JP2991750B2 - Dictionary creation method - Google Patents

Description

The present invention relates to a pattern recognition device that performs character recognition or voice recognition based on a matching distance with a pattern dictionary, and a dictionary template effective within a predetermined dictionary size. It relates to a method for creating a dictionary composed of numbers.

(Prior art)

Conventional pattern dictionaries usually include a plurality of pattern dictionary templates, because even if the target patterns to be recognized are in the same category, the patterns vary due to the influence of noise or the like.

In particular, in the case of type-printing character recognition, a dictionary template is provided for each font since the difference in shape between the character fonts is clear.

As a related technique of this type, a standard pattern creating apparatus (see Japanese Patent Application Laid-Open No. 59-59139) is used to group a large number of feature patterns and use feature patterns representing the group as standard patterns.
No. 163680).

[Problems to be solved by the invention]

However, the above-described technique has a problem that the dictionary size is large and the recognition processing speed is slow. In particular, the dictionary for recognizing various fonts with the minimum number of templates while keeping the total number of dictionary templates within a predetermined range is problematic. The configuration was not considered.

An object of the present invention is to provide a dictionary creation method capable of changing the number of dictionary templates according to a feature pattern within a predetermined dictionary size range.

[Means for solving the problem]

In order to achieve the above object, the dictionary creation method of the present invention includes a step of classifying a feature pattern of an input character into a plurality of clusters, a step of generating a standard pattern from a representative vector of each of the classified clusters, The method includes a step of evaluating the distribution of the characteristic pattern of the input character, and a step of determining the number of dictionary templates of the standard pattern for the input character based on the value evaluated in the evaluation step. In another embodiment of the present invention, the appearance frequency of the character is further evaluated in the evaluation step.

(Operation)

A feature amount is extracted from a plurality of character image data, an N-dimensional feature vector is obtained, and the clustering unit clusters the plurality of feature vectors. One cluster thus clustered is J, and two clusters are K ₁ and K ₂ . The reference pattern generator, J, K _1, K ₂
For each cluster, a vector representative of the cluster is calculated, and the representative vector, for example, an average vector in the cluster is registered as a standard pattern in the standard pattern holding unit. The evaluation value calculation unit calculates an evaluation value indicating a measure of how much a feature vector of an arbitrary character is divided into two clusters, registers the result in an evaluation value holding unit, and selects a dictionary template number selection unit. Reads out the evaluation values of all the characters registered in the evaluation value storage,
For characters from smaller characters of that evaluation value to P ₁ type and the number of dictionary template to one, for the remaining P ₂ types of characters, the number of dictionary template to two.

〔Example〕

Hereinafter, an embodiment of the present invention will be specifically described with reference to the drawings.

FIG. 1 shows a block diagram of a dictionary creation device of the present invention. In the figure, reference numeral 1 denotes a feature extracting unit for extracting features from image data for creating a dictionary, and 2 denotes a clustering unit for clustering a plurality of extracted feature vectors into one or two clusters. Reference numeral 3 denotes a standard pattern creating unit that creates a standard pattern by calculating an average vector in each cluster. Reference numeral 4 denotes a standard pattern holding unit that holds the standard pattern created by the standard pattern creating unit. An evaluation value calculation unit that calculates an evaluation value representing a measure of how much a character feature vector is divided into two clusters, 6 is an evaluation value storage unit that holds the evaluation value, and 7 is a readout from the evaluation value storage unit 6. A dictionary template number selection unit 8 for selecting the number of dictionary templates of the standard pattern held in the standard pattern holding unit according to the output evaluation value. Reference numeral 8 denotes a pattern dictionary composed of the number of templates selected by the selection unit.

FIG. 2 is a flowchart for explaining the dictionary creation operation of the present invention. The operation of the present invention will be described in detail below with reference to this flowchart. Now, the number of dictionary templates (dictionary feature patterns) of each character is calculated from a character consisting of one dictionary template and two dictionary templates. Assuming that a pattern dictionary is created with characters of the same type, when creating _one type of character consisting of one dictionary template as P and _two types of characters consisting of two dictionary templates as P, the pattern dictionary determined by the processing speed relationship Let T be the upper limit of the total number of templates. If the number of character types registered in the dictionary is C (P ₁ + P ₂ ), T = P ₁ + 2P
_{2, P 1 = 2C-T} , P 2 = T-C, although established (C <T <2C),
How to effectively allocated to _one and P ₂ or P with a C-number of character types, described below.

A plurality of image data are prepared for each character, and the feature extracting unit 1
, A feature quantity is extracted to obtain an N-dimensional feature vector (step 20). The clustering unit 2 recognizes a pattern having similar features as one class by finding a cluster of patterns, that is, clusters, from a plurality of feature vectors, and classifies (clusters) them into one cluster and two clusters. (Step 21). As a method of this clustering, for example, non-hierarchical clustering such as K-mean clustering, centroid method, Ward
There is a hierarchical clustering method such as the method. One cluster thus clustered is represented by J (entire data),
Let the two clusters be K ₁ and K ₂ . FIG. 3A is a diagram representing one cluster J, where each point represents a feature vector in the feature space, and points having similar feature amounts form a cluster in the feature space. FIG. 3B is a diagram expressing _two clusters K ₁ and K ₂ , and the feature space is classified into two categories.

The standard pattern creation unit 3 calculates, for each cluster of J, K ₁ , and K ₂ , a vector representing the cluster. As the representative vector, for example, an average vector in a cluster is used (step 22). These average vectors are registered in the standard pattern holding unit 4 as standard patterns (step 23).

Next, the evaluation value calculation unit 5 calculates an evaluation value δ representing a measure of how much a feature vector of an arbitrary character is divided into two clusters by the following equation.

δ = b− (a ₁ + a ₂ ) Here, a ₁ and a ₂ represent the variance in each cluster when there are two clusters, and b represents the variance in the case where there is one cluster. That is, in FIG.
K _1, if divided into K _2, dispersing a ₁ and the average of the cluster K ₁ to a point 42, the cluster from the average 42 clusters K ₁
The square of the distance feature vector of the K _1, cluster
It is obtained by adding all of the feature vectors in K _1, the variance a ₂ Similarly for cluster K ₂ is calculated. Assuming that the average of the clusters K ₁ and K ₂ is 41 points, the variance b is to add the square of the distance from the average 41 to an arbitrary feature vector for all the feature vectors in the clusters K ₁ and K ₂ . It is calculated by From this, the value of b increases as the two clusters are separated, that is, the evaluation value δ increases. On the other hand, the evaluation value δ decreases as the two clusters approach.

An evaluation value is calculated for each character (step 24), and the result is registered in the evaluation value holding unit 6. The above processing is performed for all characters (step 25). Next, the dictionary template number selection unit 7 reads out the evaluation values of all the characters registered in the evaluation value holding unit 6 and sorts them in ascending order of the evaluation values (step 26). Then, the number of dictionary template to one for characters from smaller characters evaluation value to P ₁ type, reads out the standard patterns corresponding to the characters from the cluster J standard pattern storage 4, as a dictionary template Register (steps 27 and 28). The remaining P ₂ types of characters, the number of dictionary template in two, from the cluster K _1, K ₂ standard pattern holding unit 4 reads the reference pattern is registered as a dictionary template (step 27, 29).

FIG. 5 is a block diagram showing another embodiment of the present invention.
This is obtained by adding an appearance frequency holding unit 9 to the configuration of the previous embodiment. Other components are the same as those described above,
Description is omitted.

In this embodiment, using the value read from the evaluation value storage unit 6 and the appearance frequency read from the appearance frequency storage unit 9, a new evaluation value for the character is calculated using the evaluation value calculation unit 5.
The calculation is made as follows. Here, the appearance frequency is an amount indicating the rate at which the character statistically appears in an actual document or the like. For example, the appearance frequency H is given as an integer value from 0 to 10. Assuming that the larger the value is, the higher the appearance frequency is, the following expression is used as a new evaluation value μ.

 μ = α · δ + β · H where α and β are parameters.

FIG. 6 is a flowchart for explaining the operation of the second embodiment. The difference from the operation of the previous embodiment is that
The difference is that a step of calculating a new evaluation value μ from the appearance frequency H and the evaluation value δ is added between steps 25 and 26, and the other steps are exactly the same, so that the description of the operation is omitted.

In the above processing, the case where there are two types of dictionary templates for each character (one dictionary template and two dictionary templates) has been described. However, the processing can be performed in a similar manner when three or more types of dictionary templates are used. When each character is composed of three types of dictionary templates (one dictionary template, two dictionary templates, and three dictionary templates), the following is performed.

When a pattern dictionary is created such that a character having one dictionary template is P ₁ type, a character having ₂ dictionary templates is P ₂ types, and a character having 3 dictionary templates is P ₃ types. Assume. Where P ₁ , P ₂ , P ₃
It is assumed that P ₁ + 2P ₂ + 3P ₃ = T and P ₁ + P ₂ + P ₃ = C are satisfied.

For each character, the result of clustering into three clusters is L ₁ , L ₂ , L ₃ , the result of clustering into two clusters is K ₁ , K ₂ , and the result of clustering into one cluster is J. Assuming that the intra-cluster variance of Li (i = 1, 2, 3) is ai, the intra-cluster variance of Ki is bi, and the intra-cluster variance of J is c, the evaluation values δ ₁ and δ ₂ are as follows.

δ ₂ = Σbi−Σai, δ ₁ = c−Σbi These calculation results are registered in the evaluation value holding unit in the same manner as described above. Number dictionary template selection unit, from those first small evaluation value [delta] ₁ to the number of dictionary template to one for characters up P ₁ kind in order, and then starting with the lowest evaluation value [delta] ₂ to P ₂ kinds in order The number of dictionary templates is set to two for the character of, and the remaining _three types of characters are set to three dictionary templates.

〔The invention's effect〕

As described above, according to the present invention, the number of dictionary templates can be changed in accordance with the distribution status of feature patterns or the appearance frequency of categories, so that the pattern distribution of the external world can be more faithfully reflected in the pattern dictionary. Is possible, and the recognition rate can be significantly improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a dictionary creating apparatus according to the present invention.
FIG. 3 is a flowchart for explaining the dictionary creation operation of the present invention. FIG. 3A is a diagram expressing one cluster J.
Figure 3 (b) is two clusters K _1, diagram expressing the K _2, Figure 4 is a diagram for explaining the dispersion when divided into two clusters K _1, K _2, FIG. 5 FIG. 6 is a block diagram of another embodiment of the present invention, and FIG. 6 is a flowchart for explaining the operation of the second embodiment. DESCRIPTION OF SYMBOLS 1 ... Feature extraction part, 2 ... Clustering part, 3 ... Standard pattern creation part, 4 ... Standard pattern holding part, 5 ... Evaluation value calculation part, 6 ... Evaluation value holding part, 7 ... Dictionary template Number selection unit 8 Pattern dictionary 9 Appearance frequency holding unit

Claims (3)

(57) [Claims] 1. A pattern dictionary creating method for creating a pattern dictionary, comprising the steps of: classifying a feature pattern of an input character into a plurality of clusters; and generating a standard pattern from a representative vector of each of the classified clusters.
A dictionary creation method, comprising: evaluating a distribution state of a characteristic pattern of the input character; and determining the number of dictionary templates of a standard pattern for the input character based on the value evaluated in the evaluation step. 2. The dictionary creation method according to claim 1, wherein said evaluating step further evaluates the appearance frequency of the character. 3. The dictionary creating method according to claim 1, wherein the total number of the determined dictionary templates is within a predetermined upper limit number.