JPH05210740A - Pattern recognition dictionary preparing system and its additional pattern preparing method - Google Patents

Abstract

PURPOSE:To efficiently prepare a reference pattern to be added to a pattern recognition dictionary for reference patterns at the time of adding a new category. CONSTITUTION:In the case of adding a reference pattern be longing to a new category to the pattern recognition dictionary 5, a similar reference pattern most similar to a sample pattern belonging to the new category is selected from the dictionary 5 and weighted means with a similar reference pattern whose variation has been already absorbed is applied to an element estimated to have large variation independent of the category out of elements constituting the sample pattern to remove variation independent of the category and prepare the reference pattern to be added.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for creating a pattern recognition dictionary used in recognition processing of categories such as characters and voices, and particularly, to efficiently add a pattern belonging to a new category to the pattern recognition dictionary. The present invention relates to a pattern recognition dictionary creating system and a method for creating an additional pattern thereof.

[0002]

2. Description of the Related Art At present, for the purpose of improving the usability of a computer, there are systems that perform input to the computer by voice, images, figures, characters, etc., such as a voice recognition system, a character recognition system, and It has been developed as a figure recognition system. In these recognition systems, an input pattern such as a character or a voice to be recognized is collated with a reference pattern registered in a pattern recognition dictionary in advance, and a reference pattern most similar to the input pattern is used as an object. Recognition processing is performed by extracting categories such as words and characters. Regarding such a recognition system, for example, "Electronic Information and Communication Handbook" edited by The Institute of Electronics, Information and Communication Engineers of Japan (1988)
(Ohm Co., Ltd., pp. 1114 to 1124), a technique relating to recognition of figures and characters is described in 119
From page 1 to page 1206, techniques relating to voice recognition are described.

In these recognition systems, the pattern to be recognized is usually composed of elements that reflect the structure of the category to which it belongs and elements that do not. Therefore, when a reference pattern belonging to a new category is added to the pattern recognition dictionary, it is necessary for the added reference pattern to have sufficient discriminating ability from reference patterns of other categories. For that, the error of the element that does not reflect the structure of the category,
Need to be absorbed.

There are roughly two types of conventional techniques for adding reference patterns of a new category to such a pattern recognition dictionary. One is to register a pattern belonging to a new category as it is as a reference pattern of the recognition dictionary, and after adding the reference pattern to the recognition dictionary in this way, recognize some patterns belonging to the new category, The registered reference pattern is changed based on the information of the pattern. The other is to prepare a plurality of different patterns belonging to the category to be added and to superimpose them to create a reference pattern for the category to be added.

As described above, in the conventional technique, in order to absorb the variation of the reference pattern to be added which does not depend on the category, after the reference pattern is created, the pattern belonging to this category is used for learning. Use the average of multiple patterns. However, when the number of times of learning is small or the number of prepared patterns is small, it is not possible to sufficiently absorb the variation that does not depend on the category from the created reference pattern. If the fluctuation is to be absorbed sufficiently, the number of times of learning or the number of prepared patterns must be increased.

[0006]

The problem to be solved by the prior art is that in the prior art, when a reference pattern belonging to a category to be newly added is added to the pattern recognition dictionary, the reference pattern to be added and other To improve the accuracy of identification with reference patterns belonging to the existing category of
The number of times of learning or the number of prepared patterns must be increased, and new categories cannot be added efficiently. An object of the present invention is to solve these problems of the prior art, to enable creation of a reference pattern that absorbs variations independent of categories by inputting one or a small number of sample patterns, and to create a new category. It is to provide a pattern recognition dictionary creating system and an additional pattern creating method thereof that enable efficient addition.

[0007]

In order to achieve the above object, the pattern recognition dictionary creating system and its additional pattern creating method of the present invention (1) register a reference pattern created by superimposing a large number of learning patterns. In the pattern recognition dictionary creation system that adds a reference pattern belonging to a new category to a pattern recognition dictionary, a sample pattern belonging to the category to be added and this sample pattern already registered in the pattern recognition dictionary It is characterized in that a reference pattern synthesizing unit for creating a reference pattern of a category to be added is provided by a weighted average with a similar reference pattern that is most similar to. (2) In the additional pattern creating method of the pattern recognition dictionary creating system according to (1), the reference pattern synthesizing unit uses a weighted average of the sample pattern and the similar reference pattern to create the similar reference pattern. Based on the distribution of all learning patterns
It is characterized in that the weight of each of the sample pattern and the similar reference pattern is changed. (3) In the additional pattern creating method of the pattern recognition dictionary creating system according to (2), the reference pattern combining unit
The weighting of each of the sample pattern and the similar reference pattern is a component of the learning pattern group used to create the similar reference pattern, and for elements with a total variance or a small variance ratio, the weight of the similar reference pattern is large, and The feature is that the weight of the sample pattern is increased for an element having a large variance or a large variance ratio. (4) In the pattern recognition dictionary creating system and the additional pattern creating method thereof according to any one of (1) to (3), the similar reference pattern is selected based on a plurality of sample patterns. ..

[0008]

In the present invention, when a reference pattern of a category not registered in the pattern recognition dictionary is added, the sample pattern of the category to be added and the existing reference pattern already registered in the pattern recognition dictionary are combined. Then, the reference pattern of the additional category is created. An existing reference pattern that is most similar to the sample pattern of the category to be added is used for this synthesis.
Furthermore, in this composition, among the constituent elements of the pattern, the element whose variance (total variance) of all learning patterns used when creating similar existing reference patterns was small, or the variance between classes (categories) A value obtained by dividing the variance within a class (category), that is, an element having a small variance ratio is heavily weighted on an existing reference pattern, and an opposite element is heavily weighted on a sample pattern. In the pattern group used to create the existing pattern recognition dictionary, the total variance or the part with a large variance ratio is an effective part for separating categories, but conversely, the small part is a category-specific variation. Compared to, the category-independent variation is relatively large. Moreover, since the existing reference pattern is created from a large number of learning patterns, variations that do not depend on the category are absorbed. So, of the sample pattern,
For a portion where the variation that does not depend on the category is large, a weighted average with a similar existing reference pattern in which the variation is already absorbed is taken, and the variation that does not depend on the category is removed from the sample pattern. As a result, it is possible to create a reference pattern having sufficient discrimination ability with respect to reference patterns belonging to other existing categories.

[0009]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a first embodiment of a configuration according to the present invention of a pattern recognition dictionary creating system according to the present invention. The pattern recognition dictionary creation system according to the first embodiment includes a preprocessing unit 1 that performs shaping processing such as normalization for a pattern that belongs to an input category to be added, such as characters and voices, and the preprocessing unit 1. The reference pattern most similar to the input pattern subjected to the shaping process is related to the present invention by the collation determination unit 2 that identifies it based on the pattern recognition dictionary 5, and the reference pattern identified by the collation determination unit 2 and the preprocessing unit 1 The pattern recognition is performed on the reference pattern synthesizing unit 3 that performs a weighted average with the processed input pattern to create a reference pattern of a category to be newly added, and the reference pattern created by the reference pattern synthesizing unit 3. The reference pattern adding unit 4 is registered in the dictionary 5. With such a configuration, the pattern recognition dictionary creation system according to the first exemplary embodiment, when adding a reference pattern belonging to a category to be newly added, such as a character, a figure, or a voice, to the pattern recognition dictionary,
For reference patterns that belong to other existing categories,
A reference pattern with sufficient discrimination ability can be created. Hereinafter, the processing operation of the pattern recognition dictionary creating system according to the present invention will be described based on the operation description of the preprocessing unit 1, the matching determination unit 2, the reference pattern synthesizing unit 3, the reference pattern adding unit 4, and the like.

First, a pattern belonging to a category to be newly added is input from an input device such as a scanner camera or a microphone, and the preprocessing unit 1 normalizes the position and size and removes isolated points. Performs shaping processing of. Next, in the collation determination unit 2, the input pattern processed by the preprocessing unit 1 is collated with the reference pattern stored in the pattern recognition dictionary 5 to determine the most similar reference pattern. The most similar reference pattern information and the input pattern information processed by the preprocessor 1 are passed to the reference pattern synthesizer 3. Then, in the reference pattern synthesizing unit 3, the reference pattern determined by the matching determination unit 2 is taken out from the pattern recognition dictionary 5, the weighted average is performed between this reference pattern and the input pattern, and the reference pattern of the category to be newly added. To create.

The weighted average performed by the reference pattern synthesizing unit 3 will be described below. When performing weighted averaging by the reference pattern synthesis unit 3, in the pattern group used when creating the existing dictionary, the pre-processing unit 1 processes the elements with a large total variance for each of the elements that make up the pattern. The weighting of the input pattern is increased, and the weighting of the reference pattern determined to be the most similar by the collation determining unit 2 is increased for the element having a small total variance. That is, when the pattern is composed of n elements, the input pattern X processed by the preprocessing unit 1 is X = {X ₁ , X ₂ , ..., Xn), and the matching determination unit 2 Then, the reference pattern Cx determined to be most similar to the input pattern X is Ci = {Ci ₁ , Ci ₂ , ..., Cin), and the reference pattern Cx of the category to which the input pattern X belongs is Cx = {Cx ₁ , Cx ₂ , ..., Cxn). If the weight used in the weighted average is f (i) (where 0 ≦ f (i) ≦ 1), the element Cxi of Cx
The weighted average for (i = 1, 2, ..., N) is Cxi = f (i) × Xi + (1-f (i)) × Cii (1). Here, if the variance of the n-th element in the pattern group used when creating the existing dictionary is σn ² , then if σi ² ≦ σj ² then f (i) ≦ f (j) With respect to, the weighting of the input pattern Xi processed by the preprocessing unit 1 is made large, and for the element of which the total variance is small, the weighting of the reference pattern Cx specified by the collation judging unit 2 is made large.

Further, when using the dispersion ratio (inter-class dispersion / intra-class dispersion) instead of using the total dispersion when performing weighted averaging, the existing dictionary is created when formula (1) is used. When the variance ratio of the n-th element in the pattern group used is Fn, if Fi ≦ Fj, then f (i) ≦ f (j).

In this way, the reference pattern created by the reference pattern synthesizing unit 3 is passed to the reference pattern adding unit 4, and the reference pattern adding unit 4 registers this received reference pattern in the pattern recognition dictionary 5. To do.

The pattern to be recognized is not only the pattern of characters, images, voice, etc., but also
It is also possible to use feature patterns extracted from those patterns, and such a system will be described with reference to FIG. 2 below.

FIG. 2 is a block diagram showing a second embodiment of the configuration according to the present invention of the pattern recognition dictionary creating system according to the present invention. The pattern recognition dictionary creating system of the second embodiment is configured by providing a feature extracting unit 6 in the pattern recognition dictionary creating system of the first embodiment in FIG. In the pattern recognition dictionary creating system having such a configuration, the input pattern of characters, figures, voices, etc., which has undergone shaping processing by the preprocessing unit 1, is extracted by the feature extraction unit 6
Then, the feature extraction unit 6 extracts features from the input pattern and creates feature patterns. Then, the feature pattern created in this manner is used as the matching determination unit 2
In the same manner as the pattern recognition dictionary creating system of the first embodiment shown in FIG. 1, the reference pattern synthesizing unit 3 creates a reference pattern belonging to a category to be newly added as a target of the collation determination in.

FIG. 3 is a flow chart showing an embodiment of the processing operation according to the present invention of the pattern recognition dictionary creating system in FIGS. Patterns that belong to the category you want to add (sample patterns) are recognized using the existing pattern recognition dictionary using the same recognition method as the normal recognition method, and it is determined that they are the most similar to the patterns that belong to the category you want to add. The selected similar reference pattern is selected (step 101). A weighted average of the similar reference pattern selected in this way and the sample pattern is performed to create a reference pattern of a category to be added (step 102). Note that this weighted average is, for example, the pattern set used when creating the selected similar reference pattern in each of the elements constituting the pattern, and the total variance or the element with a large variance ratio is the sample pattern The weight of the element is large, and conversely, the weight of the similar reference pattern is increased for the element with the total variance or the small variance ratio. The reference pattern created in this way is registered in the pattern recognition dictionary (step 103).

As described above with reference to FIGS. 1 to 3, in the pattern recognition dictionary creating system and the additional pattern creating method thereof according to the present embodiment, the reference pattern belonging to the category to be newly added is created when the reference pattern is created. Among the elements that make up the reference pattern, for the elements that are estimated to have a large variation that does not depend on the category, take the weighted average with similar reference patterns that have already absorbed the variation,
Remove the category-independent variation from the reference pattern of the added category. As a result, it is possible to create a reference pattern having sufficient discrimination ability with respect to reference patterns belonging to other existing categories.
The present invention is not limited to the embodiment described with reference to FIGS. For example, in this embodiment, the input pattern processed by the preprocessing unit is passed to the reference pattern synthesis unit via the collation determination unit, but the input pattern may be passed from the preprocessing unit to the reference pattern synthesis unit. good. Further, the selection of the similar reference pattern may be obtained from either one sample pattern or the average of a small number of sample patterns.

[0018]

According to the present invention, when a reference pattern belonging to a category to be newly added is added to the pattern recognition dictionary, only one or a small number of input patterns belong to other existing categories. It is possible to create a reference pattern having sufficient discrimination ability with respect to the reference pattern, and it is possible to efficiently add a new category.

[0019]

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of a configuration according to the present invention of a pattern recognition dictionary creating system according to the present invention.

FIG. 2 is a block diagram showing a second embodiment of the configuration according to the present invention of the pattern recognition dictionary creating system according to the present invention.

FIG. 3 is a flowchart showing an embodiment of a processing operation according to the present invention of the pattern recognition dictionary creating system in FIGS. 1 and 2.

[Explanation of symbols]

 1 Pre-Processing Section 2 Matching Determination Section 3 Reference Pattern Synthesis Section 4 Reference Pattern Addition Section 5 Pattern Recognition Dictionary 6 Feature Extraction Section

Claims (4)

[Claims] 1. A pattern recognition dictionary in which a reference pattern created by superimposing a large number of learning patterns is registered,
In a pattern recognition dictionary creating system for adding a reference pattern belonging to a new category with the addition of a new category, the input sample pattern belonging to the category to be added and the pattern recognition dictionary most similar to the sample pattern have already been added to the pattern recognition dictionary. A pattern recognition dictionary creating system characterized by comprising reference pattern synthesizing means for creating a reference pattern of the category to be added by weighted averaging with similar registered reference patterns. 2. The additional pattern creating method of the pattern recognition dictionary creating system according to claim 1, wherein the reference pattern synthesizing unit uses a weighted average of the sample pattern and the similar reference pattern to create the similar reference pattern. An additional pattern creating method for a pattern recognition dictionary creating system, characterized in that the weights of the sample pattern and the similar reference pattern are changed based on the distribution states of all the learning patterns used. 3. The additional pattern creating method of the pattern recognition dictionary creating system according to claim 2, wherein the weighting of each of the sample pattern and the similar reference pattern by the reference pattern synthesizing unit is performed to create the similar reference pattern. Among the constituent elements of the learning pattern group used, the weight of the similar reference pattern should be large for elements with a small total variance or small variance ratio, and the weight of the sample pattern should be large for elements with a large total variance or large variance ratio. An additional pattern creation method for a pattern recognition dictionary creation system characterized by. 4. The pattern recognition dictionary creating system and the additional pattern creating method thereof according to claim 1, wherein the similar reference pattern is selected based on a plurality of the sample patterns. A method for creating additional patterns in the pattern recognition dictionary creation system.