JPH11203415A - Device and method for preparing similar pattern category discrimination dictionary - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prepare a discrimination dictionary capable of attaining both of high accuracy and a low calculation cost by the small number of dimensions of feature values. SOLUTION: It is supposed that a learning sample for patterns necessary for the preparation of a discrimination dictionary is previously prepared, and its pattern information is inputted to a similar pattern category preparing means 1. The means 1 finds out similar patterns and merges these patterns. A category division part 2 divides the prepared similar pattern category into plural categories reducing misrecognition. A reintegrated category preparing means 3 integrates the divided categories to a similar gategory again. Consequently a hierarchical dictionary consisting of the divided categories and the reintegrated category is prepared, and when the hierarchical dictionary is used for the recognition of the reintegrated category and the recognition of the divided categories in order in similar pattern category discrimination processing, high recognition accuracy can be obtained and calculation volume for recognition can be reduced.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus and method for creating a similar pattern category identification dictionary, and more particularly, to previously classifying patterns or characters having similar shapes in terms of image characteristics into the same category, and finding unknown patterns. The present invention relates to a similar pattern category identification dictionary creating apparatus and method for identifying whether a feature extracted from a pattern or a character image belongs to one of these similar pattern categories or similar character categories.

[0002]

2. Description of the Related Art A document filing apparatus has been put into practical use which enables a hard copy document to be converted into an image by an image scanner, stored electronically, and retrieved later.
However, in many cases, attributes for searching such as keywords have to be manually added to each input image, which requires much labor. Originally, a full text search based on text content is desirable in document search. However, this is possible for an electronic document created by DTP (DeskTop Publishing) or the like, but cannot be performed directly for a document image. For this reason, Japanese Unexamined Patent Publication No. Sho 62-44878 discloses that a text portion in a document is subjected to character recognition, and a full-text search can be performed using coded text contents. However, in character recognition, particularly in Japanese having many character types, generally, a hundred-dimensional feature amount vector is obtained and 3
In order to perform feature amount matching with 000 character types or more, very large computer power is required for feature vector matching processing. Further, since the character recognition rate is not high, there is a problem that a keyword to be searched may be erroneously recognized.

In the electronic filing system described in Japanese Patent Application Laid-Open No. 62-44878, each character candidate obtained during the character recognition processing is held to reduce omission of retrieval due to erroneous recognition. However, basically, a large amount of computer power is required at the time of document registration in order to perform character recognition processing. If it is assumed that the final image to be obtained is a document image including a word specified at the time of search, character recognition is not performed. The results are mostly useless.

In the literature (Tanaka et al., "Implementation of a character string search function for Japanese document images", Information Processing Society of Japan, Information Technology Research Institute, Material 19-1, January 1995), the feature amount obtained from each character image is described. Instead of taking out and character recognition, the feature amount is directly converted into a 36-bit code.
Next, a character string search is realized by matching a feature amount with a search keyword image. However, it is necessary to input a search keyword as an image or to generate an image using a character font image, which is disadvantageous in that it is not easily affected by font variations.

[0005] Another document (Reynar, J. et al, "Do
cument Reconstruction: A Thousand Words from One P
icture ”, in Proc. of 4th Annual Symposium on Docu
mentAnalysis and Information Retrieval, Las Vegas,
April 1995) discloses an attempt to classify characters in a text image of a European language (English) into a small number of categories according to their size and position, and to identify them as words according to their arrangement. However, it is difficult to intuitively set features that are clues for languages including many character types such as Japanese and Chinese. Also, unlike European languages, since there is no space between words, it cannot be obtained directly from the image in word units. For this reason, it was difficult to directly identify text such as Japanese using words using the disclosed method.

The applicant has already disclosed the following method in Japanese Patent Application No. 8-274732 in order to solve the above problems. The technique is as follows. First, character types having similar shapes (for example, “way” and “tsu”, number “0”, and Roman alphabet “O”) are grouped as one category in advance. When analyzing actual images, each character image is identified by these similar character categories, and only those that can be determined as words by morphological analysis, which is a technique for extracting words from Japanese text, are extracted from the similar character category string Only the characters that remain ambiguous are identified in detail. The advantage is that the number of feature dimensions that are used for large classification in character recognition is small, and matching with a few similar character categories is sufficient, so that the calculation cost can be greatly reduced, and when morphological analysis is used Since the detailed identification is performed only on the ambiguity among words that can be accepted as words, the calculation cost of the feature amount matching in the detailed identification can be reduced. According to the present invention, clustering based on representative vectors of character types in a feature space, that is, by integrating vectors having a small distance between representative vectors (average vectors) of character types, the cluster center is regarded as a representative vector. A character category was formed, and unknown character samples were identified by the shortest distance identification with their representative vectors. However, the method of identifying a similar character category in the disclosed technique has not always been highly accurate. The distribution of the actual character feature vector will spread in the feature space as the similar character category is formed, and in the case of a sample that is far from the representative vector and located at the end of the distribution, This is considered to be due to an increase in misidentification to another similar character category. This situation will be described with reference to FIG.

FIG. 16 is a diagram for explaining a problem in similar character category identification. FIG. 16 schematically shows a feature amount space set two-dimensionally for the sake of simplicity. A sample of one similar character category has a distribution in the feature amount space. Here, three distributions A, B, and C are shown as examples, and each distribution A, B, and C is approximated by, for example, an ellipse. I have. It is assumed that the representative vector of each distribution A, B, C is at its average value, that is, at the center of the ellipse here. In the shortest distance identification, a vertical bisector connecting these representative vectors becomes an identification boundary, and a sample distributed beyond the identification boundary causes erroneous recognition. For example,
In the distribution A of a certain similar character category, both ends of the major axis of the ellipse representing the distribution are beyond the discrimination boundary with the distribution of another similar character category. It will be erroneously identified as a similar character category.

References (Ito, Endo et al., "Dictionary Construction Method for Registering Character Types in Multiple Clusters in Hierarchical Printing Kanji System", IEICE Transactions D-II, Vol.J78-D-II, No.6 ,
pp. 896-905, June 1995), as described above, in order to reduce the calculation cost of character recognition, in order to reduce the category by performing clustering on character types that are close to each other in the feature space, that is, similar in shape. A method has been proposed in which these are created in a hierarchical manner, and similar character candidates are sequentially narrowed down. Here, too, the problem of misidentification due to the spread of character sample distribution is pointed out, and in order to solve this problem, the possibility of erroneous identification occurring using learning character samples in the middle of the hierarchy is considered. Investigating, and if there is a possibility, the character type is registered in a duplicate category. However, this method is based on the premise that the correct character type can be identified with high accuracy in the final stage of collation. In the classification into the similar pattern category of the present application, even if the matching is performed with the brute force against the representative vector of the similar character category as described above, the accuracy is not so good, so that the premise cannot be used. At this time,
Although it is possible to identify a similar character category using another feature quantity having a large number of dimensions in the final stage, a new calculation cost is required for matching and extracting feature quantities for the increased number of dimensions. Statistical distance calculation such as Mahalanobis distance may be performed without using the Euclidean distance, but this also increases calculation cost. As another method for improving the accuracy, it is conceivable to allow a character type that is easily misidentified to be registered in a plurality of categories at the final stage. However, in this case, a problem occurs in word extraction in a later stage. For example, if the character type “a” is a similar character category “A”,
When the character type "b" is registered in the category "C" or "D" while being registered in duplicate with "B", the similar character category strings AC, AD, BC, and BD may all be the word "ab". Therefore, when morphological analysis is performed, the number of entries in the word dictionary increases significantly. In addition, in the case of using a bi-gram that registers all two character sequences as an index, one search word must be searched for each of a plurality of similar character category strings at the time of search. Cause an increase. Therefore, it is desirable that a similar character category to which a certain character type belongs is single.

[0009] In addition, there is disclosed a method for identifying a category including some similar characters. For example, Japanese Patent Application Laid-Open No. 63-263590 discloses a method in which character samples that are hierarchically similar are grouped and identified stepwise. In this method, a plurality of candidates are selected in order to cope with misidentification between groups of similar characters, and all the lower groups are identified. However, in this method, similar character groups are formed not in units of character types but in units of character samples, so that the same character types may eventually belong to different classifications. JP-A-4-337888, JP-A-5-174193
In the official gazette, the character type is narrowed down hierarchically using a binary tree and a three-way tree. However, in any case, the same character type may belong to a different classification, and the relief is ultimately required. It is assumed that detailed identification is performed using a feature having a large number of dimensions. For this reason, the calculation cost in the subsequent stage is increased as in the above-mentioned reason.

Therefore, when an unknown character is identified in a similar character category in which character types do not overlap, a method is required that ensures identification accuracy and has a low calculation cost.

[0011]

In any of the conventional methods, there is a problem in the method of identifying a similar character category in terms of identification accuracy and calculation cost.

SUMMARY OF THE INVENTION The present invention has been made in view of the above points. In order to uniquely identify a similar character category which does not allow a predetermined character type to be duplicated, the number of feature amount dimensions is small. An object of the present invention is to provide a similar pattern category identification dictionary creating apparatus and method for creating an identification dictionary that enables both high accuracy and low calculation cost to be achieved.

[0013]

According to the present invention, in order to solve the above-mentioned problem, a similar pattern category for creating a similar pattern category identification dictionary used for matching when a pattern included in image information is identified as a similar pattern category. In the identification dictionary creating apparatus, similar pattern category creating means for obtaining similarity between patterns from pattern information obtained by extracting feature amounts of patterns included in an image, and creating similar pattern categories by grouping similar patterns together; A category dividing means for examining a situation of misrecognition of a learning sample of a pattern belonging to the similar pattern category, dividing the similar pattern category based on the situation and creating a detailed classification identification dictionary composed of divided categories, Find a representative vector and re-similar the division category Re-integrated category creating means for creating a large classification identification dictionary composed of the re-integrated re-integrated categories, and constructing a similar pattern category identification dictionary having a hierarchical structure together with the detailed classification identification dictionary. A feature-based similar pattern category identification dictionary creating apparatus is provided.

According to such a similar pattern category identification dictionary creating apparatus, pattern information in which feature amounts of patterns included in the pattern are extracted from an image in advance is input. A similar pattern category is created by obtaining patterns and putting them together. The created similar pattern category is subdivided into categories in which misrecognition is reduced by the category dividing means. The categories thus divided are integrated again into similar categories by the reintegrated category creating means. As a result, a hierarchical dictionary of the divided categories and the reintegration categories is created. By obtaining an integrated category and obtaining the shortest-distance divided category from among the divided categories belonging to the reintegrated category, high recognition accuracy is obtained and the amount of calculation for recognition is significantly reduced.

Further, according to the present invention, in a similar pattern category identification dictionary creating method for creating a similar pattern category identification dictionary used for matching when a pattern included in image information is identified as a similar pattern category, image information is input. Extracting a feature amount of a pattern included in the image information, accumulating a learning sample, examining a similarity between the patterns of the learning sample, creating a similar pattern category in which similar patterns are put together, Based on a learning sample of a pattern belonging to the similar pattern category, each similar pattern category is divided to create a divided category so as to reduce erroneous identification, and the divided categories are collected again into similar categories, and To create reintegration categories that together form a hierarchical identification dictionary. Similar pattern category identification dictionary creating method consists of is provided.

According to such a similar pattern category identification dictionary creating method, first, a learning sample necessary for creating a dictionary is extracted from image information, and the sample is combined into a category having a similar pattern, and a similar pattern category is created. create. Next, based on the learning sample, each similar pattern category is divided so as to reduce erroneous recognition, and a dictionary of detailed classification including the divided categories is created. And
The divided categories are again grouped into a small number of similar categories, and a large classification dictionary including reintegrated categories is created. As a result, a hierarchical identification dictionary is constructed, and at the time of the identification processing, it is finally possible to uniquely identify a similar pattern category having no overlapping pattern by the constructed hierarchical identification dictionary.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram illustrating the principle of the present invention. The similar pattern category identification dictionary creation device according to the present invention includes:
Similar pattern category creating means 1 for receiving pattern information obtained by extracting feature amounts of patterns included in an image, category dividing means 2 for receiving the output of the similar pattern category creating means 1, and receiving the output of the category dividing means 2 And a reintegrated category creating means 3 for outputting a similar pattern category identification dictionary.

The similar pattern category creating means 1 examines the similarity between patterns from input pattern information and creates similar character categories by grouping similar patterns. The category dividing means 2 divides each similar pattern category based on a learning sample of a pattern belonging to the category created by the similar pattern category creating means 1 so as to reduce erroneous identification. Is subjected to principal component analysis to find an end point vector on the principal component axis. If the end point vector is erroneously identified as another category, the category is clustered and divided until no erroneous identification is found. At this time, among the end point vectors on the principal component axis that are erroneously identified as other categories, the end point vectors having the larger eigenvalue are divided. The reintegrated category creating means 3 is for re-assembling the divided categories into similar categories and constructing a hierarchical identification dictionary. The representative vectors of the categories divided by the category dividing means 2 are clustered. When a representative vector is obtained and a learning sample is erroneously identified by comparison with the obtained representative vector, an identification dictionary is constructed by registering a divided category to which the learning sample belongs to an erroneously identified category. Go.

The similar pattern category identification dictionary created in this manner is used to extract, for example, characters or word information from a document image input from an image scanner or the like. It is used for processing to identify similar character types into similar character categories.

Next, an example in which the similar pattern category identification dictionary creating apparatus is applied to create a similar character category identification dictionary for character recognition will be described. FIG. 2 is a diagram showing a hardware configuration for implementing the present invention. An apparatus embodying the present invention includes a general personal computer 10 and its peripheral devices. Personal computer 1
0 is a central processing unit (CPU) 11 and a main memory 12
And a peripheral device controller 13,
The peripheral device controller 13 includes the external storage device 1
4, a display 15, a keyboard 16, a pointing device 17 such as a mouse, an image scanner 18 as an image input device, and a network 19 are connected.

The processing of the apparatus of the present invention is entirely constituted by software, stored in the external storage device 14, loaded into the main storage memory 12 as needed, and executed by the CPU as needed.

FIG. 3 is a flowchart showing the flow of processing for creating a similar character category identification dictionary. When creating the similar character category identification dictionary, first, a document placed on the platen of the image scanner 18 is read as a binary image, and the binary image is cut out for each character, and preprocessing such as size normalization is performed. After the application, an image input process of storing the image data in the external storage device 14 together with the corresponding character type is performed (step S1). Next, for each character image for learning stored in the external storage device 14, a feature value extraction process is performed in which the feature value is represented by a multidimensional vector, and the feature value is stored in the external storage device 14 (step S2). ).
Next, an average vector of the learning sample of each character type is calculated and set as a representative vector of the character type, and the obtained representative vector is subjected to a clustering process to create a similar character category (step S3). Next, based on the distribution of training samples for each created similar character category,
The category is divided so that misrecognition to other categories is reduced, and the representative vector of the divided similar character category is numbered together with the character representative of the category and stored in the external storage device 14 (step S4). ). Then, the representative vectors of the divided similar character categories are clustered again, integrated into a small number of categories, and the representative vectors and the numbers of the divided similar character categories belonging to each are registered to create a reintegrated category. Processing is performed (step S5). Hereinafter, the dictionary creation processing will be described in more detail along the processing flow.

FIG. 4 is a flowchart showing the flow of the image input processing. First, a document for learning is read from the image scanner 18 (step S11). It is desirable to change the typeface and size of the manuscript in order to obtain a wide variety of samples. The binarization may be performed simply by using a fixed threshold, or may be captured as a grayscale multi-valued image, and may be binarized by threshold processing. In the latter case, with some binarization thresholds,
It is effective to increase the number of samples because an image with crushing and blurring can be collected. In the present embodiment, different sample images are created by changing some threshold values (step S12). Next, an area is cut out for each character from the obtained binary image (step S1).
3). The cut-out character image is subjected to removal of small isolated point noise that is considered not to constitute a character from the character image (step S14), and normalization of the size is performed based on a circumscribed rectangle of the character (step S15). . Here, it is assumed that one character is normalized to an image size of 64 × 64 pixels. There are several known techniques for noise removal, size normalization, and the like, and an appropriate one may be used. These preprocessed images are given the corresponding character types and stored in the external storage device 14 (step S16). Steps S14 to S16 are repeated for each cut out character image, and steps S13 to S1 are performed.
6 is repeatedly processed for each different threshold value.

In the following feature value extraction processing, in this embodiment, peripheral features having a small number of dimensions are used. This peripheral feature will be described with reference to FIG. FIG. 5 is an explanatory diagram of peripheral features. Peripheral features are characterized by the number of pixels from each side of the circumscribed rectangle to the point where a black pixel first appears (primary peripheral) and the point where the pixel once becomes a white pixel and then becomes a black pixel again (secondary peripheral). Things. The primary peripheral and the secondary peripheral are examined for each of 64 pixels on each side, and eight pixels are averaged, and the average is set as each element of the feature amount vector. Therefore, assuming that up to the secondary peripheral is taken in eight dimensions on each side, a total of 64-dimensional feature amount vectors are obtained.
This feature amount vector is stored in the external storage device 14 in correspondence with the character type. By this processing, the feature amount vectors are calculated for all the prepared character images for learning the typeface and size.

Upon completion of the feature value extraction process, next, a similar character category generation process is started. In the similar character category generation processing, first, from the external storage device 14,
All learning samples are extracted for each character type, their average vector is calculated, and this is used as a representative vector of the character type. Once this has been obtained for all character types, they are clustered. Clustering is described in the literature (Du
da, Hart, Pattern Classification and Scene Analy
sis ", published by Wiley-Interscience). In this method, first, hierarchical clustering is performed, and the center of each cluster, the feature vector of each learning sample, and The optimization is performed so that the sum of the square errors of the is minimized.

Hierarchical clustering actually consists of the following steps. (1) the desired number of clusters m, the total number of character type n, the initial cluster _{C = {c 1, c 2} , c 3,. . . _, c _n } and c _i
Are representative feature vectors of similar character types. As the initial value, one representative feature vector of each character type can be entered.

(2) If the current number of clusters is equal to m, the processing ends at the time point C as a result of clustering. If not, proceed to the next step. (3) Find a set of two clusters with the smallest distance d between clusters in the feature space, integrate them into one cluster, and return to (2).

Here, the desired number m of clusters is arbitrarily given. In this processing, various methods for calculating the distance d between the clusters are conceivable. Here, the distance between the centers of the two clusters, which is called the centroid method, is set as the distance d between the two clusters.

Since the result of this hierarchical clustering cannot be said to be optimal clustering, cluster optimization is performed starting from this. The optimization takes the sum of squares of the distance between the average value of the feature vectors in each cluster and each feature vector, and uses the sum of all clusters as an evaluation function. As the value of the evaluation function is smaller, the feature vectors in the cluster are denser, and it can be said that the clustering is better. It is generally difficult to find clustering that minimizes this, but it is possible to perform optimization in a pseudo manner. This is done in the following steps.

(1) Extract an arbitrary feature vector x. (2) If the cluster to which x currently belongs is c _i and the feature vector registered there is only x, the process returns to (1). Otherwise, all clusters c
_The following calculation is performed for _j .

That is, when j ≠ i,

[0032]

(Equation 1)

Is calculated, and when j = i,

[0034]

(Equation 2)

Is calculated. Where n_jIs c_jRegister with
The number of vectors_jIs c _jFeatures belonging to
It is the average of Khutor. The above equation gives x as c_jMoved to
The amount of change of the judgment function at the time is shown.

(3) If j at which a is a minimum is other than i, move x to cluster c _j and return to (4). (4) Repeat from (2) with the next feature vector as x. If the cluster cannot be moved for all the feature vectors, the processing at the time is ended with the cluster at that time as a result.

In this way, clustering of similar characters is performed. In this operation, the same processing is performed by variously changing the method of extracting an arbitrary character in (1), and an evaluation function (sum value of the sum of squares of the distance between each feature vector and the average value of the feature vectors in each cluster) is performed. Is adopted as the result.

Each cluster is stored as a similar character category table, and is used when registering a document.
Here, an example of the stored table is shown in FIG. 6A and 6B are diagrams illustrating an example of a table created by the similar character category generation process, in which FIG. 6A illustrates an example of a similar character category table, and FIG. 6B illustrates an example of a character code / category correspondence table. I have. Similar character category table
As shown in part (A), for each category,
The storage device 1 includes a character code (similar character) of a character to which the character belongs, a representative vector (representative vector) of a category feature, and a character code (representative character) representing a category.
4 The category representative vector is an average vector of the feature vectors of the characters to which the category belongs. An arbitrary one of the character codes of the characters belonging to the category is assigned to the character code representing the category. In the search process, in order to convert a search keyword into a similar character category string, a character code and a representative character code of a corresponding category are set as a reverse lookup table of the similar character category table as shown in FIG. The created character code / category correspondence table is created at the same time.

After the similar character category has been created, a category division process is started. This processing assumes the distribution of the characters included in the similar character category from the learning sample of the character type included in the similar character category, and determines the most significant point from the end point of the distribution in the direction in which the distribution of the distribution is large, that is, the representative vector that can appear statistically. Assuming points that are far apart from each other, if the point is closer to the representative vector of another category than the representative vector of the category, the category is divided.

The division of the categories is basically performed according to the literature (Omachi, Son, et al., “Construction method of multi-template dictionary for character recognition considering distribution between categories”, IEICE Transactions D-II, Vol. J79. -D-II, No. 9, pp. 1525-1533). This aims to improve the accuracy of character recognition, and there is an increase in computational cost by dividing the learning sample of each character type into several clusters and giving multiple representative vectors to the same character type. However, this is a method for reducing false positives. FIG. 7 shows the flow of the actual processing.

FIG. 7 is a flowchart showing the flow of the category dividing process. First, as preprocessing to this process, one similar character category is focused on, principal component analysis is performed on a learning sample of a character type belonging to the similar character category, and
The eigenvalues and eigenvectors corresponding to the principal components of the order are stored (step S21). Eigenvalues and eigenvectors are obtained only from the smaller of the number of dimensions of the feature vector or the number of learning samples minus one. Here, it is assumed that more learning samples are obtained than the number of dimensions. l is a constant indicating how many eigenvalues are examined in descending order, and is given when a dictionary is created. Usually, it is considered that misidentification occurs on an axis with large variance, so that about l = 5 is sufficient.

After the pre-processing is completed, two sample endpoints which can appear farthest from the center in each of the l principal component axis directions in the feature amount space are assumed for each similar character category (step S22). . The end point vector p is obtained by the following equation.

[0043]

(Equation 3)

Where m is a representative vector of a category, a
Is a constant, λ _i and Φ _i are the i-th eigenvalue and eigenvector. a is a constant and a positive real number. That is, this p
Means the position in the feature space of the sample farthest from the representative vector on the principal component axis. The meaning of this expression is
FIG. 8 schematically shows a two-dimensional feature space.

FIG. 8 is a diagram for explaining the meaning of the expression of the end point vector in the feature quantity space. In FIG. 8, small black circles represent samples in the similar character category, and their distribution is approximated by an ellipse by statistical processing called principal component analysis. The center of this ellipse is the representative vector m, which is the average of the distribution, the arrow on the long axis in the direction of the first principal component is the eigenvector Φ ₁ , and the arrow on the short axis in the direction of the second principal component is the eigenvector Φ ₂ . In this case, both end points of the long axis indicated by x are end point vectors p expressed by Expression (3). It is known that the principal component axes obtained by the principal component analysis are orthogonal to each other in the feature space and have no covariance. Therefore, each principal component axis can be treated statistically independently. The constant a indicates how many times the standard deviation of the principal component axis is considered as the range of the distribution. Assuming a normal distribution, a
If = 3.5, 99.96% of the distribution will be included in this.

When the end point vectors p are obtained on the principal component axes, the representative vectors of the categories located at the shortest distance are extracted for each of them (step S2).
3). If the representative vector at the shortest distance is the category of interest at present or a representative vector of the same category that has already been divided, nothing is performed (steps S24 and S25). Otherwise, it is stored as a division candidate along with its principal component order, corresponding eigenvalue, and eigenvector as a category in which misidentification occurs (step S26). Of the inspections in steps S21 to S26 performed for all similar character categories, a division candidate having the largest eigenvalue is extracted and divided (step S28). The division is performed by clustering learning samples of all character types belonging to the similar character category. The clustering here is
The k-means method using the Euclidean distance is used. The reason for using the k-means method is to control so as to divide the sample along the principal component axis where misidentification occurs.
Therefore, a learning sample closest to the two end point vectors p of the similar character category to be divided is given as the initial cluster center to be given. This is a document (Omachi, Son,
"Construction method of multi-template dictionary for character recognition considering distribution between categories", IEICE Transactions D-II,
Unlike Vol.J79-D-II, No.9, pp.1525-1533), the region including the end point can be divided more reliably. In addition,
If the similar character category is already one of several parts, it is not only the target cluster,
Perform clustering for all original similar character categories. At the center of the initial cluster at that time, the category to be divided gives a learning sample with the shortest distance to the end point vector, and the representative vector otherwise. FIG. 9 schematically shows an example.

FIGS. 9A and 9B are diagrams for explaining the category division by the k-means method. FIG. 9A shows an example of the first division of the similar character category, and FIG. 9B shows a further example of the division of the similar character category. Is shown. When dividing a similar character category by the k-means method first, the sample closest to the end point of the long axis is set as the initial cluster center. This sample is indicated by a white circle in the left-hand diagram before the division shown in FIG. Using this sample as the initial cluster center,
The clusters finally divided by using the clustering method of the averaging method are shown on the right side. This means that the cluster whose end point of the long axis has the longest distance and has a high possibility of erroneous recognition is divided into clusters each having a short long axis. Then, by performing principal component analysis on each cluster, a new representative vector,
An eigenvector and an end point vector are obtained. (B)
In the case where one of the similar character categories already divided into two is to be divided, as in (A), the white circle on the left indicates the initial cluster center, and the result is on the right.

When the division by the clustering is completed in this way, a representative vector of each of the divided clusters is obtained, a principal component analysis is performed, and the eigenvalues and eigenvectors of the top l principal components are obtained. Remember. This is repeated for all of the l principal components for all of the divided categories, and continues until no more division candidates can be obtained (step S27). Upon completion, an identification dictionary is created by combining a representative vector of the divided similar character category (hereinafter, referred to as a divided category) and a similar character representative character code, and registered in an external storage device.

FIG. 10 is a diagram showing a data structure of an identification dictionary by category division. This identification dictionary includes a category number, a similar character representative character code, and a representative vector. Here, the category number is used to identify a divided category belonging to a reintegrated category created in a reintegrated category creation process described later.

A category with a small number of samples may be generated during the division. In such a case, when the principal component analysis is performed, an error may increase.
For this reason, a division category including only a certain number of samples or less is not set as a division target. In this way, meaningless division can be prevented.
In the present embodiment, the threshold value of the number of samples is set so that the order of the principal component analysis does not fall below the feature dimension of 64 dimensions.
65 is set.

Using the identification dictionary based on the divided categories obtained as described above, the shortest distance identification is performed for the identification of unknown characters, thereby reducing erroneous recognition. However, as a matter of course, the number of representative vectors to be collated increases due to the category division, and may be larger than the original character type. Originally, there is no point in reducing the calculation cost by identifying the similar character categories in a number smaller than the number of character types. To solve this,
The re-clustering of the divided similar character categories is performed, and they are integrated again into fewer categories. This is the reintegration category creation processing. At the time of actual identification, first, a comparison is performed with a small number of categories obtained by this re-clustering, the best one is extracted, and a comparison is performed with a representative vector of a divided cluster belonging thereto, thereby reducing the amount of calculation. . In the reintegration category creation process, the centroid method and the optimization method in which the representative vector of the character type is first clustered to construct a similar character category are used as they are. here,
It does not matter whether the same similar character category was originally divided or not, and the one obtained as a result of the division is treated without distinction. The cluster obtained by this processing is called a reintegration category.

With the shortest distance discrimination with the reintegrated category obtained by this process alone, erroneous recognition occurs as in the case of first identifying the shortest distance with a representative vector of a similar character category. In order to avoid this, a learning sample is used for each of the divided categories, and if an erroneous recognition occurs, the learning sample is registered in the reintegrated category where the erroneous recognition has occurred. FIG. 11 shows this processing flow.

FIG. 11 is a flowchart showing the flow of the reintegration category creation process. In this reintegration category creation process, first, for each of the divided categories obtained by the category division process, all the belonging learning samples are extracted (step S31), one of which is extracted, and
A representative vector of the reintegrated category is obtained (step S3)
2). If the category is the reintegrated category to which the current divided category belongs, the next sample is checked without doing anything (step S33). If it is a different reintegration category, the currently focused division category is registered in the reintegration category (step S34). That is, the divided category of interest belongs to a plurality of reintegrated categories. This operation is performed for all the samples of all the divided categories, and when the operation is completed, the reintegrated category is stored in the external storage device as a large classification dictionary by combining the representative vector and the number of the divided category belonging thereto. (Step S35). FIG. 12 shows the data structure.

FIG. 12 is a diagram showing a data structure of a large classification dictionary by the reintegrated category creation processing. The data of the large classification dictionary created by the reintegration is composed of the representative vector and the number of the divided category belonging to it, and from this example, the divided category registered overlapping with the divided category number Is found to exist. In addition,
Here, registration is performed only for the divided category located at the shortest distance, but the divided category of interest is registered for all reintegrated categories located at a distance obtained by adding a value to the shortest distance viewed from the sample. By doing so, it is possible to cope with fluctuations in unknown characters.

As described above, the representative vector of the reintegrated category obtained as a result of the reintegrated category creation processing is used as a large classification dictionary, and the representative vector of the divided cluster is used as a detailed classification dictionary. Perform collation of. FIG. 13 shows a flowchart of the pseudo character recognition process for performing the identification into the similar character category.

FIG. 13 is a flowchart showing the flow of the similar character category identification process. First, a document image (binary image) is input from the image scanner (step S4).
1). A character block is extracted from the input image, and each character is cut out (step S42). Character block extraction processing is described in the literature (Akiyama, Masuda, "Segmentation of Document Image Using Peripheral Distribution, Line Density, and Bounding Rectangular Features" IEICE Transactions D-II, Vol.J69, No.8) For example, a region division method based on a marginal distribution disclosed in US Pat. A peripheral feature vector is calculated for the cut-out character image (step S43). First, this feature vector is compared with the representative vector of the reintegrated category obtained by re-clustering as a large classification,
The one at the shortest distance is extracted, and a representative vector of the divided category belonging to the cluster is obtained (step S4).
4). Next, comparison with the representative vector of the divided category is performed, and the one at the shortest distance is extracted (step S).
45), and output the representative character of the similar character category corresponding thereto (step S46). Steps S43 and above
The process of S46 is sequentially repeated for each extracted character.

In this embodiment, a two-stage identification dictionary is created. However, the identification dictionary can be formed in more layers. That is, the reintegration of the divided categories may be performed, and the reintegration may be performed with a smaller number of categories. As the number of stages increases, the possibility of erroneous recognition increases. Therefore, the appropriate number of stages can be experimentally confirmed by changing the number of categories and the number of stages.

As a modified example, erroneous identification is performed by extracting several categories not only from the shortest distance in the large classification but also from the shortest distance and performing matching with the divided categories belonging to them. Can be reduced. In this case, the number of times of collation increases, but the accuracy naturally increases. The number of candidates can be determined by changing the number of candidates and experimentally confirming them.

As a result, as shown in the specification of Japanese Patent Application No. 8-274732, the obtained similar character category representative code string is bi-gr for easy retrieval later.
am is extracted and registered, or a morphological analysis of a similar character category string is performed, and an allowable word is extracted and registered. At this time, if necessary, detailed identification of character candidates is performed to determine the word. The processing disclosed in Japanese Patent Application No. 8-274732 can be used as it is for the processing after the similar character category string is obtained and for the processing when there are a plurality of character cutout positions.

The present invention realizes a similar character category discrimination method that achieves both accuracy and speed in the process of extracting words from similar character category strings disclosed in Japanese Patent Application No. 8-274732. It can also be used in place of the conventional method of large classification in character recognition.

Now, a large classification process is performed for all N character types by collation with a representative vector in a round robin manner, and the major classification process is replaced by the present invention with the case of the character recognition of the conventional method in which the top n characters are classified in detail. Make a comparison with the case. In the conventional method, the dimension of the feature quantity of the large classification is x, and the dimension of the feature quantity of the detailed classification is y. The identification is the shortest distance identification based on the Euclidean distance. The amount of calculation is substantially proportional to the number of multiplications, that is, the product of the number of distance calculations and the number of dimensions. Therefore, when this is considered as an index of the calculation amount, the calculation amount for one unknown character is Nx + ny. On the other hand, assuming that the number of similar character categories in the present method is M, each similar character category includes an average of N / M character types.
It is assumed that the number of clusters at the time of re-clustering is L, and that each re-clustered cluster includes an average of K divided categories. Assuming that x and y are the same as in the above, the calculation amount is (L + K) x + (N / M) y. Here, it is assumed that the appearance probabilities of the respective character types are all equal.

The effect is shown by numerical values obtained by actually performing experiments. Here, the total character type is about 1/4 character type of the JIS first level. The feature amount used is a peripheral feature, and the number of dimensions is 64. For each character type, several fonts and sizes were changed, and about 400 learning samples were created. The given parameters are
N = 835, M = 500, L = 100. As a result, the total number of divided clusters was 3339 and K = 103.5. When the learning samples were actually identified, the identification rate for the similar character category was 99.3%, and sufficient accuracy was achieved. Assuming that the number x of dimensions of the large classification feature quantity in the conventional method is 64, the number y of the detailed classification dimensions is 256, and the number of candidates in the large classification n = 20. X64 + 20x256 = 58560.
On the other hand, in the present method, (100 + 103.5) × 6
4 + 835/500 × 256 = 1345.52. Therefore, it can be seen that the amount of calculation is about 1/4 or less, which is greatly reduced. If it is assumed that the top 10 categories of the result of classification into similar character categories are taken out in order to improve the accuracy and all the character types included in the top 10 categories are identified in detail, (100 + 103.5) × 64 + 8
35/500 × 256 × 10 = 172999.2,
Even so, the calculation amount is still 1/3 or less. When a similar character category is identified, the number of detailed classifications can be further reduced by using a method of morphologically analyzing a similar character category string disclosed in Japanese Patent Application No. 8-274732. Also, when only a single character type is registered in the identified similar character category (out of 500 similar character categories obtained in the experiment, there were 317 categories in which only one character type was registered) ), Or when the character type can be uniquely identified by morphological analysis, there is no need to perform detailed classification for the character, and therefore, feature amount extraction processing for detailed classification and its matching processing become unnecessary. Therefore, the difference in the amount of calculation is further increased.

Next, another embodiment of the category dividing process will be described. Also in the second embodiment for creating a similar character category identification dictionary, the basic processing is performed according to the flow shown in FIG. 3, and details thereof are the same except for the category division processing. Only the category division processing will be described.

FIG. 14 is a flowchart showing the flow of the category dividing process according to the second embodiment. First,
Attention is paid to one similar character category, and all the learning samples of the character types belonging to the similar character category are extracted (step S51). Next, focusing on one of the learning samples, a similar character category located at the shortest distance is obtained (step S52). If the similar character category is the similar character category currently focused on or the same similar character category already divided, nothing is performed (steps S53 and S54). Otherwise, since misidentification has occurred therein, the learning samples are compiled for each misidentified category.

FIG. 15 is an explanatory diagram of category division in the second embodiment. As shown in FIG. 15, for example, there are similar character categories A, B, C, and D, and some of the samples in the similar character category A are distributed over the identification boundary with the categories B, C, and D. That is, some of the samples belonging to the similar character category A are erroneously identified as three categories B, C, and D. In such a case, the learning samples which are erroneously identified are put together, and the similar character category A is divided into four as shown in the figure.

However, at this point, division is not performed, and only the correspondence between the erroneously identified sample and the shortest distance category is stored (step S55). The inspections in steps S52 to S55 are performed on all the samples, and the inspections in steps S51 to S55 are performed on all the similar character categories, and then the division is actually performed. That is, the average vector of the sample is calculated for each of the erroneously identified categories (step S5).
7) The cluster is divided using the representative vector of the category of interest and the average vector of the sample of the misidentified category as the initial cluster (step S58). When the division is completed, the average vector of the sample prompting the division category is obtained and set as the representative vector. This is performed for all similar character categories including the erroneously identified sample, and the same processing is repeated until erroneous identification is eliminated (step S56).

The division uses the k-means method. The center of the initial cluster collects erroneously identified samples for each erroneously identified similar character category, and sets the average vector and the representative vector of the similar character category of interest. As a result, the identification plane moves as shown on the right side of FIG. 15, and erroneous identification is reduced.

[0068]

As described above, according to the present invention, the similar pattern category identification dictionary is obtained by dividing a once created similar pattern category, a representative vector of these divided categories is used as a detailed classification dictionary, and the divided categories are reintegrated. Then, the representative vector of the reintegrated category is set to have a two-stage configuration in which a large classification dictionary is used. As a result, a similar pattern category identification dictionary capable of performing two-stage matching of a large classification and a detailed classification when matching an unknown pattern is obtained. By using such a dictionary, the identification rate for the reintegrated category is improved to 99.3% as compared with the case where the identification rate for the similar pattern category created first is about 92%. In addition, the amount of calculation is about 1/4 or less as compared with the case of the brute force collation, so that the identification speed can be improved.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a diagram showing a hardware configuration for implementing the present invention.

FIG. 3 is a flowchart illustrating a flow of a process of creating a similar character category identification dictionary.

FIG. 4 is a flowchart illustrating a flow of an image input process.

FIG. 5 is an explanatory diagram of peripheral features.

6A and 6B are diagrams illustrating an example of a table created by a similar character category generation process, wherein FIG. 6A illustrates an example of a similar character category table, and FIG. 6B illustrates an example of a character code / category correspondence table; ing.

FIG. 7 is a flowchart illustrating a flow of a category division process.

FIG. 8 is a diagram for explaining the meaning of an end point vector expression in a feature amount space.

FIGS. 9A and 9B are diagrams illustrating category division by the k-means method, wherein FIG. 9A shows an example of the first division of a similar character category, and FIG. 9B shows an example of a further division of the divided similar character category. ing.

FIG. 10 is a diagram showing a data structure of an identification dictionary by category division.

FIG. 11 is a flowchart illustrating a flow of a reintegration category creation process.

FIG. 12 is a diagram showing a data structure of a large classification dictionary by a re-integrated category creation process.

FIG. 13 is a flowchart illustrating a flow of a similar character category identification process.

FIG. 14 is a flowchart illustrating a flow of a category division process according to the second embodiment.

FIG. 15 is an explanatory diagram of category division in the second embodiment.

FIG. 16 is a diagram illustrating a problem in similar character category identification.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Similar pattern category creation means 2 Category division means 3 Reintegration category creation means 10 Personal computer 11 Central processing unit (CPU) 12 Main storage memory 13 Peripheral device controller 14 External storage device 15 Display 16 Keyboard 17 Pointing device 18 Image scanner 19 Network

Claims (7)

[Claims] 1. A similar pattern category identification dictionary creating apparatus for creating a similar pattern category identification dictionary used for matching when a pattern included in image information is identified as a similar pattern category. A similar pattern category creating means for obtaining similarities between patterns from the extracted pattern information and creating similar pattern categories by grouping similar patterns, and examining a state of misrecognition of a learning sample of a pattern belonging to the created similar pattern category. And a category dividing means for dividing the similar pattern category based on the situation to create a detailed classification identification dictionary composed of the divided categories; obtaining a representative vector of the divided category; and reintegrating the divided categories into similar ones again. Classification Dictionaries with Classified Reintegrated Categories And a reintegrated category creating means for creating a similar pattern category identification dictionary having a hierarchical structure together with the detailed classification identification dictionary. 2. A similar pattern category identification dictionary creating method for creating a similar pattern category identification dictionary used for matching when a pattern included in image information is identified as a similar pattern category, comprising the steps of: Extracting the feature amounts of the patterns included in the learning sample, accumulating the learning samples, examining the similarity between the patterns of the learning samples, creating a similar pattern category in which similar patterns are put together, and creating the similar pattern category. Based on the training sample of the pattern belonging to,
Dividing each similar pattern category to create a divided category, re-assembling the divided categories into similar categories, and creating a reintegrated category that constitutes a hierarchical identification dictionary together with the divided categories. How to create a category identification dictionary. 3. The step of creating the divided category includes performing principal component analysis on a sample belonging to the similar pattern category, obtaining an end point vector on the principal component axis, and erroneously identifying the end point vector as another similar pattern category. 3. The method according to claim 2, wherein, when performed, the similar pattern category is divided by clustering, and the clustering is repeated until no erroneous identification is found. 4. The step of creating the divided category includes: performing principal component analysis on a sample belonging to the similar pattern category; determining an end point of a sample distribution on a principal component axis in a feature space; 3. The sample which belongs to the similar pattern category is divided by clustering from those having large eigenvalues among those which are erroneously identified as other similar pattern categories, and the division is repeated until erroneous identification is eliminated. Method for creating a similar pattern category identification dictionary. 5. The step of creating the reintegrated category includes obtaining a representative vector of the divided category, and if the learning sample is erroneously identified by comparison with the previously obtained representative vector, the learning sample 3. The method for creating a similar pattern category identification dictionary according to claim 2, wherein the re-integration is performed by registering the divided category to which the. 6. The step of creating the reintegrated category includes obtaining a representative vector of the divided category, and if the learning sample is erroneously identified by comparison with the previously obtained representative vector, the learning sample The reintegration category with a large number of categories is created by registering the divided category to which the category belongs to the misidentified category, and the reintegration category with a small number of categories is created recursively using the result as an input, so that multi-stage 3. The method according to claim 2, wherein an identification dictionary is created. 7. The step of creating a divided category includes, when a learning sample belonging to a similar pattern category is erroneously identified as a representative vector of a category other than the category to which it belongs, collects the sample for each erroneously identified category. 3. The method for creating a similar pattern category identification dictionary according to claim 2, wherein the division is performed as a new cluster, and the division is performed until there is no erroneous identification.