JP2872208B1 - Method and apparatus for learning recognition dictionary - Google Patents

Abstract

An object of the present invention is to provide a learning apparatus for a recognition dictionary suitable for a normalized city block distance. A learning device according to the present invention is a learning device for a recognition dictionary used in a recognition device that performs matching based on a normalized city area distance, and includes elements of an input vector and corresponding elements of a reference vector and a weight vector. And a corresponding element of the reference vector and the weight vector according to the determining means 104 for determining the distance value with respect to
When belonging to the same category as the element of the input vector, while generating a correction vector for correcting the elements of the reference vector and the weight vector so as to enlarge the dead area,
When the corresponding elements of the reference vector and the weight vector belong to a different category from the elements of the input vector, a correction vector for generating a correction vector for correcting the elements of the reference vector and the weight vector so as to reduce the dead area It is characterized by comprising at least one pair of generating means (105, 111) and correcting means (108, 114) for correcting the elements of the reference vector and the weight vector according to the correction vector.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of learning a recognition dictionary used in a pattern recognition apparatus using a normalized city area distance, and a recognition dictionary learning apparatus to which the method is applied.

[0002]

2. Description of the Related Art A pattern recognition apparatus is an apparatus that receives a feature vector extracted from an input pattern as input and determines which category the input pattern belongs to by comparing it with a previously prepared recognition dictionary. is there. On the other hand, a recognition dictionary learning device performs statistical processing or the like on an input vector group prepared for learning, and automatically generates a recognition dictionary that reduces the rate of misrecognition of a category that should not belong to. Device.

[0003] As a learning method of a recognition dictionary, learning vector quantization (LVQ) (Kohonen, "Self-organizing map", Chapter 6, Springer Verlag Tokyo,
1996) and an improved version thereof (“Recognition Dictionary Learning Method and Recognition Dictionary Learning System Using the Same”, Japanese Patent Application No. 07/07/07)
-158069). These are the learning methods when the Euclidean distance is used as the distance between the input vector and the reference vector, and are not necessarily effective when the city block distance is used. Below, in this regard,
This will be described with reference to FIG.

First, as shown in FIG. 12, it is assumed that the number of categories is 2, and input vectors belonging to each category are distributed in a region shown by oblique lines in FIG. Under these conditions, 1 per category
It is assumed that one reference vector is prepared as a recognition dictionary, and each recognition dictionary is set at a position indicated by a black dot in FIG. In this case, the Euclidean distance between the input vector and the reference vector is represented by the following equation (1).

[0005]

(Equation 1) Also, the discrimination surface when using the Euclidean distance is shown in FIG.
As shown by a broken line in FIG. 2, a perpendicular 2 to a line connecting the two reference vectors (indicated by a dashed line in FIG. 12).
It becomes an equal line.

On the other hand, the city area distance is defined by the following equation (2).

[0007]

(Equation 2) As shown by a solid line in FIG. 12, the discrimination plane when this city block distance is used is a straight line inclined at 45 degrees with respect to the horizontal direction on the paper, inside a rectangle having two reference vectors as diagonals. On the other hand, outside the rectangle, a straight line in the vertical or horizontal direction on the paper surface causes many errors.

In order to minimize the error in the city area distance, a reference vector may be set at a position as shown in FIG. 13, for example. This reference vector is obtained by a conventional learning method based on Euclidean distance such as LVQ. Not what you get.

As can be understood from these facts, even if the reference vector minimizing the error can be obtained using the Euclidean distance, it is not necessarily the case that the error can be minimized using the city block distance. Not exclusively.

Incidentally, in addition to the Euclidean distance and the city area distance described above, a normalized city area distance obtained by improving the city area distance is known as a distance scale (Yamamoto, Mori, Mori, Shimizu, "Handwritten Katakana characters"). Recognition of Opportunities between Characters and Numbers: Phase Segmentation and Automatic Concept Formation— ”, IEICE Trans.
-69], p. 414-421, 1976). This normalized city area distance is defined by the following equation (3).

[0011]

(Equation 3) In the above (3), the element value m _i of the reference vector,
The average value of the i-th element value of the learning sample belonging to the same category is obtained, and the element value w _i of the weight vector is similarly obtained as the standard deviation of the i-th element value of the learning sample belonging to the same category.

It is known that the normalized city distance is more robust against noise than the simple city distance described above. This will be described with reference to FIG.

[0013] In FIG. 14, the vertical axis is the distance value corresponding to the i-th element of the vector, the horizontal axis represents the element values x _i of the input vector. Since the city block distance is d _i = | x _i −m _i |, a distance value as shown by a solid line in FIG. 14 is obtained. On the other hand, the normalized city area distance is d _i = max
Because it is _{(| | x i -m i -w} i, 0), the distance values as shown by the broken line in FIG. That is, for the region of ± w _i centered on m _i, the distance value becomes 0, so that the distance which the input value x _i is also stable and fluctuates around m _i is obtained.

As a method of automatically adjusting the weights related to the distance calculation, there is disclosed an invention in which a learning sample is repeatedly input and the weight value is updated so as to reduce erroneous recognition (see "Pattern Recognition"). Japanese Patent Laid-Open No. 5-233880). The present invention is limited to the case where the reciprocal of the standard deviation is used as the weight. The weight is not updated for correctly recognized inputs, and the weight is updated only for erroneously recognized inputs. Is to be reduced.

[0015]

However, in the conventional method of creating a recognition dictionary based on normalized city area distance and a recognition dictionary creation apparatus to which the method is applied, an average and its standard deviation are calculated without considering other categories. Therefore, there is a problem in that there is a limit in reducing false recognition. This will be briefly described below.

Referring to FIG. 15, there is shown an example of the i-th distance value of the vector when the number of categories is two. In FIG. 15, the curve shown on the left is a density function representing the distribution of learning samples belonging to category A, while the curve shown on the right is a density function representing the distribution of learning samples belonging to category B. is there. Here, it is assumed that the normalized city area distance when the average and the standard deviation are obtained for each category is as shown by a broken line in FIG. In FIG. 15, a dashed line a indicates a distance value for the category A, and a dashed line b indicates a distance value for the category B. In this case, the identification surface is at a position indicated by C in FIG. However, the position where the erroneous recognition is minimized is shown in FIG.
Is a position indicated by D, that is, a position where two density functions of categories A and B intersect. Therefore, when the reference vector and the weight vector are obtained without considering other categories, misrecognition is not necessarily performed. Clearly, it cannot be reduced.

In addition, the conventional automatic weight adjustment method has a problem that the recognition performance for unlearned data is not always improved because the weight is updated only for erroneously recognized inputs. This will be described below with reference to FIG.

FIG. 16 shows an example in which two categories of categories A and B are taken into consideration. A learning sample belonging to category A is represented by ◇, and a learning sample belonging to category B is represented by ×.
FIG. When the weights are updated according to the conventional automatic weight adjustment method so as to reduce erroneous recognition, FIG.
In some cases, an identification plane as indicated by a broken line is set. In this case, erroneous recognition of the learning sample is eliminated, but on the other hand, since the identification surface is set near the learning sample, an identification surface depending on the learning data is easily formed, and erroneous recognition is performed on the unlearned data. There are problems that can easily cause recognition. Also, this method relates to a method of updating the weight, and cannot update the reference vector.

As described above, each of the prior arts has a problem in creating a recognition dictionary for a normalized city area distance. That is, in the conventional recognition dictionary creation method based on the Euclidean distance and the recognition dictionary creation apparatus to which the method is applied, there is a problem that when the city block distance is used, erroneous recognition cannot always be reduced. In addition, in the case of using the conventional method of creating a normalized dictionary for urban area distance using the average and the standard deviation, there is a problem that a reduction in an erroneous recognition rate is limited because other categories are not considered. I was Further, the conventional automatic weight adjustment method has a problem that erroneous recognition is likely to occur for unlearned.

Therefore, the present invention solves these problems and provides a method of learning a recognition dictionary suitable for normalized city area distances, which aims to reduce erroneous recognition of learning patterns and reduce unrecognized patterns. It is another object of the present invention to provide a learning method for automatically generating a recognition dictionary capable of performing highly accurate recognition, and to provide a learning apparatus for a recognition dictionary to which the learning method is applied.

[0021]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention considers other categories while taking into account the area in which the element value of the input vector is included for both the correct category and the competing category that is likely to be misrecognized. The reference vectors and their weights are modified according to.

Specifically, the present invention provides the following recognition dictionary learning method and an apparatus to which the method is applied.

That is, according to the present invention, a recognition apparatus for comparing an input vector obtained by extracting a feature from a learning pattern with a reference vector prepared as a recognition dictionary and its weight vector based on a normalized city area distance. A learning method of the recognition dictionary used in the initial value setting step of performing initial setting of a reference vector and a weight vector, and an input step of selecting one input vector from a group of input vectors prepared for learning, A first reference vector and a first weight vector and a first distance value that minimize the normalized city area distance from a reference vector group belonging to the same category as the input vector, and a category different from the input vector. , A second reference vector and a second weight vector that minimize the normalized city area distance, and a second distance thereof. , A region determination step of determining a region including an input value for each element of the input vector, and a first distance value that reduces the first distance value using the result of the input determination step. A first reference vector correction vector calculation step for obtaining a reference vector correction vector, and a first weight for obtaining a first weight vector correction vector such that the first distance value decreases using the result of the area determination step. A vector correction vector calculation step, a first reference vector correction step of correcting the first reference vector using the first reference vector correction vector, and a first weight using the first weight vector correction vector. A first weight vector correction step of correcting a vector, and a second reference for obtaining a second reference vector correction vector such that a second distance value increases using a result of the area determination step. A vector correction vector calculation step; a second weight vector correction vector calculation step of obtaining a second weight vector correction vector such that the second distance value increases using the result of the area determination step; A second reference vector correction step of correcting the second reference vector using the reference vector correction vector, and a second reference vector correction step using the second weight vector correction vector.
A second weight vector correcting step of correcting the weight vector and a termination determining step of determining the termination of the learning.

Further, according to the present invention, a recognition apparatus for comparing an input vector obtained by extracting a feature from a learning pattern with a reference vector prepared as a recognition dictionary and its weight vector based on a normalized city area distance. A learning device for a recognition dictionary used in the, the initial value setting means for initial setting of the reference vector and the weight vector, and input means for selecting one input vector from a group of input vectors prepared for learning, A first reference vector and a first weight vector and a first distance value that minimize the normalized city area distance from a reference vector group belonging to the same category as the input vector, and a category different from the input vector. , A second reference vector and a second weight vector that minimize the normalized city area distance, and a second distance thereof. , A region determining unit that determines a region including an input value for each element of the input vector, and a first distance value that reduces a first distance value using the result of the input determining unit. First reference vector correction vector calculation means for obtaining a reference vector correction vector, and first weight for obtaining a first weight vector correction vector such that the first distance value decreases using the result of the area determination means. Vector correction vector calculation means, first reference vector correction means for correcting the first reference vector using the first reference vector correction vector, and first weighting using the first weight vector correction vector. A first weight vector correcting unit for correcting a vector, and a second reference for obtaining a second reference vector correction vector for increasing a second distance value using a result of the region determining unit. Vector correction vector calculation means, second weight vector correction vector calculation means for obtaining a second weight vector correction vector such that the second distance value increases using the result of the area determination means, A second reference vector correcting means for correcting the second reference vector using the reference vector correction vector, and a second reference vector correcting means using the second weight vector correction vector.
A second weight vector correcting means for correcting the weight vector and an end determining means for determining the end of the learning.

The method and apparatus for learning a recognition dictionary according to the present invention can also be represented by the following expressions.

That is, according to the present invention, an input vector obtained by extracting a feature from a learning pattern is compared with a reference vector and a weight vector prepared as a recognition dictionary by using a normalized city block distance. A learning method for learning a recognition dictionary used in a recognition apparatus that performs pattern recognition by using a distance value for each element of the input vector and corresponding elements of the reference vector and the weight vector, and determining the distance value. If the corresponding element of the reference vector and the weight vector belongs to the same category as the element of the input vector, the processing is performed so as to enlarge the dead area. Thus, a method of learning a recognition dictionary characterized by performing processing to reduce a dead area can be obtained.

Further, according to the present invention, in the method for learning the recognition dictionary, the processing for enlarging the dead area is performed so that the element value of the reference vector approaches the element value of the input vector. The correction is performed by increasing the element value of the weight vector, and the processing for reducing the dead area is performed so that the element value of the reference vector is kept away from the element value of the input vector, A method of learning a recognition dictionary, which is performed by reducing the element value of the weight vector, is obtained.

According to the present invention, there is provided a method of learning the recognition dictionary, wherein one of processing for expanding or reducing the dead area is performed over a plurality of categories according to the result of the distance value determination. Thus, a method of learning a recognition dictionary characterized by moving the identification surface to a point where the number of inputs of each of the plurality of categories becomes equal is obtained.

Further, according to the present invention, an input vector obtained by extracting a feature from a learning pattern, and a reference vector and a weight vector prepared as a recognition dictionary are used in a recognition device that performs the normalized city block distance. A learning device for a recognition dictionary, wherein each element of the input vector is:
Determining means for determining a distance value with respect to the corresponding element of the reference vector and the weight vector, and according to the determining means of the determining means, the corresponding element of the reference vector and the weight vector is the same as the element of the input vector. When belonging to the category, while generating a correction vector for correcting the elements of the reference vector and the weight vector so as to enlarge the dead area, the corresponding elements of the reference vector and the weight vector, the elements of the input vector and When belonging to a different category, according to a correction vector generating means for generating a correction vector for correcting the elements of the reference vector and the weight vector so as to reduce the dead area, and a correction vector generated by the correction vector generating means. To modify the elements of the reference vector and the weight vector Learning device recognition dictionary, characterized in that it comprises a correction means at least one set is obtained.

As described above, according to the present invention, in order to obtain a reference vector and a weight vector suitable for performing recognition based on the normalized urban area distance, the input vector is determined while determining the distance value for each element of the vector. For the same category as above, the reference vector and the weight vector are corrected so that the distance value becomes smaller, while for different categories, the reference vector and the weight vector are corrected so that the distance value becomes larger.

The recognition dictionary learned and generated according to the present invention having the above features has the following reasons.
The recognition performance is improved.

[0032] That is, in the present invention, the use of the normalized urban area distance, the distance values d _i for the element m _i of the reference vector will become as shown by the solid line in FIG. 17, m _i
Range of ± w _i from, x _i belongs to the region B, having 0 as the distance value, that is, a dead zone. Further, if x _i is included in the area A, a distance value _{d i = -x i + m i} -w i, whereas, when included in the region _{C, d i = x i -m} i -
w _i .

First, in the present invention, the reference vector and the weight vector belonging to the same category as the input vector are corrected in the direction in which the normalized city area distance decreases as shown in the following equation (4).

[0034]

(Equation 4) On the other hand, the reference vector and the weight vector belonging to a different category from the input vector are corrected in the direction in which the normalized city area distance increases as shown in the following equation (5).

[0035]

(Equation 5) In the expressions (4) and (5), α is a small positive real number. The element of the reference vector of distance values d _i values m _i
And the partial differentiation by weight vector w _i are
The following is obtained for each area.

That is, when x _i is included in the area A,
Since d _i = −x _i + m _i −w _i , the partial differentiation of the distance value d _i with the element value m _i of the reference vector and the weight vector w _i
Are expressed as in the following equation (6).

[0037]

(Equation 6) When _xi is included in the area B, since d _i = 0, they are respectively expressed by the following equations (7).

[0038]

(Equation 7) When _xi is included in the area C, d _i = x _i −m _i
Because it is -w _i, respectively, are shown as follows (8).

[0039]

(Equation 8) More clarity, reference vectors and weight vectors, along with if they belong to the same category as the input vector corrects an element value m _i of the reference vector so as to approach the element values x _i of the input vector, the elements of the weight vector By increasing the value w _i , the insensitive area is enlarged, while, if it belongs to a different category from the input vector,
As well as modified away the element value m _i of the reference vector from the element values x _i of the input vector, by reducing the element values w _i of the weight vector, to reduce the dead region.

Next, the result of processing based on the concept of the present invention will be described with reference to FIG.

For example, similar to LVQ2.1, the identification surface C
Referring to the case where a window W as shown by a dashed line is set in the vicinity, the reference vector m of the category A is included in the window W and the input belonging to the category A
_A moves rightward in FIG. 18 and the dead area widens, and the reference vector m _B of category B moves rightward and the dead area narrows in FIG. Will be moved to On the other hand, by the input included in the window W and belonging to the category B, the reference vector m _A of the category A moves to the left in FIG. 18, the dead area becomes narrower, and the reference vector m _B of the category B becomes Moving to the left in FIG. 18, the dead area becomes wider, and as a result, the identification plane C moves to the left. Here, since the frequency of movement of the identification plane C is proportional to the number of inputs included in the window W, the identification plane C is asymptotically determined by the number of inputs of the categories A and B included in the window W. Are approached, that is, the position indicated by D in FIG. 18, and as a result, erroneous recognition is improved.

As can be understood from the above brief description, the present invention considers not only input vectors belonging to the same category as the reference vector but also input vectors belonging to a different category. It is possible to reduce erroneous recognition as compared with a conventional recognition dictionary creation method of obtaining an average and a standard deviation using learning samples belonging to a category. Furthermore, in the present invention, the reference vector and the weight vector are corrected even if the recognition is correct, so that the weight is updated only when the recognition is incorrect. It is possible to improve recognition performance for data.

A method of efficiently creating a recognition dictionary with a high recognition system in consideration of not only the same category but also the relationship between different categories is disclosed in Japanese Patent Laid-Open No. 6-76061.
The main object of the present invention is to reduce the number of reference patterns and reduce the capacity of the recognition dictionary to speed up the recognition process. Neither does it focus on the specific problems in using it. It should be noted that some words may be similar, so that they may be misunderstood, and trying to understand the present invention while referring to the present invention may lead to unexpected misunderstandings. That is, it is needless to say that the technique cannot solve the problems raised by the present invention.

[0044]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A recognition dictionary learning method according to an embodiment of the present invention and a recognition dictionary learning apparatus to which the method is applied will be described below with reference to the drawings.

(First Embodiment) As shown in FIG. 1, a recognition dictionary learning apparatus according to a first embodiment of the present invention has an initial value setting unit 101, an input unit 102, and a dictionary search unit 10.
3. An area determination unit 104, a first correction vector correction vector calculation unit 105, a first dictionary correction unit 108, a second correction vector correction vector calculation unit 111, a second dictionary correction unit 114, and an end determination unit 117. ing. Here, the region determining means 104 surrounded by a broken line in FIG.
5, and a correction vector calculation means 1 for a second correction vector
11 is a feature of the present invention. Hereinafter, each component will be described.

The initial value setting means 101 initializes a reference vector and a weight vector used as a recognition dictionary.

The input means 102 selects one input vector from an input vector group prepared for learning.

The dictionary search means 103 selects, from among the recognition dictionaries belonging to the same category as the input vector selected by the input means 102, a first reference vector which minimizes the normalized city area distance from the input vector. And a first weight vector, and a first distance value which is a distance value between the vector and the input vector. Further, the dictionary search unit 103 includes a second reference vector and a second reference vector that minimize the normalized city area distance of the input vector from the recognition dictionaries belonging to a category different from the category of the input vector selected by the input unit 102. And a second distance value which is a distance value between the vector and the input vector.

The area determining means 104 determines an area including an input value for each element of the input vector.

The first correction vector calculation means 105 calculates the first correction vector
A correction vector calculation means for reference vector 106 and a correction vector calculation means 107 for first weight vector are provided. First reference vector correction vector calculation means 106
Calculates a correction vector for the first reference vector according to the determination result of the area determination unit 104,
The first weight vector correction vector calculation unit 107 calculates a correction vector for the first weight vector according to the determination result of the area determination unit 104.

The first dictionary correction means 108 includes a first reference vector correction means 109 and a first weight vector correction means 11
0. First reference vector correction means 109
Is for correcting the first reference vector using the first reference vector correction vector, and the first weight vector correcting means 110 uses the first weight vector correction vector to perform the first weight vector correction. Is to correct.

The second correction vector calculation means 111
A correction vector calculation means for reference vector 112 and a correction vector calculation means for second weight vector 113 are provided. Correction vector calculation means for second reference vector 112
Calculates a correction vector related to the second reference vector according to the determination result of the area determination unit 104,
The second weight vector correction vector calculation unit 113 calculates a correction vector related to the second weight vector according to the determination result of the area determination unit 104.

The second dictionary correction means 114 includes a second reference vector correction means 115 and a second weight vector correction means 11
6 is provided. Second reference vector correction means 115
Is used to correct the second reference vector using the second reference vector correction vector, and the second weight vector correction means 116 uses the second weight vector correction vector to generate the second weight vector. Is to correct.

The end determination means 117 is for determining the end of learning. In the end determination means 117,
If it is determined that the learning of the recognition dictionary has not been completed, the process returns to the input means 102, and the next input vector is selected.

Hereinafter, the operation processing in the present embodiment having the above-described configuration will be described with reference to the flowcharts shown in FIGS. 2 to 8, assuming that the number of categories is K and the total number of learning patterns is N. It is to be noted that as a premise of the following operation processing, an n-dimensional input vector is created in advance for each pattern using some kind of feature extraction method, and a set of reference vectors and weight vectors is used as a recognition dictionary with M sets per category. It shall be prepared.

FIG. 2 is a flowchart showing the operation processing of the initial value setting means 101.

The initial value setting means 101 first sets 1 as an initial value to a variable c representing a category number (step S201).
M reference vectors belonging to the category c are obtained (step S202). The M reference vectors are represented as in the following equation (9).

[0058]

(Equation 9) After obtaining the M reference vectors, the initial value setting means 101 sets M weight vectors belonging to the category c (step S203). Here, the M weight vectors are represented as in the following equation (10), and using the notation, the M weight vectors are set, for example, as in the following equation (11).

[0059]

(Equation 10)

[0060]

[Equation 11] Next, the initial value setting means 101 compares the value of c with the number of categories K (step S204), and when they are not equal, adds 1 to the value of c (step S205),
The process returns to step S202. If the values are equal, the process is terminated, and an initial value is output to the input unit 102.

FIG. 3 is a flowchart showing the operation processing after the input means 102.

The input means 102 first sets 1 as an initial value to a variable t representing the number of times of learning (step S30).
1), 1 is also set as an initial value for a variable j representing a learning pattern number (step S302). Then
The input unit 102 selects an input vector corresponding to the j-th learning pattern (Step S303). here,
The input vector is represented by the following equation (12).

[0063]

(Equation 12) The dictionary search means 103 searches the recognition dictionaries belonging to the same category as the input vector for the one having the smallest normalized city area distance (step S304). Note that the first reference vector, the first weight vector, and the first distance value used here are represented as in the following equation (13).

[0064]

(Equation 13) In addition, the dictionary search unit 103 searches the recognition dictionaries belonging to a category different from the input vector for the one having the smallest normalized city block distance (step S305). Note that a second reference vector, a second weight vector,
In addition, the second distance value is represented as in the following equation (14).

[0065]

[Equation 14] Given the element value of the input vector, the element value of the reference vector, and the element value of the weight vector, the area determination unit 104 determines whether the element value of the input vector is the area A, the area B, or the area A shown in FIG. It operates according to a program describing a function that returns which area of C is included (step S306).

The first correction vector calculation means 105 calculates a correction vector for the first reference vector, which is a correction vector for the first reference vector, and a correction vector for the first weight vector based on the determination result in step S306. A correction vector for the first weight vector is set (step S307). Here, the correction vector for the first reference vector and the correction vector for the first weight vector are respectively represented by the following equation (15).

[0067]

(Equation 15) The first dictionary correction unit 108 corrects the first reference vector and the first weight vector using the first reference vector correction vector and the first weight vector correction vector set in step S307 (step S307). S30
8).

On the other hand, the second correction vector calculation means 111
Sets a correction vector for the second reference vector, which is a correction vector for the second reference vector, and a correction vector for the second weight vector, which is a correction vector for the second weight vector, based on the determination result in step S306. (Step S309). Here, the correction vector for the second reference vector and the correction vector for the second weight vector are respectively represented by the following equation (16).

[0069]

(Equation 16) The second dictionary correction unit 114 corrects the second reference vector and the second weight vector using the second reference vector correction vector and the second weight vector correction vector set in step S309 (step S309). S31
0).

The end determination means 117 first compares the value of j with the number of learning patterns (step S311), and if they are not equal, adds 1 to the value of j (step S31).
2), the process returns to step S303. The end determination means 117 further compares the number of learnings t with a preset number of learnings T when it is determined that they are equal in step S311 (step S313). Is added to (step S31).
4) The process is returned to step S302, and if it is determined in step S313 that they are equal, the learning is terminated.

Next, the area determination means 104 (step S3)
The area determination process in step 06) will be described in more detail with reference to FIG.

The area determining means 104 first receives the element value x of the input vector, the element value m of the reference vector, and the element value w of the weight vector (step S401), and the element value x of the input vector is It is determined whether the difference is smaller than the difference m−w obtained by subtracting the element value of the weight vector from the element value (step S402). If the result of determination in step S402 is x <m−w, the area determination unit 10
4 assigns A to a variable R representing the type of region (step S403), and proceeds to the process of step S407 described later. On the other hand, as a result of the determination in step S402, x <
If it is not m−w, the area determination unit 104 determines whether or not the element value x of the input vector is greater than the sum m + w of the element value of the reference vector and the weight vector (step S404). As a result of the determination in step S404, x
If> m + w, the area determination unit 104 substitutes C for a variable R representing the type of the area (step S40).
5) The process proceeds to the process of step S407 described below. On the other hand, as a result of the determination in step S404, x> m + w
If not, the area determination unit 104 substitutes B for the variable R (step S406), and proceeds to the next step S407.
Proceed to. Finally, the area determination unit 104 determines in step S4
03, R in which any of A, B, and C is substituted in step S405 or step S406 is output as a region determination result (step S407).

Next, the first correction vector calculation means 105
Regarding the operation processing in (step S307), FIG.
This will be described in detail with reference to FIG.

First, the first modified vector calculation means 105 sets 1 as an initial value for a variable i representing a vector element (step S501). Next, the first correction vector calculation means 105 calculates the element value x of the input vector.
[i] is substituted for the variable x, the element value m ₁ [i] of the first reference vector is substituted for the variable m, and the element value w ₁ [i] of the first weight vector is substituted for the variable w. Then, the area determination means 1
04 (step S401), and R representing the type of region is obtained as an output (step S502).
The first correction vector calculation means 105 determines in step S502
It is determined whether or not R obtained in is equal to A (step S503). If it is determined that R is equal, -1 is substituted for the element value m ₁ [i] of the first reference vector correction vector. At the same time, 1 is substituted for the element value w ₁ [i] of the first weight vector correction vector (step S504),
The process proceeds to step S508 described below. On the other hand, when it is determined in step S503 that R = A is not satisfied,
The first correction vector calculation means 105 determines whether or not R is equal to C (step S505). If it is determined that R is equal to C, the first correction vector calculation means 105 determines the element value m ₁ [i] of the first reference vector correction vector. In addition, 1 is substituted for the element value w ₁ [i] of the first weight vector correction vector (step S506), and a step S50 described later is performed.
Proceed to step 8. Further, in step S505,
When it is determined that R = C is not satisfied, the first correction vector calculation unit 105 substitutes 0 for the element value m ₁ [i] of the first reference vector correction vector and sets the first weight vector correction vector Also, 0 is substituted for the element value w ₁ [i] (step S507), and a later-described step S508 is performed.
Processing proceeds to Finally, the first correction vector calculation means 1
05 compares the variable i with the number of dimensions n of the vector (step S508).
While the process is returned to step S2, if they match, the process is terminated and the first dictionary correction unit 108 (step S30)
The processing shifts to 8).

Next, the process of updating the recognition dictionary in the first dictionary correction means 108 (step S308) will be described with reference to FIG.
This will be described in detail with reference to FIG.

First, the first dictionary correction means 108 sets 1 as an initial value to a variable i representing an element of a vector (step S601). Next, the first dictionary correction unit 108
First correction vector calculation means 105 (step S307)
The element value m ₁ [i] of the first reference vector and the element value w ₁ [i] of the first weight vector are corrected using the two correction vectors obtained in (Step S602). Here, α is a real number from 0 to 1 and, for example, 0.1
Etc. Finally, the first dictionary correction means 108
Compares the variable i with the number of dimensions n of the vector after correcting the first reference vector and the first weight vector (step S603).
02, and if they are equal, the process is terminated, and the process proceeds to the end determination unit 117.

Next, the second correction vector calculation means 111
Regarding the operation processing in (Step S309), FIG.
This will be described in detail with reference to FIG.

First, the second modified vector calculation means 111 sets 1 as an initial value for a variable i representing a vector element (step S701). Next, the second correction vector calculation means 111 calculates the element value x of the input vector.
[i] is substituted for the variable x, the element value m ₁ [i] of the second reference vector is substituted for the variable m, and the element value w ₁ [i] of the second weight vector is substituted for the variable w. Then, the area determination means 1
04 (step S401), and R representing the type of region is obtained as an output (step S702).
The second correction vector calculation means 111 determines in step S702
It is determined whether or not R obtained in is equal to A (step S703). If it is determined that R is equal to -1, -1 is substituted for the element value m ₂ [i] of the second reference vector correction vector. At the same time, 1 is substituted for the element value w ₂ [i] of the correction vector for the second weight vector (step S704),
The process proceeds to step S708 described below. On the other hand, if it is determined in step S703 that R = A is not satisfied,
The second correction vector calculation means 111 determines whether or not R is equal to C (step S705). If it is determined that R is equal to R, the second correction vector calculation means 111 determines the element value m ₂ [i] of the second reference vector correction vector. In addition, 1 is substituted for the element value w ₂ [i] of the second weight vector correction vector (step S706), and a later-described step S70 is performed.
Proceed to step 8. Further, in step S705,
When it is determined that R = C is not satisfied, the second correction vector calculation unit 111 substitutes 0 for the element value m ₂ [i] of the second reference vector correction vector, and Also, 0 is substituted for the element value w ₂ [i] (step S707), and step S708 described later is performed.
Processing proceeds to Finally, the second correction vector calculation means 1
11 compares the variable i with the number of dimensions n of the vector (step S708).
While the process returns to step S2, if the values match, the process is terminated and the second dictionary correction unit 108 (step S31)
The processing shifts to 0).

Next, the process of updating the recognition dictionary in the second dictionary correction means 114 (step S310) will be described with reference to FIG.
This will be described in detail with reference to FIG.

First, the second dictionary correcting means 114 sets 1 as an initial value to a variable i representing a vector element (step S801). Next, the second dictionary correction unit 114
Second correction vector calculation means 111 (step S309)
The element value m ₂ [i] of the second reference vector and the element value w ₂ [i] of the second weight vector are corrected using the two correction vectors obtained in (Step S802). Again, α is a real number from 0 to 1 and, for example, 0.1
Etc. Finally, the second dictionary correction means 114
Compares the variable i with the number of dimensions n of the vector after correcting the second reference vector and the second weight vector (step S803).
02, and if they are equal, the process is terminated, and the process proceeds to the end determination unit 117.

According to the recognition dictionary learning apparatus of the present embodiment having a configuration capable of performing such processing, a reference vector and a weight vector suitable for the normalized city area distance can be obtained. It is possible to sufficiently reduce the erroneous recognition rate and create a recognition dictionary having high recognition performance even for unlearned patterns.

(Second Embodiment) The second embodiment of the present invention is a modification of the above-described first embodiment.
This is an application example of the present invention to the case where the normalized city area distance is defined by the following equation (17).

[0083]

[Equation 17] In the present embodiment, in the processing shown in FIGS. 2 and 3, in step S304, the first reference vector and the first weight vector are searched based on the distance obtained from Expression (17). At the same time, in step S305,
A point where the second reference vector and the second weight vector are searched based on the distance obtained from the equation (17);
Step S306, step S307, and step S
309 is the same as the above-described first embodiment, except for the difference in the processing in 309.

First, the processing in the area determining means 104 (step S306) in the present embodiment will be described with reference to FIG.
This will be described with reference to FIG.

The area determining means 104 receives the element value x of the input vector, the element value m of the reference vector, and the element value w of the weight vector (step S901). The value obtained by multiplying the difference obtained by subtracting the element value of the vector by the element value of the reference vector (1-
w) It is determined whether it is smaller than m (step S90)
2). As a result of the determination in step S902, x <(1
-W) If m, the area determination unit 104 substitutes A for a variable R representing the type of area (step S903), and proceeds to the processing of step S907 described below. On the other hand, if the result of determination in step S902 is that x <(1−w) m, the area determination unit 104 determines that the element value x of the input vector is 1 It is determined whether or not the value is multiplied by (1 + w) m (step S904). As a result of the determination in step S904, if x> (1 + w) m, the area determination unit 104 substitutes C for a variable R representing the type of the area (step S905), and proceeds to step S907 described below. move on. On the other hand, if the result of the determination in step S904 is not x> (1 + w) m,
04 substitutes B for a variable R (step S906),
The process proceeds to the next step S907. Finally, area determination means 1
In step S04, R in which any of A, B, and C is substituted in step S903, step S905, or step S906 is output as a region determination result (step S9).
07).

Next, the operation processing in the first modified vector calculation means 105 (step S307) in the present embodiment will be described in detail with reference to FIG.

First, the first modified vector calculating means 105 sets 1 as an initial value for a variable i representing a vector element (step S1001). Next, the first correction vector calculation means 105 calculates the element value x of the input vector.
[i] is substituted for the variable x, the element value m ₁ [i] of the first reference vector is substituted for the variable m, and the element value w ₁ [i] of the first weight vector is substituted for the variable w. Then, the area determination means 1
04 (step S901), and R representing the type of region is obtained as an output (step S100).
2). The first correction vector calculation means 105 determines in step S
It is determined whether or not R obtained in 1002 is equal to A (step S1003).
Substituting −1 + w for the element value m ₁ [i] of the first reference vector correction vector and substituting w for the element value w ₁ [i] of the first weight vector correction vector (step S1004), and the process proceeds to step S1008 described below. On the other hand, in step S1003, R =
If it is determined that it is not A, the first correction vector calculation means 1
05 determines whether or not R is equal to C (step S1005). If it is determined that R is equal to C, 1+ is added to the element value m ₁ [i] of the first reference vector correction vector.
w and w to the element value w ₁ [i] of the first weight vector correction vector (step S1).
006), and the process proceeds to step S1008 described below. Further, if it is determined in step S1005 that R is not C, the first correction vector calculating unit 105
Substituting 0 for the element value m ₁ [i] of the first reference vector correction vector, and also substituting 0 for the element value w ₁ [i] of the first weight vector correction vector (step S1007), and the process proceeds to step S1008 described below. Finally, the first correction vector calculation means 105 calculates
The variable i is compared with the number of dimensions n of the vector (step S1).
008) If they do not match, the process returns to step S1002, while if they do match, the process ends and the process moves to the first dictionary correction unit 108 (step S308).

Next, the operation processing in the second correction vector calculation means 111 (step S309) in this embodiment will be described in detail with reference to FIG.

First, the second correction vector calculation means 111 sets 1 as an initial value for a variable i representing a vector element (step S1101). Next, the second correction vector calculation means 111 calculates the element value x of the input vector.
[i] is substituted for the variable x, and the element value m of the second reference vector
₁ [i] is substituted for a variable m, and further, the element value w ₁ [i] of the second weight vector is substituted for a variable w, so that the area determination means 10
4 (step S401) and output as
An R representing the type of the area is obtained (step S1102). The second correction vector calculation means 111 determines whether or not R obtained in step S702 is equal to A (step S702).
1103), when it is determined that they are equal to each other, the element value m ₂ [i] of the correction vector for the second reference vector is −1 + w
And w is substituted for the element value w ₂ [i] of the second weight vector correction vector (step S11).
04), the flow proceeds to the process in step S1108 described later.
On the other hand, if it is determined in step S1103 that R is not A, the second correction vector calculation unit 111 determines whether R is equal to C (step S110).
5) If it is determined that they are equal, 1 + w is substituted for the element value m ₂ [i] of the second reference vector correction vector, and the element value w of the second weight vector correction vector
₂ w is substituted for [i] (step S1106), and the process proceeds to step S1108 described below. Further, if it is determined in step S1105 that R = C is not satisfied, the second correction vector calculation unit 111 substitutes 0 for the element value m ₂ [i] of the second reference vector correction vector, and 0 is also substituted for the element value w ₂ [i] of the weight vector correction vector (step S110)
7), the process proceeds to step S1108 described below. Finally, the second correction vector calculation means 111 compares the variable i with the number of dimensions n of the vector (step S1108),
If they do not match, the process returns to step S1102, while if they match, the process ends and the second
The processing shifts to the dictionary correction means 108 (step S310).

In the second embodiment for performing such processing, the same effect as in the first embodiment described above is obtained, though the handling of the distance is different.

In the first and second embodiments,
Although the number of recognition dictionaries per category has been described as M, a different number of recognition dictionaries may be used for each category.

Further, it is needless to say that various modifications and applications can be made to the above-described two embodiments under the concept of the present invention.

[0093]

As is apparent from the above description, according to the present invention, it is possible to obtain a reference vector and a weight vector suitable for the normalized city area distance, and to sufficiently reduce the false recognition rate for the learning pattern. And a recognition dictionary having high recognition performance even for unlearned patterns can be created.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a recognition dictionary learning device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing an operation process of an initial value setting unit in FIG. 1;

FIG. 3 is a flowchart showing an operation process in a block after an input unit in FIG. 1;

FIG. 4 is a flowchart showing a process in an area determining unit in FIG. 1;

FIG. 5 is a flowchart showing a process in a first correction vector calculating means in FIG. 1;

FIG. 6 is a flowchart showing a process in a first dictionary correction unit in FIG. 1;

FIG. 7 is a flowchart showing a process in a second correction vector calculation means in FIG. 1;

FIG. 8 is a flowchart showing a process in a second dictionary correction unit in FIG. 1;

FIG. 9 is a flowchart showing a process in an area determining unit provided in the recognition dictionary learning device according to the second embodiment of the present invention.

FIG. 10 is a flowchart showing processing in a first correction vector calculating means provided in the recognition dictionary learning device according to the second embodiment of the present invention.

FIG. 11 is a flowchart showing a process in a second correction vector calculating means provided in the recognition dictionary learning device according to the second embodiment of the present invention.

FIG. 12 is a diagram for explaining a difference between a city block distance and a Euclidean distance.

FIG. 13 is a diagram for explaining an identification plane based on a city area distance.

FIG. 14 is a diagram for explaining a normalized city area distance.

FIG. 15 is a diagram for explaining a problem of the related art regarding a normalized city area distance.

FIG. 16 is a diagram for explaining a problem of a conventional automatic weight adjustment method.

FIG. 17 is a diagram for explaining an area at a normalized city area distance.

FIG. 18 is a diagram for explaining an operation when the present invention is applied.

[Explanation of symbols]

Reference Signs List 101 initial value setting means 102 input means 103 dictionary search means 104 area determination means 105 first correction vector calculation means 106 first reference vector correction vector calculation means 107 first weight vector correction vector calculation means 108 first dictionary correction means 109 First reference vector correction means 110 First weight vector correction means 111 Second correction vector calculation means 112 Second reference vector correction vector calculation means 113 Second weight vector correction vector calculation means 114 Second dictionary correction means 115 Second reference Vector correction means 116 Second weight vector correction means 117 End determination means

Claims (6)

(57) [Claims] An input vector obtained by extracting features from a learning pattern is compared with a reference vector prepared as a recognition dictionary and its weight vector based on a normalized city distance. A learning method, comprising: an initial value setting step of initially setting a reference vector and a weight vector; an input step of selecting one input vector from a group of input vectors prepared for learning; and the same category as the input vector. Among the reference vector group belonging to the first reference vector and the first weight vector and the first distance value that minimize the normalized city area distance;
A dictionary search step for obtaining, from a group of reference vectors belonging to a category different from the input vector, a second reference vector, a second weight vector, and a second distance value that minimize the normalized city area distance; An area determining step of determining an area including an input value for each element of the input vector; and a step of obtaining a first reference vector correction vector such that the first distance value decreases using the result of the input determining step. 1
Calculating a correction vector for a reference vector, and obtaining a correction vector for a first weight vector such that the first distance value decreases using the result of the area determination step.
Calculating a correction vector for a weight vector; correcting a first reference vector using the correction vector for a first reference vector; and performing a first correction using a correction vector for the first weight vector. A second weight vector correction step of correcting a weight vector; and a second reference vector correction vector for increasing a second distance value using a result of the area determination step.
Calculating a correction vector for a second weight vector using a result of the step of calculating a correction vector for a reference vector and a second distance value using a result of the area determination step;
Calculating a correction vector for the weight vector, a second reference vector correction step of correcting the second reference vector using the second correction vector for the reference vector, and a second step using the correction vector for the second weight vector. A method for learning a recognition dictionary, comprising: a second weight vector correction step of correcting a weight vector; and an end determination step of determining the end of learning. 2. A recognition dictionary for use in a recognition apparatus used in a recognition device for comparing an input vector obtained by extracting a feature from a learning pattern with a reference vector prepared as a recognition dictionary and its weight vector based on a normalized city area distance. A learning device, comprising: initial value setting means for initial setting of a reference vector and a weight vector; input means for selecting one input vector from an input vector group prepared for learning; and the same category as the input vector. Among the reference vector group belonging to the first reference vector and the first weight vector and the first distance value that minimize the normalized city area distance;
Dictionary search means for obtaining, from a group of reference vectors belonging to a category different from the input vector, a second reference vector, a second weight vector, and a second distance value that minimize the normalized city area distance; Area determining means for determining an area including an input value for each element of the input vector; and obtaining a first reference vector correction vector such that the first distance value decreases using the result of the input determining means. 1
A first correction vector for a weight vector that reduces the first distance value using a result of the reference vector correction vector calculation unit and the area determination unit;
Weight vector correction vector calculation means, first reference vector correction means for correcting the first reference vector using the first reference vector correction vector, and first weight vector correction vector for the first reference vector. A first weight vector correction unit for correcting a weight vector; and a second reference vector correction vector for increasing a second distance value using a result of the area determination unit.
A second correction vector for obtaining a second weight vector that increases the second distance value using a result of the reference vector correction vector calculation unit and the area determination unit;
Weight vector correction vector calculation means; second reference vector correction vector using the second reference vector correction vector; second reference vector correction means using the second weight vector correction vector; A learning device for a recognition dictionary, comprising: a second weight vector correction unit for correcting a weight vector; and an end determination unit for determining the end of learning. 3. An input vector obtained by extracting a feature from a learning pattern, and a reference vector and a weight vector prepared as a recognition dictionary are obtained by using a normalized city block distance.
A method of learning a recognition dictionary used in a recognition device that performs pattern recognition by performing matching, wherein a distance value is determined for each element of the input vector and a corresponding element of the reference vector and the weight vector. As a result of the determination, if the corresponding element of the reference vector and the weight vector belongs to the same category as the element of the input vector, processing is performed to enlarge the dead area.
A method for learning a recognition dictionary, characterized in that if the element belongs to a category different from the element of the input vector, processing is performed to reduce the dead area. 4. The method for learning a recognition dictionary according to claim 3, wherein the processing for expanding the dead area corrects the element value of the reference vector so as to approach the element value of the input vector. The processing for reducing the dead area is performed by increasing the element value of the weight vector, and correcting the element value of the reference vector so as to be away from the element value of the input vector. A method of learning a recognition dictionary, which is performed by reducing an element value of a weight vector. 5. The method for learning the recognition dictionary according to claim 3, wherein one of processing for expanding or reducing the dead area is performed over a plurality of categories according to a result of the distance value determination. A method of learning a recognition dictionary, characterized by moving an identification surface to a point where the number of inputs of each of a plurality of categories is equal. 6. A recognition dictionary learning apparatus used in a recognition apparatus for performing an input vector obtained by extracting a feature from a learning pattern and a reference vector and a weight vector prepared as a recognition dictionary by using a normalized city area distance. Determining means for determining a distance value for each element of the input vector and a corresponding element of the reference vector and the weight vector; and a correspondence between the reference vector and the weight vector according to the determining means of the determining means. When the element to be included belongs to the same category as the element of the input vector, a correction vector for correcting the element of the reference vector and the weight vector so as to enlarge the dead area is generated, and the reference vector and the weight vector are If the corresponding element belongs to a different category from the element of the input vector, Correction vector generating means for generating a correction vector for correcting the elements of the reference vector and the weight vector so as to be reduced; and, according to the correction vector generated by the correction vector generating means, the elements of the reference vector and the weight vector. A learning device for a recognition dictionary, comprising: at least one set of correction means for correcting the following.