JPH07111732B2 - Dictionary creation device for character and figure recognition - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to character / graphic recognition, and in particular, a recognition feature is extracted from input character / graphic data, and the extracted features and a plurality of stored features are stored in a dictionary. The present invention relates to a method of creating a dictionary for a character / graphics recognizing device which compares standard characteristics of character / graphics data with character / graphics data having a standard characteristic having a high degree of coincidence between the characteristics as a recognition result.

2. Description of the Related Art One of the conventional character / graphic recognition methods is to extract some features from a character / graphic to be recognized, and to combine a set of features (feature vectors or simply features) into multiple criteria stored in a dictionary. There is a method in which the reference pattern that has the most similar feature vector is used as the recognition result as compared with the feature vector of the pattern.

However, even if it is the same character figure, it is simple to create a dictionary for handwritten character figure recognition in which the figure varies depending on the writer or each time it is written and a multi-font character recognition that allows multiple glyphs. No reference pattern can be set for. Therefore, in general, a reference feature vector is determined for each type (category) of each character graphic as follows, and a dictionary is created. That is, a feature vector is extracted from a plurality of patterns in the same category (for example, the same character pattern handwritten by several people), and the average of the extracted feature vectors is used as the reference feature vector for that category, and the dictionary of the feature vector is extracted. Store in the dictionary as data. The pattern used to create the dictionary is called learning data.

Learning data A1, A2, ..., Am belonging to two categories Ca and Cb
Two features (f1, f2) obtained for each of (belonging to Ca) and B1, B2, ..., Bn (belonging to Cb) are
A case where the feature space defined by f1 and f2 exhibits the distribution shown in FIG. 2 will be described. In FIG. 2, the characteristic vector of the data belonging to Ca is represented by a black dot, and the characteristic vector of the data belonging to Cb is represented by a white triangular point. The solid line indicates the boundary between both category areas.

Here, for the sake of explanation, the feature vector is 2 with (f1, f2).
However, in general, more features are used instead of two features.

The reference feature vector of each category obtained by the above method is represented by × points Ao and Bo in the figure.

At this time, the feature vector of B3 is closer to Ao than Bo, that is, more similar to Ao than Bo, even though B3 belongs to category Cb. Therefore, if Ao and Bo are recognized as dictionary data, B3 will be erroneously recognized as Ca.

In such a case, a category such as Cb is divided into two groups Cb1 and Cb2 shown by dotted lines in the figure, and average feature vectors B1o and B2 are obtained from the learning data belonging to each group.
If o is obtained and both are registered in the dictionary as category Cb dictionary data, misrecognition can be prevented.

Therefore, by dividing the categories into a plurality of groups and increasing the dictionary data of the categories, it is possible to create a dictionary for handwritten character recognition that gives a higher recognition rate.

When a dictionary is created by this method, generally, the larger the number of groups in each category, that is, the larger the number of dictionary data, the higher the recognition rate of the dictionary.

However, when performing recognition using the created dictionary, the feature vector of the recognition target pattern is compared (distance calculation) with all the dictionary data stored in the dictionary, so the total number of dictionary data stored in the dictionary is When the number is large, the time required for recognition increases accordingly. Therefore, considering the recognition time, the smaller the dictionary data for one category, the better.

Under the above two requirements, the minimum number of groups that can achieve a high recognition rate is the optimum number of groups. Note that, in general, the feature distribution state of the learning data is different for each category, and therefore, the optimal number of groups is naturally different for each category.

The optimal number of groups in a certain category is determined not only by the distribution of the characteristics of the learning data belonging to that category but also by the relationship with the distribution of the data of other categories that are adjacently distributed in the feature space. In order to determine the optimum number of groups, it is necessary to understand the distribution of the characteristics of the learning data in all categories. In general, a huge amount of learning data is used to create a dictionary, so the work of grasping the mutual relationships and determining the optimal number of groups for each category requires a huge amount of calculation. ，
It was impossible to realize.

Therefore, in the past, each category could not be grouped according to the optimal number of groups for that category, and the grouping was performed with a uniform number of groups for each caterigo.

In addition, when actually recognizing using the dictionary created by the above method and the recognition rate is not good, manually
I was adjusting the number of groups.

Problems to be Solved by the Invention In the above-mentioned conventional method of grouping with a uniform number of groups in each category, generally, redundant grouping is performed, and therefore the number of groups depends on the recognition rate. Often increases. On the contrary, when the number of groups is reduced, the recognition rate is significantly reduced. Also, the work of manually adjusting the results requires a great deal of time.

In view of the above-mentioned conventional art, the present invention provides a dictionary creating apparatus for character / figure recognition which has a simple structure and can perform high-speed processing.

Means for Solving Problems In order to solve the above problems, the character / graphic recognition dictionary creation device of the present invention extracts recognition features from a plurality of evaluation character / graphic data prepared in advance and extracts them. By comparing the features with the standard features of multiple character / graphic data stored in the dictionary, the character / graphic data with standard features with a high degree of matching between the features is used as the recognition result. A recognition means for obtaining a recognition rate and a recognition feature are extracted from a plurality of learning character / graphic data for each category, a standard feature for each category is calculated, and an initial dictionary in which the standard feature is stored as dictionary data is created. As a result of recognition of the evaluation character / graphic data by the initial dictionary creating means and the recognizing means, for a category having a lower recognition rate than a predetermined recognition rate
Category data dividing means for increasing the number of groups in the category and grouping learning character / graphic data belonging to the category, and recognition of learning character / graphic data belonging to each group divided by the category data dividing means Dictionary updating means for calculating a standard characteristic of each group from the use characteristics and creating a dictionary in which the standard characteristic in the dictionary of the category to be divided is replaced with the standard characteristic of each group divided by the category data dividing means In the first recognition, the recognition means uses the initial dictionary created by the initial dictionary creation means, and thereafter, the dictionary update means performs recognition using the latest dictionary,
The recognition by the recognition unit, the division of the low recognition rate by the category data division unit into categories, and the update of the dictionary data of the low recognition rate category by the dictionary update unit are repeated until a preset end condition is satisfied. It is a thing.

Effect The present invention creates a dictionary for handwritten character / figure recognition that can realize a high recognition rate by adjusting the number of groups in each category while feeding back the recognition result by the above method and having a small number of dictionary data. To do.

Embodiment An embodiment of the present invention will be described below with reference to FIG.

FIG. 1 is a block diagram of a character / graphic recognition dictionary creating apparatus of the present embodiment. 1 is an initial dictionary creation unit, 2 is a recognition unit, 3
Is a dictionary update unit, 4 is a control unit, 5 is a learning data storage memory, 6 is a recognition data storage memory, 7 is a dictionary memory, 8
Is a recognition result memory.

Hereinafter, each function will be described in detail.

The initial dictionary creating unit 1 creates an initial dictionary by the following procedure.

(1) Features are extracted from the learning character / graphics data stored in the learning data storage memory 5 according to a predetermined rule. It is assumed that a plurality of learning data for each category is prepared in advance in the learning data storage memory 5. Since the extracted feature is generally composed of a plurality of elements, it is treated as a multidimensional feature vector in this embodiment.

(2) The feature vector obtained for each learning data is averaged for each caterigo and used as the reference feature.

(3) The reference feature vector for each caterigo obtained in (2) is stored as dictionary data in the dictionary memory 7 in a predetermined format to create a dictionary. The recognition unit 2 performs recognition using the latest dictionary in the dictionary memory 7 and calculates the recognition rate. As the character / graphic data to be recognized, the learning data in the learning data storage memory 5 may be used, or the character / graphic data for recognition may be separately prepared and used. When preparing the recognition data separately, the recognition data storage memory 6 is provided and the prepared recognition data is stored therein. However, in this case, in order to obtain the recognition rate, it must be known to which category the data belongs.

The recognition rate is the ratio of correctly recognized data to all recognition target data. In a narrow sense, the correct recognition is a category of dictionary data that is most similar to the feature vector of the recognition target data (small distance), that is, the case where the first-rank recognition result matches the category to which the recognition target belongs. However, it may be defined as follows in a broad sense. Based on the similarity (distance) between the feature vector of the recognition target data and the dictionary data, multiple categories are selected as candidates in descending order of similarity (small distance), and among the control categories,
If there is a category to which the recognition target data belongs, it is regarded as correctly recognized. For example, the recognition rate based on the candidate categories up to the third rank at the time of recognition is called the third rank recognition rate.

In the present invention, any recognition rate may be adopted.

In the present invention, the recognition rate is calculated for each category.
That is, each category recognition rate is the ratio of correctly recognized data among all recognition target data belonging to each category.

The obtained recognition rate is stored in the recognition result memory 8.

The dictionary updating unit 3 creates a new dictionary by the following procedure.

(1) Referring to the recognition result by the recognition unit 2 stored in the recognition result memory 8, for a category with a poor recognition rate,
In the dictionary used for recognition, the number of groups is increased from the number of groups in that category, and the learning data belonging to that category is grouped again. The process of focusing on similarities between a plurality of pieces of data, collecting similar pieces and dividing the pieces of data into groups is called clustering. As a clustering algorithm,
The shortest distance method, the longest distance method, the center of gravity method, the median method, the Ward method, etc. are known ("multivariate statistical analysis method", Yutaka Tanaka / Kazumasa Wakimoto, Hyundai Mathematics). In the present invention, since the clustering algorithm does not matter, the grouping is performed by any of the above algorithms.

Further, as the method for increasing the number of groups, it is possible to increase the number of groups by a certain number (for example, increase) or multiply the number of groups by a certain number (for example, double the number of groups). I don't care.

(2) For each group formed as a result of (1), the average of the feature vectors of the learning data belonging to that group is used as the reference feature, and these are used as new dictionary data for that category.

(3) For categories that have not been changed, the dictionary data of the dictionary used for recognition is used as it is as dictionary data.

(4) The dictionary data of (2) and (3) are stored in the dictionary memory 7 in a predetermined format to create a new dictionary. Control unit 4
Determines the end condition and controls the repetition.

The operation procedure of the character / graphic recognition dictionary creating apparatus of the present embodiment will be described below with reference to FIG.

(1) The initial dictionary creating unit 1 creates an initial dictionary and stores it in the dictionary memory 7.

(2) The recognition unit 2 performs recognition using the initial dictionary in the dictionary memory 7, and stores the recognition result in the recognition result memory 8.

(3) The control unit 4 determines the ending condition. If the end condition is not satisfied, the process proceeds to (4), and if the end condition is satisfied, the process proceeds to (7).

(4) The dictionary updating unit 3 creates a new dictionary in which the number of dictionaries is increased based on the recognition result stored in the recognition result memory 8 and stores it in the dictionary memory 7.

(5) The recognition unit 2 is newly created by the dictionary updating unit 3,
The latest dictionary in the dictionary memory 7 is used for recognition, and the recognition result is stored in the recognition result memory 8.

(6) After that, in (3), the processes of (3), (4), and (5) are repeated until the preset termination condition is satisfied.

(7) If the ending condition is satisfied in (3), the iteration is ended, and the latest dictionary in the dictionary memory 7 at that time is output as the final result.

The termination conditions are set for the recognition rate and the number of repetitions. For example, a condition such as "end the repetition when the recognition rate of all categories is 99.5% or more" or "repeat up to 10 times" can be considered. In addition, these conditions may be combined.

EFFECTS OF THE INVENTION It is very difficult to know the optimal number of groups in each category in advance, but according to the present invention, the optimal number of groups can be determined by adjusting the number of groups in each category while feeding back the recognition result. It is possible to realize grouping based on a close number of groups. As a result, it is possible to create a dictionary that realizes a high recognition rate with a small number of dictionary data, which is essential for a high-speed and high-performance character recognition device.

[Brief description of drawings]

FIG. 1 is a functional diagram of the present invention, and FIG. 2 is a diagram showing an example of distribution of characteristics of learning data. 1 …… Initial dictionary creation section, 2 …… Recognition section, 3 …… Dictionary update section, 4
...... Control unit, 5 ...... Learning data storage memory, 6 ...... Recognition data storage memory, 7 ...... Dictionary memory, 8 ...... Recognition result memory.

Claims (2)

[Claims] 1. A feature for recognition is extracted from a plurality of character / graphic data for evaluation prepared in advance, and the extracted feature is compared with standard features of a plurality of character / graphic data stored in a dictionary to match the features. A recognition means that obtains the recognition rate for each type (category) of character and figure by using the character and figure data having a high standard characteristic as the recognition result and determining the correctness of the recognition result. An initial dictionary creating unit that extracts a recognition feature from data, calculates a standard feature for each category, and creates an initial dictionary that stores the standard feature as dictionary data, and a result of recognition character graphic data recognition by the recognition unit. , Category data having a recognition rate lower than a predetermined recognition rate, the number of groups of the category is increased to classify learning character / graphic data belonging to the category. The standard feature of each group is calculated from the dividing unit and the recognition feature of the learning character / graphic data belonging to each group divided by the category data dividing unit, and the standard feature in the dictionary of the category to be divided is calculated. Dictionary updating means for creating a dictionary in which is replaced with the standard features of each group divided by the category data dividing means, and the recognizing means, in the first recognition,
Using the initial dictionary created by the initial dictionary creating means, thereafter, recognition is performed by using the latest dictionary by the dictionary updating means, and recognition by the recognizing means until a preset ending condition is satisfied, A dictionary creating apparatus for character / figure recognition, characterized in that the division of the low recognition rate category by the category data dividing means and the updating of the dictionary data of the low recognition rate category by the dictionary updating means are repeated. 2. A patent characterized in that, in the category data dividing means, grouping is performed based on similarity or distance in a feature space of recognition features extracted from each learning character / graphic data of a division target category. The character dictionary recognizing dictionary creating apparatus according to claim 1.