JP2868770B2 - Structure of pattern dictionary - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to pattern data, in particular, a large number of characters.
The present invention relates to a configuration method of a pattern dictionary for classifying and grouping hierarchically according to its appearance characteristics, increasing the speed of searching and recognizing input unknown characters, and improving the recognition rate.

[Conventional technology]

Conventionally, as a character identification method, the similarity between the input unknown character pattern feature and all the features contained in the pattern dictionary is calculated, and the largest similarity among these similarities is calculated. Is generally used as a pattern in the pattern dictionary that gives the identification character of the input unknown character.

In recent years, there has been an increasing number of schemes in which the procedure of extracting a characteristic of a character pattern is performed in stages by processes such as large classification, intermediate classification, and individual recognition. In this method, first, in a rough feature extraction process, all patterns whose similarity to an input unknown character is equal to or greater than a certain threshold are selected as recognition character candidates (major classification). Next, for all the candidate characters selected in this way, pattern matching with the input unknown character is performed by extracting a larger amount of features than in the previous stage. For example, 10
Narrow down to character types (middle classification). These procedures are repeated in the same manner to finally narrow down to one individual character, and the finally narrowed down one pattern character type is set as a recognition character of the input unknown character.

[Problems to be solved by the invention]

However, according to the conventional character recognition method described above, regardless of which method is adopted, the similarity is obtained by brute-force matching of the input unknown character with all the patterns in the pattern dictionary, and these similarities are obtained. Since the identification processing is performed based on the degree, the amount of calculation becomes enormous, and the processing time required until individual recognition becomes extremely long.

The present invention focuses on the problems of the related art, and has been made in order to solve the problems.In order to solve the problems, a high-speed recognition processing system is implemented, and the pattern dictionary has a hierarchical structure. Also, in order to realize a highly accurate recognition processing system, for example, if pattern data is simply grouped hierarchically, the final constituent elements of each group become independent from each other. As shown in the figure, overlapping portions sharing character features often occur due to practical similarity between group 1 and group 2. However, since overlapping portion 1a is included in group 1 and exceeds the boundary, group 2 overlaps. And 2a is a pattern dictionary configuration method that takes into account the group configuration method that also eliminates the problem of high misclassification rates that are included in group 2 but are identified as group 1. It is intended to provide.

[Means for solving the problem]

In order to achieve this object, a method of constructing a pattern dictionary according to the present invention classifies character patterns into a plurality of groups, and forms a first group having a certain characteristic with the first group when the character patterns are grouped. A character pattern existing in a shared portion with a second group having another characteristic different from that of the first group, belonging to each of the first group and the second group, and grouping the character patterns to form a hierarchically structured pattern In the method of constructing a dictionary, a large number of character patterns are classified into any one of independent groups based on a clustering result for each, and a group to which the user belongs is set as an original group, and a representative pattern is created for each of the original groups. The distance between the character pattern and the representative pattern of the original group is defined as a first distance, and the distance between the character pattern and the original group is determined. The distance between the other group of representative patterns other than the second distance, the said first distance second
If the difference from the distance is smaller than the size of the arbitrarily set shared portion, the character pattern is also assigned to the other group, and the difference between the distances described above is calculated for each character pattern for each group. [Function] As described above, a plurality of pattern groups are created as described above, and a portion where similar overlapping occurs between groups belongs to any group. In other words, the erroneous classification rate is greatly reduced, and highly accurate recognition of the input unknown character can be obtained. Furthermore, by subdividing each group into a plurality of groups, a pattern dictionary can be configured hierarchically, character classification / recognition processing can be realized with a small number of matchings, and processing speed can be increased.

〔Example〕

Next, an embodiment of the present invention will be described in detail with reference to FIGS. 1 to 4C.

FIG. 1 is an explanatory view showing the concept of the gist of the configuration system of the pattern dictionary according to the present invention.
In particular, assuming that the character data has a normal distribution in the distributions 11 and 12 of the pattern group due to its appearance characteristics,
When 1 and 12 have a shared portion in the vicinity, compared to an arbitrarily defined boundary, the 11a portion is the distribution 1 while being the distribution 11.
It also belongs to 2, while 12a part is distribution 12 but distribution 11
The correction operation is performed on the components in each of the distributions 11 and 12 by belonging to. That is, although the data size is slightly larger, the shared portions 11a and 12a belong to both the groups 11 and 12.

FIG. 2 is a flowchart showing a schematic flow of a process for forcibly making a shared part belong to each group as described above. Each of the areas E1 to E4 separated by chain lines in FIG. 3A to 4C will be specifically described later. First, a large number of pattern data (especially character types) to be grouped are input to create a representative pattern of each target group, and then a distance between a standard pattern of each character type and a representative pattern in each group is determined. ,
The difference between the distance between the standard pattern of each character type and the representative pattern of the group to be included and the distance between the standard pattern of each character type and the representative pattern of each group is determined. Then, the size of the shared portion ε is compared with the difference obtained above, and if the size of the shared portion ε is equal to or larger than the difference, the standard pattern shall belong to the group that calculated the difference,
If it is smaller, the standard pattern does not belong to the group for which the difference has been calculated.

Here, the areas E1 to E4 of the flowchart shown in FIG. 2 will be specifically described. FIG. 3A to FIG.
5 is a detailed flowchart of the first embodiment. In this flowchart, n types of characters are specifically grouped.
Here, the clustering result is input, and the number GG that should contain the I-th standard pattern is replaced with CLNUM (I),
This operation is repeated n times the number of characters (FIG. 3A). When this processing is completed, in FIG. 3B, CDAT and NDAT (both variable names used for operation) are set to “0” (initialization), and the feature vector of the standard pattern of each character type in the α feature and the β feature is set. By inputting (DAT), the feature vectors of the standard patterns of the character types included in each group are added (CDAT). During this process, the vector is a combination of α and β features (BDAT). Here, the α feature is p-dimensional, and the β feature is q-dimension. During this process, the feature becomes p + q = r-dimension, and this process is repeated n times the number of characters. When this process is completed, an average feature vector (representative pattern) is created for each group in FIG. 3C, and CDAT (I, J) is the sum of feature vectors of character types included in the group in each group. NDAT (I) is a character type in each group. Here, I is the group number, and J is the number of characteristic items.
Thus, the average feature vector (representative pattern) of each group can be created by using CDAT / NDAT. The number of repetitions gn of the process is the number of groups. When the representative pattern is created in this manner, the processing shifts to E2 in FIG. 2, and the E2 is as shown in FIG. 3D. In FIG. 3D, ADAT is a variable representing the square distance. In this process, the square distance between the standard pattern of each character type and the representative pattern of each group is obtained (operation A). The difference between the square distance and the square distance between the standard pattern of each character type and the representative pattern of the included group is determined and set as EDAT (operation B). This EDAT is output and input to E3 in FIG.

E3 in FIG. 2 corresponds to the processing shown in FIGS. 4A and 4B. In FIG. 4A, the EDAT (difference in the squared distance) obtained in E2 is input. This is repeated n times the number of characters. This ED
When the input of AT is completed, the size of the shared portion ε between the groups is input to SH as shown in FIG. 4B. The value of ε is obtained by search. If EDAT is smaller than SH, “1” is substituted for ADAT, and if it is larger, “0” is substituted. Therefore, this 4B
The ADAT in the figure is different from the one used up to E2.
This processing operation is repeated for each character, that is, n times.
The processing in FIG. 4B allows each character type to share the character type with a group other than the group determined by the clustering result described above, and whether or not to share the character depends on the size of EDAT. Distance threshold SH
(The magnitude of ε). When the group to which each character type belongs is determined in this way, the processing shifts to E4 in FIG. 2, and this output is as shown in FIG. 4C. Data output ADAT (I, J) is assigned "0" or "1" and is output, where I is a character type number and J is a group number. ADAT (I,
When J) is “1”, it indicates that the I-th character type is included in the J-th group.

In the processing described above, pattern data is grouped and classified according to features. The above-described processing is performed on the classified pattern data in each group by a new characterization, and the groups are subdivided. In this way, the groups are subdivided into a hierarchical structure and a dictionary of pattern data is created.

〔The invention's effect〕

As described above, according to the configuration method of the pattern dictionary according to the present invention, since the shared portion in the distribution of each pattern belongs to a plurality of groups, the misclassification rate decreases, and the input unknown character Recognition rate increases. Furthermore, the same grouping is performed for each group and subdivided into a hierarchical structure, so that input unknown characters can be recognized and searched at higher speed.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the concept of the gist of the pattern dictionary construction method according to the present invention, FIG. 2 is a flowchart showing the outline of the flow of the processing, and FIGS. 3A to 3D are E1 in FIG. ~ E2
4A to 4C are flowcharts showing the details of E3 to E4, and FIG. 5 is a diagram for explaining a case where the information is simply classified hierarchically. 11 ・ 12 …… Pattern distribution 11a ・ 12a …… Shared part

Claims (1)

(57) [Claims] 1. Classifying character patterns into a plurality of groups,
When grouping character patterns, the first
And a character pattern existing in a shared portion between the group of the first group and the second group having another characteristic different from that of the first group is assigned to each of the first group and the second group to form a group, In the method of constructing a hierarchically structured pattern dictionary, a large number of character patterns are classified into any one of independent groups based on a clustering result, and a group to which the character pattern belongs is defined as an original group. A representative pattern is created for each, a distance between the character pattern and the representative pattern of the original group is a first distance, a distance between the character pattern and a representative pattern of a group other than the original group is a second distance, If the difference between the first distance and the second distance is smaller than the size of the arbitrarily set shared portion, A pattern dictionary, wherein the character patterns belong to the other group, and the individual character patterns are grouped by calculating the difference in distance described above for each group and determining whether to belong to another group. Configuration method.