JP3017325B2 - How to create a dictionary for pattern recognition - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for creating a dictionary for pattern recognition of characters and voices, and more particularly to a technique capable of efficiently creating a recognition dictionary with high recognition accuracy.

[0002]

2. Description of the Related Art In a conventional character or voice pattern recognition method, a reference pattern, which is a standard pattern of each category, is prepared in advance as a recognition dictionary, and a pattern to be recognized is compared with a reference pattern of each category. After calculating the similarity using an evaluation function such as a Euclidean distance or a weighted Euclidean distance, a category having the highest similarity is output as a recognition result.

There are two types of recognition dictionaries prepared in advance: a single template dictionary in which the number of reference patterns is one for each category, and a multi-template dictionary in which the number of reference patterns is plural for each category. The following creation methods have been proposed for each recognition dictionary.

[0004] 1. Single template dictionary creation method 1-. There is a method of calculating an arithmetic mean of several patterns for each category and using the arithmetic mean as a reference pattern. This is a very simple method used in the early days of pattern recognition technology.

1-. After the recognition dictionary is obtained by the above method 1, a process of recognizing a learning pattern with the recognition dictionary and optimizing the position of the reference pattern so as to minimize erroneous recognition is repeatedly performed. As an optimization method, a method employing a simulated annealing method applied to layout design of LSI (large-scale semiconductor integrated circuit) has been proposed [Matsunaga, Abe, Kida: "Simulated annealing method Optimization of the character recognition dictionary used ”, IEICE, PRU90-39 (1990)].

[0006] 2. Multi-template dictionary creation method 2-. After obtaining the recognition dictionary by the method 1-,
In this method, a dictionary is created by recognizing a learning pattern with the recognition dictionary and newly adding a pattern that cannot be read correctly to the recognition dictionary. In a similar way,
In order to make the dictionary size as small as possible, instead of providing one reference pattern for each misread pattern, a method of adding an average of a plurality of misread patterns has also been proposed [Ide, Wakahara: "Character Recognition" Study of Sub-Category Generation Method in IEICE ", Proc. Of the IEICE Autumn National Convention 1990, D-348, p.6-350 (1990)].

2-. In order to faithfully represent the pattern distribution of each category, this is a method of clustering learning patterns for each category, taking an arithmetic mean of patterns in each group, and creating a recognition dictionary using the patterns as a reference pattern. A dictionary creation method using the Ward method or the like has been proposed as a clustering method [Okura, Shiono:
"Study on Dictionary Pattern Creation by Clustering Multiple Dictionaries Similarity Method in Category", IEICE D-II, Vol.J72-DI
I, No. 4, pp. 499-506 (1989)].

2-. This is a method of optimizing the recognition dictionary by moving the reference pattern so as to minimize erroneous recognition after creating the recognition dictionary by the method 2-. As a typical optimization method, if a recognition pattern is recognized by this recognition dictionary and it cannot be read correctly, the closest reference pattern in the category that should be read correctly is brought close to the learning pattern and read incorrectly. An LVQ method (learning vector quantization method) has been proposed which creates a recognition dictionary so as to minimize misreading by repeatedly performing a process of keeping a category reference pattern away [T. Kohonen, GB
arna and R. Chrisley: “Statistical pattern recog
nition with newral networks ”, benchmarking studie
s, IEEE, Proc. of ICNN, Vol. 1, pp. 61-68 (1988)].

[0009]

In the conventional method of simply using an average pattern as a reference pattern as described in 1 above, there is an advantage that the dictionary size is compact, but when the pattern distribution between categories is different, the dictionary size is small. There is a problem that recognition accuracy is significantly reduced. That point, 1-
The method of (1) is excellent in that the dictionary is optimized so as to increase recognition accuracy in consideration of the difference in pattern distribution between categories while maintaining the advantage that the dictionary size is small. However, when the pattern distribution is complicated and intermingled between categories, or when the pattern distribution is not continuous but separated, there is a problem that a single reference pattern cannot cope with the pattern distribution, which also lowers the recognition accuracy. is there.

On the other hand, the second method can have a plurality of reference patterns for each category, and thus can solve the problems as in the first method. However, in the method (2), basically, a reference pattern must be added for each misreading pattern, and a very large number of reference patterns are required to obtain sufficient recognition accuracy. In addition, there is a problem that the size of the recognition dictionary created for the same reason becomes very large.
In the method (2), although a complex pattern distribution can be expressed by a plurality of reference patterns using a clustering method, the recognition accuracy is not necessarily improved because the pattern distribution between categories is not considered. Does not lead to In addition, although the number of reference patterns should be determined in consideration of the pattern distribution within and between categories, an appropriate number of reference patterns must be set in advance for each category. The only way to obtain it is to determine it by trial and error, so that there is a problem that it takes an enormous amount of time to create a dictionary.
In the method (2), the position of the reference pattern is optimized so that the misread pattern can be correctly recognized, and the method is conscious of the pattern distribution between the categories. Is obtained by the method 2-, and the number of reference patterns is invariant in the algorithm. Therefore, there is a problem in setting the number of reference patterns as in the method 2-.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to provide a technique capable of efficiently creating a recognition dictionary with high recognition accuracy.

Another object of the present invention is to automatically optimize the position of the reference pattern and set the number of reference patterns for each category while taking into account not only the pattern distribution within the category but also the pattern distribution between the categories. It is to provide various technologies.

The above and other objects and novel features of the present invention will become apparent from the description of the present specification and the accompanying drawings.

[0014]

In order to achieve the above object, the means (1) of the present invention is a recognition method which is a set of reference patterns of each category used when recognizing patterns such as characters and voices. a use dictionary creation method, a first step of creating an initial recognition dictionary having one of the reference pattern in advance for each category, a second step of updating the recognition dictionary by performing the movement of references pattern When the function indicating the shortage of reference patterns exceeds the specified value,
And a third step of newly adding the moved reference pattern to the recognition dictionary, and repeating the second step and the third step until erroneous reading of the learning pattern is eliminated or the set number of times of learning is reached. The most important feature.

The above means (2) uses the number of movements of the reference pattern as a function indicating the shortage of the reference pattern in the method of creating a dictionary for pattern recognition, and the number of movements reaches a predetermined number of times . Then given
It is characterized in that a reference pattern that has been moved a number of times is newly added to the recognition dictionary.

[0016]

According to the above-mentioned means (1), learning is performed by using a recognition dictionary having one reference pattern for each category as an initial dictionary, and reference patterns are assigned only to portions (categories) where reference patterns are insufficient. By adding
An appropriate number of reference patterns for each category can be automatically set in consideration of the pattern distribution within and between categories. This means that it is possible to efficiently create a recognition dictionary, and it is also possible to express a complicated distribution in which patterns in a category are separated. In addition, the position of the reference pattern is optimized to a position where misrecognition is minimized as much as possible.
Significant improvement in recognition accuracy can also be expected.

According to the means (2), in addition to the above-described operation, a category having a shortage of reference patterns is considered to have a large number of movements of the reference pattern of the category. By using the number as the evaluation function, a portion (category) where the reference pattern is insufficient can be set accurately and automatically. In addition, the number of movements of the reference pattern need only be the number of movements each time the movement is performed. Therefore, no trouble is required. Furthermore, whether or not the number of reference patterns is insufficient is determined by the number of movements ≧ k (k is predetermined. It is only necessary to evaluate a reference value (a reference value determined to be insufficient for the number of reference patterns), so that it is possible to shorten the dictionary creation time.

[0018]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a pattern recognition dictionary creation method according to the present invention. FIG. 2 is a flowchart showing a processing procedure of the pattern recognition dictionary creation method of the embodiment. FIG. 3B is a block diagram illustrating a schematic hardware configuration of a pattern information processing apparatus to which the pattern recognition dictionary creation method according to the present embodiment is applied. FIG. 3B is a diagram illustrating a pattern recognition process using the pattern recognition dictionary creation method according to the embodiment. FIG. 3 is a block diagram illustrating a functional configuration.

As shown in FIG. 3A, the pattern information processing apparatus to which the pattern recognition dictionary creating method of the present embodiment is applied includes a pattern input apparatus 10, an information processing / control apparatus 20,
It comprises a file storage device 30 and an output device 40 such as a display device, a printer, and a disk. For example, a computer is used as the information processing / control device 20, and a central processing unit (CPU) 20
1. It is composed of a main memory 202. This information processing
As shown in FIG. 3B, the functional configuration in the case of performing the pattern recognition process using the control device 20 includes a preprocessing unit 203, a feature extraction unit 204, a dictionary creation unit 205, and a recognition unit 206.

As shown in FIG. 1, the functional configuration of the dictionary creating unit 205 of the present embodiment is as follows: an initial dictionary creating unit 1, a learning pattern selecting unit 2, a pattern recognizing unit 3, a reference pattern movement determining unit 4, Reference pattern movement execution unit 5, reference pattern addition determination unit 6, reference pattern addition execution unit 7, control unit 8
It is composed of The control unit 8 controls each unit, and performs processing end determination and data output.

Next, a method for creating a dictionary for pattern recognition according to the present embodiment will be described with reference to FIGS.

There are several possible methods for optimizing the recognition dictionary. Here, the method used in the LVQ method is used. The LVQ method is a method used to create a multi-template dictionary in pattern recognition. After an initial dictionary is created by a clustering method, a learning pattern is recognized and cannot be correctly recognized. Means that the closest reference pattern among the reference patterns of the category to be correctly recognized is brought closer and the reference pattern of the erroneously recognized category is moved away when the reference pattern exists near the discrimination boundary between the reference pattern and the misreading side reference pattern. Is performed repeatedly to optimize the position of the recognition dictionary. The specific movement of the reference pattern is performed by the following equation. Since this equation is a one-dimensional moving equation of the feature vector expressing the pattern, it is necessary to execute this moving equation for each feature axis. [Regular reading side] m (t) = m (t-1) + α (t) * (x−m (t−1)) ··· [Incorrect reading side] m (t) = m (t−1) − α (t) * (x−m (t−1)) m (t): Reference pattern at time t m (t−1): Reference pattern at time t−1 x: Learning pattern α (t) : Decreasing function [0 <α (t) <1] FIG. 2 shows a processing flow of the present embodiment under the above conditions, and the algorithm thereof will be described below.

In the initial dictionary creating section 1, 1 for each category
Create a number of reference patterns (101), which is the initial recognition dictionary M ₀ (102). The learning pattern selection unit 2 selects one learning pattern x for learning the recognition dictionary M (103), and recognizes the selected learning pattern x in the pattern recognition unit 3 (104). It is checked whether or not the recognition of the learning pattern x has been correctly recognized (1).
05) If it is correctly recognized (YES), it is checked whether the processing is completed (112). If the processing is completed, the recognition dictionary M is registered and the processing is ended. If it is not correctly recognized in the step 105 (NO), the following processing is performed for each of the reference patterns m ₁ , m ₂ ... Mi. That is, the reference pattern movement determination unit 4 determines whether the condition for moving the reference pattern is satisfied based on the recognition result of the selected learning pattern x and the positional relationship with the pattern (106).
If the condition is satisfied, the reference pattern movement execution unit 5 executes the movement of the reference pattern. Further, the reference pattern movement execution unit 5 stores the number of times of movement of each reference pattern. The reference pattern addition determining unit 6 determines whether the number of movements of the reference pattern mi moved by the reference pattern movement executing unit 5 has reached k times (107), and the condition is satisfied. The reference pattern is added by the reference pattern addition execution unit 7 (108). For those which do not satisfy the conditions, the reference pattern mi is only moved by the reference pattern movement executing unit 5 (10).
9), the number of movements of the reference pattern mi is incremented (110). Next, it is checked whether or not the reference pattern mi to be moved does not exist (111). If it does not exist (YES), it is checked whether or not the processing is completed (112). Register and exit. If there is (NO), the flow returns to step 106 to eliminate the erroneous reading of the learning pattern or to set the learning number k (k
Is repeated until a predetermined reference value (a reference value for determining that the number of reference patterns is insufficient) is reached.

The control unit 8 terminates learning of the recognition dictionary as to whether the number of times of learning has reached k or whether there has been no misreading of the learning pattern, and registers the recognition dictionary at that time.

Next, a specific example of handwritten character recognition according to this embodiment will be described.

The category is “1
There are 0 types, and the stroke feature integration method [Akamatsu, Kawatani, Shio, Iida: "Hand-reading kanji character reader", Research and practical report, Vol.36, No.4 (1987). . FIG. 4 shows the recognition accuracy of the learning pattern changed as the learning of the recognition dictionary is performed. As the learning progresses, the number of misreads decreases, and it can be seen that all the learning patterns are finally correctly recognized. FIG. 5 shows the number of reference patterns in each category in the recognition dictionary at the end of learning, and it can be seen that an appropriate number of reference patterns is automatically set for each category.

Although the present invention has been described in detail with reference to the embodiments, it is needless to say that the present invention is not limited to the above-described embodiments and can be variously modified without departing from the scope of the invention. .

[0029]

As described above, according to the present invention, in a pattern recognition method for recognizing a category based on which reference pattern the input pattern is close to in the pattern space, the position of the reference pattern is optimized. In addition, since reference patterns are added depending on the category, a recognition dictionary with high recognition accuracy can be created.
In addition, since the reference pattern is added automatically by evaluating a function indicating the lack of the reference pattern, it is not necessary to determine the number of reference patterns by trial and error unlike the conventional multi-template dictionary creation method. Creation time can be greatly reduced.

Further, by using the number of movements of each reference pattern up to that point as a function representing the lack of the reference pattern, a portion (category) where the reference pattern is lacking can be accurately and automatically found at a high speed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a pattern recognition dictionary creation method according to the present invention;

FIG. 2 is a flowchart showing a processing procedure of a pattern recognition dictionary creation method according to the embodiment;

FIG. 3A is a block diagram showing a schematic configuration of a pattern information processing apparatus to which the pattern recognition dictionary creation method of the present embodiment is applied;

FIG. 3B is a block diagram illustrating a functional configuration of a pattern recognition process using the pattern recognition dictionary creation method according to the embodiment;

FIG. 4 is a graph showing learning pattern recognition accuracy according to the embodiment;

FIG. 5 is a graph showing the number of templates for each category according to the embodiment.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Initial dictionary preparation part, 2 ... Learning pattern selection part, 3 ... Pattern recognition part, 4 ... Reference pattern movement determination part, 5 ... Reference pattern movement execution part, 6 ... Reference pattern addition determination part, 7
... Reference pattern addition execution unit, 8 control unit, 10 pattern input device, 20 information processing / control device, 30 file storage device, 40 output device, 201 central processing unit (CPU), 202 main Memory, 203: pre-processing unit, 2
04: feature extraction unit, 205: dictionary creation unit, 206: recognition unit.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-4-205389 (JP, A) JP-A-3-42767 (JP, A) JP-A-3-18983 (JP, A) IEICE Technology Research Report Vol. 91 No. 156 pp.63-68 “Study of a method for creating a handwritten kanji recognition dictionary using LV Q”, et al., IEICE Transactions on Electronics, Information and Communication Engineers D- ▲ II ▼ Vol. 73 No. 8 pp. 1186-1191 “Neural Network Model Adapting to Severe Environmental Changes” Okamoto (58) Fields investigated (Int. Cl. ⁷ , DB name) G06T 7/00 G06K 9/68 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A method for creating a recognition dictionary, which is a set of reference patterns for each category, used for recognizing patterns such as characters and voices.
A first step of creating an initial recognition dictionary having reference patterns, a second step of updating the recognition dictionary by moving the reference patterns, and a function representing a shortage of reference patterns being equal to or greater than a prescribed value. References made when moved
A third step of newly adding a pattern to the recognition dictionary, wherein the second step and the third step are repeated until erroneous reading of the learning pattern is eliminated or the set number of times of learning is reached. Dictionary creation method. 2. The pattern recognition dictionary creating method according to claim 1, wherein the number of movements of the reference pattern is used as a function indicating the shortage of the reference pattern, and the number of movements reaches a predetermined number . If the predetermined times
A method for creating a pattern recognition dictionary, characterized by newly adding a reference pattern that has been moved several minutes to a recognition dictionary.