JPH05233880A - Dictionary preparing method for pattern recognition - Google Patents

Abstract

PURPOSE:To prepare a recognizing dictionary having the high rate of recognition by optimizing weight for correcting position relation between a reference pattern and a recognizing object pattern concerning a pattern recognizing method to recognize a category corresponding to which reference pattern an inputted pattern is adjacent to on a pattern space. CONSTITUTION:An initial dictionary preparation part 1 calculates the reference pattern as an arithmetic mean for each category and calculates the weight as the inverse of standard deviation. This is defined as the initial dictionary for recognition. A learning pattern selection part 2 selects one of patterns for learning the dictionary for recognition, and the selected pattern is recognized by a pattern recognition part 3. A weight correction judge part 4 judges whether the weight is corrected or not corresponding to the recognized result of the selected pattern and concerning the weight satisfying conditions, the weight of the dictionary is changed by a weight correction execution part 5. Further, strength for mutual operation between the learning pattern and the reference pattern is calculated, and the weight is changed so as to entirely reduce the mutual operation.

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 for efficiently creating a pattern recognition dictionary with high recognition accuracy. ..

[0002]

2. Description of the Related Art In pattern recognition of characters and voice,
A reference pattern that is a standard pattern for each category and a weight vector that corrects the positional relationship between the reference pattern and the recognition target pattern according to the distribution shape of the category are prepared in advance as a recognition dictionary. After calculating the similarity with the reference pattern of, using some evaluation function, the category with the highest similarity is output as the recognition result.

Usually, the reference pattern is given as the average value of the patterns to be learned, and the weight is given as the reciprocal of the standard deviation.

As a method of sequential learning, there has been proposed a method of deciding weights by correcting errors during learning, such as the potential function method.

The above-mentioned prior art is widely used because of its simple algorithm and, in many cases, its effect can be expected.

[0006]

However, the reciprocal of the standard deviation is an optimum weight only when the Gauss distribution is assumed, and the reciprocal of the standard deviation has a complicated distribution as seen in an actual character pattern. On the other hand, it is not always the best dictionary.

[0007] Further, in many cases, a dictionary learning method such as the potential function method that assumes linear separability between pattern sets, and the linear separability is generally detected. Since there is no method, the effectiveness of the method is evaluated only by the success or failure of creating the dictionary, which is not practical.

The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to automatically optimize the weight while considering not only the category distribution but also the pattern distribution between categories. It is to provide the technology that can be done in.

Another object of the present invention is to provide a technique capable of efficiently creating a pattern recognition dictionary having a high recognition rate.

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

[0011]

In order to achieve the above object, the means (1) of the present invention is used to correct the positional relationship according to the standard pattern of each category and the distribution shape of each category. A method for creating a pattern recognition dictionary having a set of weights, wherein an arithmetic mean and a standard deviation of each feature are obtained in advance for each of a plurality of learning patterns for each category and each feature axis. The average is used as the reference pattern, and an initial pattern recognition dictionary in which the reciprocal of the standard deviation is used as the weight is created, and the weight is moved so that misrecognition is reduced while repeatedly giving the learning pattern. The feature is that the recognition dictionary is updated.

According to a second aspect of the present invention, in the method for creating a pattern recognition dictionary according to the first aspect, regarding a pattern that causes an erroneous recognition, the reference pattern is used as a method for updating the pattern recognition dictionary. , The weight is changed so that the interaction becomes small, and when the discriminant hyperplane created by the discriminant function becomes the equilibrium point of the interaction, the weight is converged as a correct answer. It is characterized by

That is ,. First, create a reference pattern that is the arithmetic mean of several patterns and weights that are the reciprocal of the standard deviation for each category, and register these as the initial dictionary. Next, focusing on the erroneous recognition of the learning pattern, we assume an interaction that is inversely proportional to the distance between the learning pattern and the reference pattern, and change the weights so that the interaction decreases. By finding the equilibrium point of the interaction from the distribution, the weight learning is terminated at the optimum value even if the linear separability is not satisfied.

[0014]

According to the above-mentioned means, the optimum value is guaranteed for the pattern set according to the Gauss distribution by using the pattern recognition dictionary whose weight is the reciprocal of the standard deviation, By focusing on the erroneous recognition pattern and changing the weight in the direction in which the interaction from the reference pattern decreases, a more accurate discrimination boundary can be obtained even for a pattern distribution deviating from the Gaussian distribution.

Further, by using the equilibrium point of the interaction as a learning end condition, the pattern recognition dictionary creating device creates a dictionary in which the learning pattern is excessively specialized in order to forcibly relieve a region that cannot be actually separated. Since it works so as to prevent such a situation, it is possible to create a pattern recognition dictionary having a high recognition rate with a relatively small processing amount and a simple algorithm.

[0016]

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

FIG. 1 is a block diagram showing the functional arrangement of an embodiment of a pattern recognition dictionary creating apparatus for carrying out the pattern recognition dictionary creating method of the present invention. 1 is an initial dictionary creating unit, and 2 is a learning pattern. The selection unit 3, the pattern recognition unit 4
Is a weight correction determination unit, 5 is a weight correction execution unit, and 6 is an end determination unit.

The initial dictionary creating section 1 obtains a reference pattern as an arithmetic mean for each category, and obtains a weight as an inverse of the standard deviation. This is the initial recognition dictionary. The learning pattern selection unit 2 selects one pattern for learning the recognition dictionary and recognizes the pattern selected by the pattern recognition unit 3.

The weight correction judging section 4 judges whether or not the weight is changed according to the recognition result of the selected pattern, and if the conditions are satisfied, the weight correction executing section 5 is executed.
At, the weight of the dictionary is changed. Further, the weight correction execution unit 5 calculates the strength of the interaction acting between the learning pattern and the reference pattern, and changes the weight so that the interaction is reduced as a whole.

The termination judgment unit 6 judges whether the discrimination boundary surface formed by the weight and the reference pattern has reached the equilibrium point of the interaction between the pattern sets or whether the misreading of the learning pattern has disappeared and the conditions are satisfied. In this case, the learning of the recognition dictionary is finished and the recognition dictionary at that time is registered. If it is determined that the convergence is too slow, the learning is terminated when the number of learning reaches a predetermined number.

There are several possible interactions assumed for optimizing the weight of the recognition dictionary, but here, an embodiment using an interaction that satisfies Gauss's law will be described.

Gauss's law is a force that is inevitably introduced only from geometrical features of space, and in natural phenomena it is an electric field,
It exists universally as an interaction such as the gravitational field.

Further, from the viewpoint of identification ,. The larger the distance, the smaller the magnitude of the interaction ,. This is in line with our intuition in that the interaction decreases as the number of dimensions increases.

In a general N-dimensional feature space, Gauss's law is

[0025]

[Equation 1]

Can be written as Here, E _i ″ (hereinafter “” indicates a superscript arrow → of a letter and means a vector) is the strength of interaction, s ″ is an area element, and f _i (x ”)
Is the probability density function of the distribution, r is the distance from the center of the distribution,
x ″ is the coordinate of the pattern to be interacted with, ρ is a parameter that determines the strength of the interaction, and v is a volume element of the feature space.

Here, regarding the two distributions, which are considered only in the two-category problem, it can be expected that the equilibrium point of the exerted force will divide the distribution well into two. That is,

[0028]

[Equation 2]

Therefore,

[0030]

[Equation 3]

The condition of the equilibrium point can be defined as Here, assuming that the shapes of the distribution functions of the two categories are the same, equation (3) becomes

[0032]

[Equation 4]

It becomes This is nothing but the condition for Bayes maximum likelihood estimation. That is, in at least the two-category problem, Gaussian (Ga
It can be seen that finding the equilibrium point of the interaction satisfying the law of Uss) is equivalent to Bayes' estimation.
In the formulas (2), (3), and (4), “′”
Means another distribution or place.

In the actual movement of the weights, an equation in which the pattern moves in a direction that reduces the interaction obtained by differentiating the interaction given by the equation (1) is obtained, and this solution is obtained. The dictionary is updated so that the learning pattern is not erroneously recognized by calculating a weight that gives the same coordinates as.

Under the above conditions, an operation flow of one embodiment of the pattern recognition dictionary creating method of the present invention is shown in FIG. 2, and its algorithm will be described below.

First, a recognition dictionary M0 having one reference pattern, which is an arithmetic mean of several patterns, and one weight vector, which is the reciprocal of the standard deviation, is created for each category (step 201). Is registered as the recognition dictionary M (step 202). Next, set the learning pattern to 1
An individual is selected (step 203), and this pattern is recognized using the recognition dictionary M (step 204). Then, it is determined whether the recognition result is a misread or a correct reading (step 205).
Return to 03. If the reference pattern and weight are not correctly recognized, the following process is performed.

First, for the reference pattern on the correct category side, the interaction such that the interaction with the learning pattern takes a negative value and the differential value for each feature axis are calculated, and the learning pattern becomes the correct reference pattern. The amount of movement relative to that is calculated (step 206), and the weight is modified so that the learning pattern feels the interaction having the same value as the coordinate after movement (step 207). For the weight on the misread side, the interaction with the learning pattern takes a positive value and the differential value for each feature axis is calculated, and the learning pattern moves to the misread reference pattern. Is calculated, and the weight is corrected so that the learning pattern feels the interaction having the same value as the coordinate after movement (step 207).

After the above processing is performed, it is judged whether the learning pattern has been erroneously read, whether the discrimination hyperplane has reached the equilibrium of interaction, or has reached the set number of times of learning (step 208), and the condition is set. If satisfied, the learning ends, the recognition dictionary at that time is registered, and the processing ends.

Although the present invention has been specifically described based on the embodiments, it is needless to say that the present invention is not limited to the above embodiments and various modifications can be made without departing from the scope of the invention. Absent.

[0040]

As described above, according to the present invention, the positions of the reference pattern and the recognition target pattern are recognized in the pattern recognition method for recognizing the category according to which reference pattern the input pattern is closer to in the pattern space. A recognition dictionary with a high recognition rate can be created by optimizing the weights that correct the relationship.

Further, since the learning can be surely succeeded by evaluating the equilibrium point of the assumed interaction between the patterns, the linear separability is evaluated by the learning itself like the conventional sequential learning method. It is possible to eliminate the risk of circular reasoning such as performing, and to create an efficient and reliable dictionary.

[Brief description of drawings]

FIG. 1 is a block diagram showing a functional configuration of an embodiment of a pattern recognition dictionary creating apparatus that implements a pattern recognition dictionary creating method of the present invention;

FIG. 2 is a flowchart of an algorithm executed by the pattern recognition dictionary creation device of the present embodiment.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Initial dictionary creation part, 2 ... Learning pattern selection part, 3 ... Pattern recognition part, 4 ... Weight correction determination part, 5 ... Heavy correction execution part, 6 ... End determination part.

Claims (2)

[Claims] 1. A standard pattern of each category (hereinafter,
(Referred to as reference pattern) and a pattern recognition dictionary having a weight set used to correct the positional relationship according to the distribution shape of each category. For each of them, an arithmetic mean and a standard deviation of each characteristic are obtained, the arithmetic mean is used as the reference pattern, and an initial pattern recognition dictionary is created in which the reciprocal of the standard deviation is weighted, and the learning pattern is created. The method for creating a pattern recognition dictionary is characterized in that the pattern recognition dictionary is updated by moving the weights so as to reduce erroneous recognition while repeatedly giving. 2. The pattern recognition dictionary creation method according to claim 1, wherein a pattern causing an erroneous recognition is:
As a method of updating the pattern recognition dictionary, an interaction from the reference pattern is assumed, weights are changed so that the interaction becomes small, and an identification hyperplane created by an identification function becomes an equilibrium point of the interaction. A method for creating a dictionary for pattern recognition, which is characterized by using a method in which the weight is converged as a correct answer.