JP2937235B2 - Recognition dictionary learning method and apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a recognition dictionary learning method and apparatus, and more particularly to a recognition dictionary learning method used in a pattern recognition apparatus based on a matrix matching method and a recognition dictionary learning apparatus to which the method is applied.

[0002]

2. Description of the Related Art In a matrix matching method, a plurality of binary mask patterns (templates) are prepared as a recognition dictionary, and an input pattern is determined based on the degree of overlap with a binary input pattern (for example, Kosho 60-2
714, "Printed character recognition device". Taking character recognition as an example, as shown in FIG. 5, a plurality of dictionary patterns each comprising a set of a template called a positive feature and a template called a negative feature for a binary input pattern are prepared for each category. The number of pixels where the black pixel of the feature and the white pixel of the input pattern match is defined as the positive feature mismatch degree, and the number of pixels where the black pixel of the negative feature matches the black pixel of the input pattern is defined as the negative feature mismatch degree. The category to which the dictionary pattern having the smallest sum of the feature mismatch degree and the negative feature mismatch degree belongs is set as the recognition result.

It is known that in the case of a dictionary pattern in which the number of pixels is the same as the input pattern and the pixel value is a real number, a template can be automatically created by a learning method called learning vector quantization (LVQ) by Kohonen. [T. Kohonen: "The Neural P
Honetic Typewriter ”, IEEE
Omputer, Vol. 21, No. 3, pp. 11
-22], this LVQ cannot be used to create a binary template used in the matrix matching method.

Therefore, a method of creating a dictionary pattern by the procedure shown in FIG. 6 is used as a known technique. That is, a contraction processing unit 302 is applied to the binary original image 301.
Performs the contraction processing of the black pixel by using the template 30 of the positive feature.
3 and the expansion processing unit 30
After performing the black pixel expansion processing in step 4, the black-and-white inversion processing is performed in the inversion processing unit 305 to create a template 306 having a negative feature.

[0005]

However, in the above-described conventional method for creating a dictionary pattern, no consideration is given to patterns belonging to other categories, and dictionary patterns are created for each category. Many,
There was a problem that a human had to correct the dictionary pattern by trial and error. This will be briefly described below.

An area in which both the positive feature template and the negative feature template are white pixels will be referred to as a dead area. In the dead area, the number of mismatches is always 0 irrespective of whether the input pixel is black or white. Therefore, by expanding the dead area, the degree of mismatch can be made close to 0 even for a slightly deformed input pattern. However, if the dead area is excessively widened, the degree of inconsistency becomes small even for patterns belonging to other categories, so that recognition errors increase. In order to increase recognition accuracy, it is necessary to adjust this dead area so as to reduce recognition errors.However, the conventional method of creating a recognition dictionary based on dilation / contraction processing does not consider patterns belonging to other categories. There were many recognition errors, and humans had to correct the dictionary pattern by trial and error.

An object of the present invention is to eliminate the problem that the recognition accuracy is not improved with a dictionary pattern created by a conventional recognition dictionary creation method based on expansion / shrinkage processing, and it is necessary to correct the dictionary pattern by human trial and error. Another object of the present invention is to provide a recognition dictionary learning method for automatically generating a high-precision recognition dictionary.

It is another object of the present invention to provide a recognition dictionary learning apparatus to which the above-described recognition dictionary learning method is applied.

[0009]

A recognition dictionary learning method according to a first aspect of the present invention is a method for learning a recognition dictionary used in a pattern recognition apparatus based on a matrix matching method, wherein initialization of a dictionary pattern is performed. An initial value setting step to be performed, an input step of selecting one input pattern from a group of input patterns prepared for learning, and a degree of mismatch between the input pattern and a dictionary pattern group belonging to the same category as the input pattern. Is the first dictionary pattern and the first degree of inconsistency thereof, and a dictionary pattern group belonging to a category different from the input pattern and the second dictionary pattern and the second dictionary pattern having the least degree of inconsistency with the input pattern A dictionary search step for obtaining a second degree of inconsistency; a correction frequency calculation step for calculating a frequency for correcting the dictionary pattern; A dictionary correction step of correcting the dead area of the dictionary pattern to be large and a dead area of the second dictionary pattern to be small, and an end determination step of determining the end of learning of the dictionary pattern. Features.

A recognition dictionary learning apparatus according to a second aspect of the present invention is a recognition dictionary learning apparatus to which a recognition dictionary learning method used in a pattern recognition apparatus based on a matrix matching method is applied. Means for selecting one input pattern from a group of input patterns prepared for learning, and a mismatch between the input pattern and a dictionary pattern group belonging to the same category as the input pattern. A first dictionary pattern having a minimum degree of mismatch and a first mismatch degree thereof, and a second dictionary pattern having a minimum degree of mismatch with the input pattern from a dictionary pattern group belonging to a category different from the input pattern; Dictionary search means for obtaining the second degree of inconsistency; correction frequency calculation means for calculating the frequency of correcting the dictionary pattern; Dictionary correction means for correcting the dead area of the first dictionary pattern to be larger and the dead area of the second dictionary pattern to be smaller in accordance with the frequency; and end determination means for determining the end of learning of the dictionary pattern. It is characterized by having.

Further, the recognition dictionary learning method according to the third invention of the present application is the recognition dictionary learning method according to the first invention, wherein the difference between the first degree of mismatch and the second degree of mismatch is smaller. The method further includes the step of calculating a correction frequency so as to increase the frequency of correcting the dictionary pattern.

Furthermore, a recognition dictionary learning device according to a fourth invention of the present application is the recognition dictionary learning device according to the second invention, wherein the difference between the first degree of mismatch and the second degree of mismatch is small. The correction frequency calculating means for calculating the correction frequency so as to increase the frequency of correcting the dictionary pattern.

[0013]

Next, an embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a basic configuration of a recognition dictionary learning apparatus to which a recognition dictionary learning method according to one embodiment of the present invention is applied. The recognition dictionary learning apparatus according to the present embodiment includes an initial value setting means (step) 101 for initial setting of a dictionary pattern, and an input means for selecting one input pattern X from an input pattern group prepared for learning. (Step) 102, a first dictionary pattern W1 and a first dictionary pattern W1 having a minimum degree of inconsistency with the input pattern X from a dictionary pattern group belonging to the same category as the input pattern X;
, A second dictionary pattern W2 having the minimum degree of mismatch with the input pattern X and a second degree of mismatch D2 thereof from a dictionary pattern group belonging to a category different from that of the input pattern X. Dictionary search means (stage)
103, a correction frequency calculating means (step) 105 for calculating a frequency of correcting the dictionary pattern, and a dead area of the first dictionary pattern W1 is increased according to the correction frequency.
And dictionary correction means 104 (step) for correcting the dead area of the second dictionary pattern W2 so as to be small, and end determination means (step) 10 for determining the end of learning of the dictionary pattern.
6 is comprised. When the end determination unit 106 determines that the learning of the dictionary pattern has not been completed, the process returns to the input unit 102.

FIG. 2 is a flowchart showing the processing of the dictionary correction means 104 and correction frequency calculation means 105 enclosed by broken lines in FIG.

FIG. 3 is a diagram showing pixel values of an input pattern and a dictionary pattern used in the recognition dictionary learning method and apparatus according to the present embodiment. That is, if the total number of pixels of the input pattern is N, and if the pixels of the input pattern are ○, −1, and if the pixel is ●, +1, then the input pattern is (−
It can be regarded as an N-dimensional array having (1, 1) as an element value, and this is expressed as X [i]. Where i is 1 to N
The value up to. For dictionary patterns, (Positive feature,
If the set of (negative features) is (○, ●), −1, if (○, ○) is 0, and if (●, ○) is +1, the dictionary pattern is {-1, 0, 1}. Can be regarded as an N-dimensional array in which W1 [i] or W2
Notated as [i].

In the dictionary pattern group belonging to the same category as the input pattern X, the dictionary pattern having the lowest degree of mismatch is W1 and the degree of mismatch is D1. Similarly, among the dictionary pattern groups belonging to a category different from the input pattern X, the dictionary pattern with the smallest degree of mismatch is W2, and the degree of mismatch is D2. In the matrix matching method, the category to which the dictionary pattern with the lowest degree of mismatch belongs is the recognition result. Therefore, if D1 <D2, correct recognition is performed, and if D1 ≧ D2, erroneous recognition is determined.

This will be described in detail with reference to FIG. Focusing on pixels having an input pattern, there are a total of six combinations of pixel values of the positive and negative features of the dictionary pattern with respect to the pixel values of the input pattern, as shown in FIG. here,
● and ○ represent the case where the pixel value is black and white, respectively, and it is assumed that a combination in which both the positive feature and the negative feature are black does not occur. The degree of inconsistency for these combinations is 1 only when the input is 特 徴 and the positive feature is ●, and the input is ● and the negative feature is ●.
Becomes

The dictionary correction means 104 includes a dictionary pattern W1
Is modified as follows. The pixel value X [i] of the input pattern X is ○ and the positive feature of the dictionary pattern W1 is ●
In the case of, the positive feature of the dictionary pattern W1 is modified from ● to ○, and when the pixel value X [i] of the input pattern X is ● and the negative feature of the dictionary pattern W1 is ●, the negative feature of the dictionary pattern W1 is changed from ●. Correct to ○. Portions other than the above, that is, portions indicated by-in FIG. 4 are not changed. Since the degree of mismatch for the corrected pixel decreases from 1 to 0, the value of the degree of mismatch D1 decreases.

The dictionary correction means 104 corrects the dictionary pattern W2 as follows. When the pixel value X [i] of the input pattern X is ○ and the positive feature of the dictionary pattern W2 and the negative feature of the dictionary pattern W2 are both ○, the positive feature of the dictionary pattern W2 is corrected from ○ to ●, If the pixel value X [i] is ● and both the positive feature of the dictionary pattern W2 and the negative feature of the dictionary pattern W2 are ○, the negative feature of the dictionary pattern W2 is corrected from ○ to ●.
Portions other than the above, that is, portions indicated by-in FIG. 4 are not changed. The degree of inconsistency for the corrected pixel is 0
From 1 to 1, the value of the degree of mismatch D2 increases.

The dead area is represented by ○ together with positive and negative features.
, The dictionary pattern W1 increases the dead area, and the dictionary pattern W2 decreases the dead area. However, if the above correction is performed for all the pixels of the input pattern X, the dictionary pattern greatly changes for each input pattern X, and there is a problem that erroneous reading for the entire learning data is not necessarily reduced. The above correction is made with a probability of 1 or less. That is, the dictionary correction unit 104 corrects only a part of the pixels satisfying the above-described correction condition by performing the correction with the correction frequency calculated by the correction frequency calculation unit 105.

Next, the operation of the thus configured recognition dictionary learning apparatus according to the present embodiment will be described together with its recognition dictionary learning method.

First, the initial value setting means 101 performs initial setting of a dictionary pattern.

Next, the input means 102 selects one input pattern X from an input pattern group prepared for learning.

Subsequently, the dictionary search means 103 selects a first dictionary pattern which minimizes the degree of mismatch with the input pattern X from a group of dictionary patterns belonging to the same category as the input pattern X selected by the input means 102. W1 and its first inconsistency D1 and a second dictionary pattern W2 and its second inconsistency that minimize the inconsistency between the input pattern X and a dictionary pattern group belonging to a category different from the input pattern X The degree D2 is obtained.

Next, the correction frequency calculating means 105 calculates the quantities a1 and a2 representing the correction frequency by predetermined functions F1 and F2 using the first degree of mismatch D1 and the second degree of mismatch D2, respectively (FIG. 2). Step S201). The quantities a1 and a2 are real numbers from 0 to 1 and, for example, when D1 <D2, a1 = 0.01, a2 = 0,
When D1 ≧ D2, a1 = 0 and a2 = 0.01 are set.

Next, the dictionary correction means 104 sets the counter i
Is set to 1 (step S202).

Subsequently, the dictionary correction means 104 outputs the element value W1 [i] of the first dictionary pattern W1 and the input pattern X
Is compared with the element value X [i] (step S20).
3) If the element value W1 [i] is equal to the element value -X [i], the process proceeds to step S204; otherwise, the process proceeds to step S206.

In step S204, the dictionary correction means 10
4 generates a uniform random number from 0 to 1 and obtains an obtained value r
and is compared with the amount a1, and if the value rand is smaller than the amount a1, the process proceeds to step S205; otherwise, the process proceeds to step S206.

In step S205, the dictionary correction means 10
No. 4 corrects the element value W1 [i] of the first dictionary pattern W1. Specifically, W1 [i] = W1 [i] + X
[I] is calculated.

In step S206, the dictionary correction means 10
In step S4, if the counter i is equal to the number N of pixels of the input pattern X, the process ends; otherwise, the value of the counter i is increased by 1 (step S207), and step S207 is performed.
Return to 203.

On the other hand, in step S208, the dictionary correction means 104 sets the value of the counter i to 1.

Next, the dictionary correction means 104 compares the element value W2 [i] of the second dictionary pattern W2 with 0 (step S209), and when it is equal to 0, proceeds to step S210. Goes to step S212.

In step S210, the dictionary correction means 10
4 generates a uniform random number from 0 to 1 and obtains an obtained value r
and is compared with the quantity a2. If the value rand is smaller than the quantity a2, the process proceeds to step S211. If not, the process proceeds to step S212.

In step S211, the dictionary correction means 10
No. 4 corrects the element value W2 [i] of the second dictionary pattern W2. Specifically, W2 [i] = W2 [i] -X
[I] is calculated.

In step S212, the dictionary correction means 10
4 terminates the processing if the counter i is equal to the number N of pixels of the input pattern X, otherwise, the counter i
Is incremented by 1 (step S213), and step S209
Return to

Regarding the setting of the quantities a1 and a2 representing the correction frequency in step S201, in the recognition dictionary learning method and apparatus according to the first embodiment, the correction frequency calculation means 105 determines that a1 = 0 when D1 <D2, for example. .0
1, when a2 = 0 and D1 ≧ D2, a1 = 0 and a2 =
By setting 0.01, when the input pattern X is correctly recognized (D1 <D2), only the first dictionary pattern W1 is corrected at a low frequency, and when the input pattern X is incorrectly recognized (D1 ≧ D2), the first dictionary pattern W1 is corrected. Only the second dictionary pattern W2 is modified at a low frequency.

Further, regarding the setting of the quantities a1 and a2 representing the correction frequency in step S201, in another numerical example, the correction frequency calculation means 105 determines that a1 = 0.0
1, a2 = 0.01 is set. In this case, both the first dictionary pattern W1 and the second dictionary pattern W2 are corrected irrespective of whether the input pattern X is correct recognition or incorrect recognition.

By the way, it is known that many erroneously recognized input patterns mainly exist near the category boundary.
In the matrix matching method, the inconsistencies D1 and D1
2 is used to determine the input pattern, and the degree of mismatch D1
It is conceivable that the input pattern in which D2 and D2 have close values exists near the category boundary. If the dictionary pattern is corrected with a higher frequency as the difference between the values of the degrees of mismatch D1 and D2 is smaller, a pattern that is likely to be erroneously recognized near the category boundary is learned with a higher frequency, and misreading is further reduced. You.

Therefore, regarding the setting of the quantities a1 and a2 representing the correction frequency in step S201, in another numerical example, the correction frequency calculating means 105 determines that a1 = max (0, 0.01 × (0, 0.01 × (D1 <D2)) 1- | D1-D
2 | / 10)), when a2 = 0 and D1 ≧ D2, a1
= 0, a2 = max (0, 0.01 × (1- | D1-D
2 | / 10)). Here, | x | represents the absolute value of x, and max (a, b) is a function that outputs the larger value of a and b. In this case, when the input pattern X is correctly recognized (D1 <D2), only the first dictionary pattern W1 is corrected more frequently as the difference between the degrees of mismatch D1 and D2 is smaller, and the input pattern X is incorrectly recognized. When (D1 ≧ D2)
Include only the second dictionary pattern W2 with inconsistencies D1 and D2.
The smaller the difference, the higher the frequency of correction.

Further, regarding the setting of the quantities a1 and a2 representing the correction frequency in step S201, in another numerical example, the correction frequency calculation means 105 determines, for example, that a1 = max
(0,0.01 × (1- | D1-D2 | / 10)), a
2 = a1 is set. In this case, regardless of whether the input pattern X is correct recognition or erroneous recognition, the mismatch degree D1
The smaller the difference between D1 and D2, the more frequently both the first dictionary pattern W1 and the second dictionary pattern W2 are corrected.

In the above embodiment, the input pattern X is an array having {-1, 1} as element values, and the dictionary patterns W1 and W2 are arrays having {-1, 0, 1} as element values. However, the present invention is, of course, effective even if a combination of white pixels and black pixels is used.

Also, in the above embodiment, step S2
Although a uniform random number is used for calculating the value rand in steps 04 and S210, an arbitrary probability distribution may be used. Therefore, the present invention is not limited to the above embodiment.

[0044]

As is apparent from the above description, according to the recognition dictionary learning method of the present invention, an initial value setting step for initial setting of a dictionary pattern and one of input pattern groups prepared for learning are performed. An input step of selecting one input pattern; a first dictionary pattern having a minimum degree of inconsistency with the input pattern from a dictionary pattern group belonging to the same category as the input pattern; and a first degree of inconsistency thereof; A dictionary search step of obtaining a second dictionary pattern that minimizes the degree of inconsistency with the input pattern from a group of dictionary patterns belonging to a category different from the input pattern and a frequency of correcting the dictionary pattern And calculating the dead area of the first dictionary pattern so that the dead area of the first dictionary pattern is increased in accordance with the correction frequency. By including a dictionary correction step to correct the size of the dictionary pattern and an end determination step to determine the end of the learning of the dictionary pattern, the insensitive area of the dictionary pattern is automatically set. Thus, there is an effect that a high-precision recognition dictionary that does not need to correct the dictionary pattern can be automatically generated.

Further, according to the recognition dictionary learning apparatus of the present invention, initial value setting means for initial setting of a dictionary pattern,
An input unit for selecting one input pattern from a group of input patterns prepared for learning; and a first unit for minimizing the degree of mismatch with the input pattern from a dictionary pattern group belonging to the same category as the input pattern. A dictionary pattern and a first degree of inconsistency thereof, and a second dictionary pattern and a second dictionary pattern thereof in which the degree of inconsistency between the input pattern and a dictionary pattern belonging to a category different from the input pattern is minimized.
Search means for calculating the degree of inconsistency of the first dictionary pattern, correction frequency calculation means for calculating the frequency of correcting the dictionary pattern, and a second dictionary pattern in which the dead area of the first dictionary pattern is increased according to the correction frequency. By providing a dictionary correction means for correcting the dead area of the dictionary to be smaller and an end determination means for determining the end of learning of the dictionary pattern, the dead area of the dictionary pattern is automatically set. There is an effect that a highly accurate recognition dictionary that does not require a human to correct a dictionary pattern by trial and error can be automatically generated.

[Brief description of the drawings]

FIG. 1 is a block diagram and a flowchart showing a basic configuration and an operation processing procedure of a recognition dictionary learning apparatus to which a recognition dictionary learning method according to an embodiment of the present invention is applied.

FIG. 2 is a flowchart showing processing of a dictionary correction unit and a correction frequency calculation unit in FIG.

FIG. 3 is a diagram illustrating values of input patterns and dictionary patterns used in the recognition dictionary learning method and apparatus according to the present embodiment.

FIG. 4 is a diagram for explaining a dictionary correction method in the recognition dictionary learning method and apparatus according to the present embodiment.

FIG. 5 is a diagram for explaining a matrix matching method.

FIG. 6 is a principle block diagram for explaining a conventional dictionary pattern creation method.

[Explanation of symbols]

 Reference Signs List 101 Initial value setting means (step) 102 Input means (step) 103 Dictionary search means (step) 104 Dictionary correction means (step) 105 Correction frequency calculation means (step) 106 Termination determination means (step)

Claims (4)

(57) [Claims] 1. A method for learning a recognition dictionary used in a pattern recognition apparatus based on a matrix matching method, comprising: an initial value setting step for initial setting of a dictionary pattern; and one of input pattern groups prepared for learning. An input step of selecting an input pattern; and a first step of minimizing a degree of mismatch with the input pattern from a dictionary pattern group belonging to the same category as the input pattern.
And the second dictionary pattern and the second degree of mismatch that minimize the degree of mismatch with the input pattern from a group of dictionary patterns belonging to a category different from the input pattern. , A correction frequency calculation step of calculating a frequency of correcting the dictionary pattern, and a dead area of the first dictionary pattern is increased according to the correction frequency. A recognition dictionary learning method, comprising: a dictionary modification step of modifying the dictionary pattern to be smaller; and an end determination step of determining an end of learning of the dictionary pattern. 2. A recognition dictionary learning apparatus to which a recognition dictionary learning method used in a pattern recognition apparatus based on a matrix matching method is applied, wherein initial value setting means for performing initial setting of a dictionary pattern, and an input prepared for learning. An input unit that selects one input pattern from a group of patterns; and a first unit that minimizes a degree of mismatch with the input pattern from a group of dictionary patterns belonging to the same category as the input pattern.
And the second dictionary pattern and the second degree of mismatch that minimize the degree of mismatch with the input pattern from a group of dictionary patterns belonging to a category different from the input pattern. , A correction frequency calculation means for calculating the frequency of correcting the dictionary pattern, and a dead area of the first dictionary pattern is increased according to the correction frequency. A recognition dictionary learning apparatus, comprising: a dictionary correction unit for correcting a dictionary pattern to be smaller; and an end determination unit for determining the end of learning of a dictionary pattern. 3. The recognition dictionary learning method according to claim 1, wherein the correction frequency is calculated such that the smaller the difference between the first degree of mismatch and the second degree of mismatch, the higher the frequency of correcting the dictionary pattern. A learning method for learning a recognition dictionary. 4. The recognition dictionary learning apparatus according to claim 2, wherein the correction frequency is calculated so that the smaller the difference between the first degree of mismatch and the second degree of mismatch, the higher the frequency of correcting the dictionary pattern. A recognition dictionary learning device comprising the correction frequency calculating means.