictionary by a distance calculation method that further increases the recognition rate, the second large classification dictionary creation unit 15
And the total distance calculation unit 16 creates a secondary large classification dictionary.

FIG. 5 is a flow chart showing an example of the processing of the second large classification dictionary creation unit 15 in the first operation example in the embodiment of the large classification dictionary creation device of the present invention. Here, only the processing for each character type will be described. However, the processing shown in FIG. 5 is actually performed for the entire character type, and the processing is repeated until the movement of the character type between categories stops. The distance used in the following processing uses the distance obtained by the second distance calculation method.

First, the target character type is updated in S31, and the category to which the target character type belongs is stored in the variable J in S32. Next, in S33, the average feature amount of all the character types in the category is used to calculate the distance between the average of the category and all the character types in the category, and the sum is calculated. This process is performed by the total distance calculation unit 16. At this time, in the category to which the target character type belongs, the sum of the distance between the average of the category and all the character types in the category is calculated on the assumption that the target character type is excluded from the category. For a category to which the target character type does not belong, the case where the target character type is added to the category is assumed, and the sum of the distance between the average of the categories and all the character types in the category is calculated. The sum thus calculated is defined as A _i .

In S34, the total sum A _i calculated in S33 is compared, and the cluster S having the smallest total distance in the category is selected. In S35, the selected cluster S
Is determined to be a cluster to which the target character type belongs. If the cluster S is not a cluster to which the target character type belongs, the process moves to the cluster S in which the target character type is selected in S36, and the average of all categories is updated in S37. I do.

By performing such processing for the entire character type, a secondary large classification dictionary in which the category is optimized based on the distance calculated by using the second distance calculation method is constructed. The second-order large separation dictionary configured as described above constitutes a large classification dictionary having a high recognition rate in the recognition processing by using the distance calculation method actually used in the recognition processing as the second distance calculation method. Can be.

Next, a second operation example in the embodiment of the large classification dictionary creating apparatus of the present invention will be described. In the second operation example, a case where the projection distance method is used as the second distance calculation method will be described.

The projection distance is defined by a residual between a feature vector of a character type and a subspace composed of higher-order principal component axes calculated from the distribution of the character type. FIG. 6 is an explanatory diagram of the projection distance. In the figure, 41 is a subspace, 42 is a feature vector, 43 is an average vector, and 44 is a residual. here,
A subspace composed of higher-order principal component axes calculated from the distribution of character types is referred to as a subspace 41. The average vector in that category is indicated by an average vector 43. Now, when a character type feature vector 42 is given, a residual 44 which is the shortest distance from the point C to the subspace 41 is obtained.
Is the projection distance.

The Euclidean distance in the example shown in FIG. 6 is the length of the vector MC, and is a value larger than the residual 44 which is the projection distance. For this reason, a category that is determined to be far from the Euclidean distance may be determined to be located very close to the projection distance.

FIG. 7 is an explanatory diagram of the distribution of two categories according to the Euclidean distance, and FIG. 8 is an explanatory diagram of the occurrence of erroneous recognition due to the projection distance. In the figure, 51 and 52 are categories, 5
Numerals 3 and 54 are subspaces, and 55 is an erroneous recognition location. For example, as shown in FIG. 7, consider a case where there are two categories 51 and 52 classified by Euclidean distance using a two-dimensional feature amount. In this case, the subspace used for the projection distance is one-dimensional, and the subspaces 53 and 54 for the categories 51 and 52 are each represented as a straight line. When these two categories are reconstructed by the projection distance, a part of the category 52 is classified into the category 51 at the intersection of the two subspaces 53 and 54 as shown in FIG. Is generated.

That is, the distribution of the character types belonging to the category 52 that exists near the subspace 53 of the category 51 has a very small residual from the subspace 53 of the category 51, and the distribution of the character types belonging to the category 52 is very small. Some will be classified into category 51. Since the actual distribution of character types has a certain degree of spread due to the variation in the feature amount, a character type located near the misidentification occurrence position 55 may cause misidentification.

In order to create a category that is more stable with respect to fluctuations in the feature amount and reduce the above-described misidentification, movement of character types between categories when reconstructing the secondary large classification dictionary is limited. Specifically, a threshold value is provided for the distance obtained by the first distance calculation method, and only for the character type whose distance between the feature amount of the character type and the average of the category by the first distance calculation method is within this threshold value, the secondary It can be configured to move between character category categories when reconstructing the large classification dictionary.

FIG. 9 is a flowchart showing an example of the processing of the second large dictionary creation unit 15 in the second operation example in the embodiment of the large dictionary creation device of the present invention. Here, as in the above-described first operation example, only the processing for each character type will be described. However, in practice, the processing shown in FIG. 9 is performed on the entire character type until the character type does not move between categories. Repeat the process.

First, in S61, the target character type is updated, and in S62, the category to which the character type belongs is stored in the variable J. Next, in S63, the distance between the feature vector of the target character type and the average of all categories is set to the first distance.
In step S64, a category whose distance by the first distance calculation method is within a threshold value is selected. In S65, for the category selected in S64, the process of creating the secondary large classification dictionary in the above-described first operation example shown in FIG. 5 is performed using the second distance calculation method.

By such processing, some or all of the character types of the misrecognition occurrence position 55 in FIG. 8 do not move to the category 51 even when reconstructing the secondary large classification dictionary so that they belong to the category 52. A large classification dictionary is created. This makes it possible to create a large classification dictionary that is resistant to variations in the feature amount. Of course, similar to the first operation example described above,
By using the distance calculation method actually used in the recognition processing as the second distance calculation method, a large classification dictionary having a high recognition rate in the recognition processing can be configured.

In the recognition process using the secondary large classification dictionary configured as described above, first, the distance between the character type and the category is calculated by the first distance calculation method, and the distance between the character type and the category is within the threshold value. On the other hand, further recognition is performed by the second distance calculation method.

[0038]

As is apparent from the above description, according to the present invention, a primary large classification dictionary is created by the first distance calculation method using the average of character types, and each primary classification dictionary in the primary large classification dictionary is created. The categories are checked, and the sum of the distances between the average of the categories and the average of the features of each character type is calculated by the second distance calculation method. By creating a secondary large classification dictionary, a large classification dictionary having the same calculation scale as the distance calculation method actually used in the identification processing can be obtained, and a large classification dictionary with a high final recognition rate can be created. There is an effect that can be.

[Brief description of the drawings]

FIG. 1 is a basic configuration diagram showing an embodiment of a large classification dictionary creation device of the present invention.

FIG. 2 is an explanatory diagram of a primary peripheral feature.

FIG. 3 is an explanatory diagram of an example of a feature amount stored in a feature amount storage unit.

FIG. 4 is a flowchart illustrating an example of a hierarchical clustering process.

FIG. 5 is a flowchart illustrating an example of a process of a second large classification dictionary creation unit in a first operation example in the embodiment of the large classification dictionary creation device of the present invention.

FIG. 6 is an explanatory diagram of a projection distance.

FIG. 7 is an explanatory diagram of distribution of two categories according to Euclidean distance.

FIG. 8 is an explanatory diagram of occurrence of erroneous recognition due to a projection distance.

FIG. 9 is a flowchart illustrating an example of processing of a second large classification dictionary creation unit in a second operation example according to the embodiment of the large classification dictionary creation device of the present invention.

[Explanation of symbols]

11: character image data file, 12: feature amount extraction unit,
13: feature amount storage unit; 14: first large classification dictionary creation unit;
15: second large classification dictionary creation unit, 16: total distance calculation unit,
17 central control device, 18 storage device, 41 partial space, 42 feature vector, 43 average vector, 44
Residual, 51, 52 ... category, 53, 54 ... subspace,
55: Misidentification occurrence location.

Claims (2)

[Claims] 1. A feature extracting means for extracting a multidimensional feature amount from a character image, and a first distance calculation method using the average of the features of each character type to classify the character image into a category composed of a plurality of character types. A first-class classification dictionary
A second major classification dictionary creating means and a total calculating means for calculating the sum of the distances between the average of the categories in the first major classification dictionary created by the first major classification dictionary creating means and the character types in the categories by the second distance calculating method. A distance calculating means, and a second category in each category in the large classification dictionary created by the primary large classification dictionary creating means.
A secondary classification dictionary creating means for creating a secondary classification dictionary by moving each character type between categories so as to minimize the sum of the distances according to the distance calculation method. Dictionary creation device. 2. The method according to claim 1, wherein the second large classification dictionary creating unit determines a category of the character type whose distance between the characteristic of the character type to be moved and the average of the destination category is within a threshold value in the first distance calculation method. 2. The apparatus for creating a large classification dictionary according to claim 1, wherein the dictionary is moved.