JPH07311846A - Pattern recognition dictionary preparation device - Google Patents

Abstract

PURPOSE:To provide a pattern recognition dictionary preparation device which and eliminates a need of the processing of affine part spaces of categories contending with a noticed category while taking the erroneous recognition rate as the evaluation reference and is easily applied to a multi-category pattern recognition dictionary. CONSTITUTION:A standard feature vector update part 13 refers to a category limiting learning pattern recognition result storage part 11 to count the number of erroneously recognized patterns and takes out these patterns one by one to generate standard feature vectors, which are moved so as to extend the distances between erroneously recognized patterns and erroneous recognition destination categories, or partial space base vectors, which are moved so as to reduce the distances between erroneously recognized patterns and their correct answer categories, and transfers these vectors to a category limiting learning pattern recognition part 10. The part 10 uses transferred standard feature vectors instead of standard feature vectors stored in a standard feature vector storage part 3 to execute the category limiting recognition of the noticed category and categories contending with it.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention uses a set of feature vectors extracted from a learning pattern to recognize a recognition dictionary used by a pattern recognition device, that is, a set of quantities representing a typical shape for each recognition target category. The present invention relates to a pattern recognition dictionary creating device for creating a pattern.

[0002]

2. Description of the Related Art A pattern recognition dictionary creating device is a device for performing various statistical processes on a set of feature vectors extracted from a learning pattern and outputting a recognition dictionary. The pattern recognition device recognizes the unknown pattern by calculating the distance or similarity between the unknown pattern and each category stored in the recognition dictionary using the discriminant function. The structure of the recognition dictionary depends on the discriminant function used.

When the discriminant function is the Euclidean distance,
A standard feature vector representing each category is used as a recognition dictionary. That is, the distance De between the unknown pattern and the category of interest is

[0004]

[Equation 1] Given in. Here, x is a feature vector of an unknown pattern, μ is a standard feature vector of the target category, and ‖‖ is a Euclidean norm. Generally, as the standard feature vector μ, the average vector of the feature vectors of the learning patterns of the target category:

[0005]

[Equation 2] Is used. Here, x _n represents the feature vector of the n-th learning pattern in the target category, and N represents the number of learning patterns in the target category.

Usually, there is a correlation between the components of the feature vector. Therefore, even in the case of pattern sets of the same category, there are many cases in which the direction in which the pattern distribution spreads in the feature space is not parallel to the feature axis. When the Euclidean distance is used, such a correlated distribution is represented by one point in the feature space called the standard feature vector, and therefore its discriminating ability has a limit.

Therefore, the expression by "points" of each category is expanded to the expression by an affine subspace such as a straight line or a plane, and the distance between the unknown pattern and each category is calculated by using the feature vector of the unknown pattern and the affine part of each category. A method of defining the shortest distance from the space has been proposed (Ikeda, Tanaka,
Motooka: “Projection distance method in handwritten character recognition”, Journal of Information Processing Society of Japan, Vol. 24, No. 1, pp. 106
-112, 1983-01).

That is, as shown in FIG. 6, it is composed of a subspace L designated by basis vectors Ψ ₁ , Ψ ₂ , ..., Ψ _p (p: dimensionality of subspace) and a standard feature vector μ. Affine subspace:

[0009]

[Equation 3] Is used to express the distribution of each category. The distance Dr between the unknown pattern and each category is defined as the feature vector x of the unknown pattern.
And the shortest distance between each affine subspace S of each category

[0010]

[Equation 4] Define in. The minimum of ‖x−y‖ is the vector x
Since -y and the affine subspace S are orthogonal to each other, Dr can be expressed by the following equation.

[0011]

[Equation 5] In the projection distance method, μ is the average vector

[0012]

[Outer 1] To use. Further, the basis vectors Ψ ₁ , Ψ ₂ , ..., Ψ _p are determined using principal component analysis. That is, the covariance matrix A for each category:

[0013]

[Equation 6] Eigenvalue problem: eigenvectors Ψ ₁ , Ψ ₂ ,…, Ψ _K obtained by solving A Ψ = λΨ
(K: number of feature vector dimensions), corresponding eigenvalue λ
Let p selected in descending order of _{j be} the subspace basis vectors of the target category.

The projection distance is an excellent discriminant function as compared with the Euclidean distance, but since the principal component analysis is used, the projection distance is distributed in the learning pattern near the average vector which is high in the appearance frequency although it is not likely to be misrecognized. Affin subspace is strongly influenced. Therefore, the recognition accuracy is greatly affected. There is a problem that the learning pattern information existing near the boundary with other categories is difficult to be reflected.

As one method for solving this problem,
There is a learning subspace method proposed by Kohonen et al. (Translated by Erki Oya, Hidemitsu Ogawa, Makoto Sato; “Pattern Recognition and Subspace Method”, Sangyo Tosho, 1986, pp. 91-
118). This is because the learning pattern is recognized with the subspace determined using the principal component analysis as the initial subspace, and for each misrecognized pattern, the distance from the correct category becomes small, and the misrecognized destination category This is a subspace optimization method in which the subspaces of the correct category and the misrecognized category are rotated so that the distance between and becomes larger.

The learning subspace method of Kohonen et al. Uses a simple subspace L instead of the affine subspace S, that is, an affine subspace including the origin, and the projection length to the subspace instead of the distance to the subspace:

[0017]

[Equation 7] Although the format is slightly different from the projection distance method described above, the same idea can be easily applied to the projection distance method.
The procedure is as follows.

As shown in FIG. 7, the standard features of the target category are
The characteristic vector is μ and the basis vector is Ψ ₁ , Ψ₂ ,…, Ψ
_p , X is the feature vector of the misrecognition pattern used for learning
When doing, follow the steps below to perform affine subspace
Is rotated around the standard feature vector μ. (1) To the affine subspace of the target category of vector x
The projection y of is calculated by the following equation.

[0019]

[Equation 8] (2) After replacing the vector y-μ with one of the subspace basis vectors, for example, Ψ ₁ , using the Gram-Schmidt orthogonalization method for y-μ, Ψ ₂ , ..., Ψ _p. Orthonormal basis vector

[0020]

[Equation 9] To create. (3) New basis vectors Φ ₁ , Φ ₂ , ..., Φ _p after rotation
Is calculated by the following formula.

[0021]

[Equation 10] here,

[0022]

[Equation 11] Is a parameter corresponding to the rotation angle. The distance between x and the affine subspace has a minimum value of 0 when θ = ∞,

[0023]

[Equation 12] Is the maximum

[0024]

[Equation 13] become. In particular,

[0025]

[Equation 14] In this case, correct recognition of the learning pattern x is guaranteed. here,
Dr is the distance between x and the affine subspace of the target category:

[0026]

[Equation 15] De is the distance between x and the standard feature vector of the target category:

[0027]

[Equation 16] Represents Also, Dr ′ is the distance from the affine subspace of the misrecognized category when x is the pattern of the target category that is erroneously recognized by another category, and x is the pattern of another category that is erroneously recognized by the target category. Is the distance to the affine subspace of the correct category:

[0028]

[Equation 17] Represents μ ′ is a standard feature vector of another category, and y ′ is a projection of another category of x onto an affine subspace.

Although θ is not the angle itself, it has a one-to-one correspondence with the angle, and henceforth, it will be called "rotation angle" for convenience.
Further, θ ° will be referred to as a rotation angle reference value. When the rotation angle reference value θ ° is used as θ, other learning patterns are erroneously recognized even if the focused learning pattern x is correctly read, and as a whole,
On the contrary, it often happens that the misrecognition pattern increases. To avoid this, Kohonen et al. Have proposed a method in which θ ° is equally distributed to both the correct category and the misrecognized category, and both affine subspaces are rotated simultaneously. If this method is used, the amount of rotation of each category should be small, and false recognition of other learning patterns should be small.

[0030]

[Problems to be Solved by the Invention] However, Koh
Even if the rotation angle is shared by both the correct answer category and the erroneous recognition destination category as in the method of onen et al., there is no guarantee that erroneous recognition of other learning patterns does not occur. This is because, depending on the distribution of patterns in the feature space, it is still possible that a slight rotation of the affine subspace will cause a large number of false recognitions. On the other hand, even if a rotation larger than the rotation angle reference value θ ° is performed and the correct recognition is possible without causing misrecognition of other learning patterns, the method of Kohonen et al. Absent.

A method for surely reducing the false recognition is to repeat a minute rotation with the rotation angle reference value as a guide, recognize the learning pattern every time the rotation angle is updated, and change the number of false recognition patterns from a decrease to an increase. It is a method of stopping the rotation of the affine subspace. This method is a preferable affine subspace optimization method in terms of using the false recognition rate as an evaluation criterion, but it requires a large amount of processing because recognition of learning patterns is repeated. for that reason,
When the number of categories is large, the evaluation of the number of false recognition patterns is
There is no choice but to limit the attention category to a category (hereinafter, referred to as a competition category) in which misrecognition easily occurs between the attention category.

In this case, when trying to execute the rotation of the affine subspace of the competitive category according to the method of Kohonen et al., It is necessary to use the learning pattern of the competitive category and the recognition dictionary for the competitive category to evaluate the number of misrecognized patterns. Therefore, even if there are 10 competing categories with respect to one category, the subspace optimization must be performed in consideration of the maximum 100 categories per focused category, which requires a large amount of processing.
Therefore, the application of this method to multi-category pattern recognition is limited.

Therefore, if the rotation of the affine subspace is limited to the rotation of the target category, some patterns that are erroneously recognized as the target category may not be correctly read. Specifically, De <Dr ', that is, the Euclidean distance De between the feature vector of the pattern of interest and the standard feature vector of the misrecognized category which is the category of interest is the feature vector of the pattern of interest and the affine of the correct category. The pattern is smaller than the distance Dr 'from the subspace. This means that no matter how the affine subspace of the misrecognized category is rotated, the distance Dr between the feature vector of the pattern of interest and the affine subspace of the misrecognized category does not become larger than De: Dr <De Therefore, Dr <Dr 'always holds. Therefore, the correct recognition of such patterns can only be expected from the optimization process in which the correct category is set as the focused category.

An object of the present invention is to provide a pattern recognition dictionary creating apparatus which does not need to process an affine subspace of a competing category of a target category while using an erroneous recognition rate as an evaluation criterion and which can be easily applied to a multi-category pattern recognition dictionary. To provide.

[0035]

A pattern recognition dictionary creating apparatus of the present invention uses a set of feature vectors extracted from a learning pattern of a learning pattern feature vector and an affine subspace for each category, In a pattern recognition dictionary creating apparatus for determining an affine subspace, a learning pattern recognition means for recognizing the learning pattern using a distance between a feature vector of the learning pattern and the affine subspace for each category, and the learning pattern recognition means. Based on the recognition result of the learning pattern set recognized by using, a competing category set determining means for determining a set of categories with high misrecognition frequency for each target category, and a category of the affine subspace used for recognition , The category of the learning pattern to be recognized, the target category, and the competing category set for the target category With respect to the limited category recognition pattern recognition means and the learning pattern of the competing category that is erroneously recognized by the category limitation learning pattern recognition means as the focused category, the feature vector of the learned pattern and the affine subspace of the focused category The affine subspace is rotated in the direction in which the distance between the two becomes larger, and the learning pattern of the target category erroneously recognized as a competitive category by the category-limited learning pattern recognizing means is the affine of the feature vector of the learning pattern and the target category. Regarding the affine subspace rotating means for rotating the affine subspace in the direction in which the distance to the subspace becomes smaller, and the learning pattern of the competitive category erroneously recognized as the focused category by the category limited learning pattern recognition means, Of the pattern feature vector and the target category For the learning pattern of the target category that is erroneously recognized as a competitive category by the category-limited learning pattern recognition means, the feature vector of the learning pattern is translated by parallel moving the affine subspace in the direction in which the distance to the fin subspace increases. And affine subspace translation means for translating the affine subspace in a direction in which the distance between the affine subspace and the affine subspace of the target category becomes smaller.

[0036]

The present invention is most characterized in that it has a means for moving the affine subspace in addition to a means for rotating the affine subspace. That is, when the projection vector of the vector x-μ onto the affine subspace is y-μ, the standard feature vector μ is moved in the vector xy direction as shown in FIG. The distance between the vector x and the affine subspace of the target category is increased or decreased. The standard feature vector v after movement is represented by the following equation.

[0037]

[Equation 18] τ (−∞ <τ <∞) is a parameter representing the amount of movement of the standard feature vector, and the distance between x and the affine subspace is

[0038]

[Formula 19] Takes a minimum value of 0. In particular,

[0039]

[Equation 20] In this case, correct recognition of the learning pattern x is guaranteed. here,
Dr is the distance between x and the affine subspace of the target category, and Dr ′ is the distance between the affine subspace of the misrecognized category when x is the feature vector of the pattern of the target category that is erroneously recognized by another category. , X is a feature vector of a pattern of another category that is erroneously recognized as the focused category, it represents the distance from the affine subspace of the correct category. Hereinafter, this τ ° will be referred to as a movement amount reference value.

Like τ, τ is gradually increased from 0 with τ ° as a guide. The learning pattern is recognized every time the standard feature vector is updated with τ, and the parallel movement of the affine subspace is terminated when the number of erroneous recognition patterns changes from decreasing to increasing.

Unlike the rotation angle θ, there is no limit on the range of the movement amount τ, so that any erroneous recognition pattern is correctly recognized only by processing the target category unless the erroneous recognition of other learning patterns increases. be able to. Therefore, it is not necessary to process the affine subspace of the competing category of the target category while using the recognition error rate as an evaluation criterion, and it is possible to provide a pattern recognition dictionary creating apparatus that can be easily applied to a pattern recognition dictionary of multiple categories.

[0042]

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

FIG. 1 is a block diagram of a pattern recognition dictionary creating apparatus according to an embodiment of the present invention.

The pattern recognition dictionary creating apparatus of the present embodiment includes a learning pattern feature vector storage unit 1, an initial standard feature vector creation unit 2, a standard feature vector storage unit 3, an initial subspace basis vector creation unit 4, and a subspace basis vector. A storage unit 5, a learning pattern recognition unit 6, a learning pattern recognition result storage unit 7, a competitive category set determination unit 8, a competitive category set storage unit 9, a category limited learning pattern recognition unit 10, a category limited learning pattern recognition result storage unit 11, It is composed of a subspace basis vector updating unit 12 and a standard feature vector updating unit 13. In FIG. 1, a portion surrounded by a broken line is a portion added by the present invention.

The feature vector of the learning pattern used for creating the recognition dictionary is stored in the learning pattern feature vector storage unit 1. The initial standard feature vector creation unit 2 reads the feature vector stored in the learning pattern feature vector storage unit 1 for each category, creates a standard feature vector by averaging, and stores it in the standard feature vector storage unit 3. The initial subspace basis vector creation unit 4 reads the feature vector stored in the learning pattern feature vector storage unit 1 for each category, calculates the eigenvector by principal component analysis, and stores it in the subspace basis vector storage unit 5.
The learning pattern recognition unit 6 uses the standard feature vector stored in the standard feature vector storage unit 3 and the subspace basis vector stored in the subspace basis vector storage unit 5 to learn pattern feature vector storage unit 1 The feature vector stored in is recognized, and the result is stored in the learning pattern recognition result storage unit 7. Competitive category set determination unit 8
Calculates the erroneous recognition frequency between categories from the recognition result stored in the learning pattern recognition result storage unit 7, determines the competitive category set for each category, and stores it in the competitive category set storage unit 9.

The subspace basis vector updating unit 12 carries out, for each focused category, limited category recognition of the focused category and its competing category (recognizing the recognized category by limiting it to some categories). Request the recognition unit 10. The category-limited learning pattern recognition unit 10 reads the competitive category set for the focused category from the competitive category set storage unit 9, and learns the learned pattern feature vector, the standard feature vector, and the subspace basis vector of the focused category and the competitive category, respectively. It reads from the feature vector storage unit 1, the standard feature vector storage unit 3, and the subspace basis vector storage unit 5 to perform category limited recognition, and stores the recognition result in the category limited learning pattern recognition result storage unit 11.

When the category limited recognition is completed, the subspace basis vector updating unit 12 reads the recognition result of each learning pattern stored in the category limited learning pattern recognition result storage unit 11, counts the number of misrecognized patterns, The misrecognition patterns are taken out one by one and rotated so that the distance between the misrecognition pattern and the misrecognition destination category becomes large, or rotated so that the distance between the misrecognition pattern and the correct answer category becomes small. The generated subspace basis vector is created and transferred to the category limited learning pattern recognition unit 10. The category-limited learning pattern recognition unit 10 uses the transferred subspace basis vector instead of the subspace basis vector stored in the subspace basis vector storage unit 5, and uses the transferred subspace basis vector to classify the category of interest and its competition category. Limited recognition is performed, and the recognition result is stored in the category limited learning pattern recognition result storage unit 11.

The subspace basis vector updating unit 12 reads the recognition result of each learning pattern stored in the category limited learning pattern recognition result storage unit 11, counts the number of erroneous recognition patterns, and reduces the number of erroneous recognition patterns. As long as it continues, the rotation of the subspace basis vector of the focused category and the category limited recognition are repeated. When the number of false recognition patterns starts to increase, the above process is repeated for the next false recognition pattern. When the predetermined learning misrecognition pattern number is reached for the focused category, the subspace basis vector of the focused category is stored in the subspace basis vector storage unit 5.

The standard feature vector updating unit 13 recognizes the category-limited recognition of the category of interest and its competing category for each category of interest.
Ask. The category-limited learning pattern recognition unit 10 refers to the competition category set storage unit 9 to acquire the learning pattern feature vector, the standard feature vector, and the subspace basis vector of the focused category and the competition category, respectively. , Category characteristic recognition is performed by reading from the standard feature vector storage unit 3 and the subspace basis vector storage unit 5, and the category limited learning pattern recognition result storage unit 11 stores the recognition result.

When the category limited recognition is completed, the standard feature vector updating unit 13 reads the recognition result of each learning pattern stored in the category limited learning pattern recognition result storage unit 11, counts the number of misrecognized patterns, and then The recognition patterns were taken out one by one and moved so that the distance between the misrecognized pattern and the misrecognized destination category was increased, or the distance between the misrecognized pattern and the correct answer category was decreased. A subspace basis vector is created and transferred to the category limited learning pattern recognition unit 10. The category-limited learning pattern recognition unit 10 uses the transferred standard feature vector instead of the standard feature vector stored in the standard feature vector storage unit 3 to perform category-limited recognition of the focused category and its competitive category. Execution, category limited learning pattern recognition result storage unit 1
The recognition result is stored in 1.

The standard feature vector updating unit 13 reads the recognition result of each learning pattern stored in the category limited learning pattern recognition result storage unit 11, counts the number of erroneous recognition patterns, and the number of erroneous recognition patterns continues to decrease. As long as the movement of the standard feature vector of the target category and the category-limited recognition are repeated. When the number of false recognition patterns starts to increase, the above process is repeated for the next false recognition pattern. When the number of learning misrecognition patterns of the target category is reached, the standard feature vector of the target category is stored in the standard feature vector storage unit 3.

2 to 4 are flow charts showing the operation of the recognition dictionary creating apparatus shown in FIG. First, the feature vectors of all learning patterns are read (step 21), and the initial recognition dictionary, that is, all categories Cn (n = 1, 2, ...,).
N) initial standard feature vector μn and initial subspace Ln≡
{Ψ ₁ , Ψ ₂ , ..., Ψ _p } (Ψ j: j-th basis vector)
And are created (steps 22 and 23). Step 22,
Recognition processing of all learning pattern data is performed using the initial recognition dictionary created in 23, and the frequency of misrecognition between categories is checked based on the recognition result (step 24), and the competitive category set of each category Cn: Cn ₁ , Cn ₂ , ..., Cns
(S: number of registered competition categories for category Cn) is determined (step 25). Then, the counter n of the category number of interest is set to 0 (step 26), and step 27
The subsequent processing for each category is started.

First, 1 is added to n (step 2
7). Next, the categories of the learning pattern and the recognition dictionary are set to the focused category Cn and the competition categories Cn ₁ , Cn ₂ , ..., Cn.
The recognition process is limited to s (step 28), and the variable E
The number of misrecognized patterns is stored in 1 (step 29). Then, the learning number counter k is set to 0 (step 30),
Learning of misrecognition patterns due to rotation of the subspace is executed (steps 31 to 43).

In the recognition processing of step 28, an arbitrary one pattern is selected as a learning misrecognition pattern from the patterns belonging to the competitive category misrecognized by the focused category and the patterns belonging to the focused category misrecognized by the competitive category. Select (step 31). Next, the rotation angle reference value θ ° is calculated (step 32), the iteration counter m is set to 0 (step 33), and the steps 34 to 40 are repeated using the selected misrecognition pattern. Learning is executed After adding 1 to m (step 34), the rotation angle θ: θ ← θ ° * m / M (M: constant) is calculated (step 35), and the subspace L of the target category is calculated.
Rn (m) is created by rotating n by θ (step 36). Standard feature vector μn, μn ₁ , μn ₂ , ..., μ
ns and subspaces Rn (m), Ln ₁ , Ln ₂ , ..., Ln
The learning pattern of the focused category and its competing category is recognized using s (step 37), and the number of misrecognized patterns is stored in the variable E2 (step 38). The number of erroneous recognition patterns E1 and E2 before and after the subspace rotation is compared (step 3
9) If E2 ≦ E1, false recognition is reduced or leveled, so that a larger rotation angle is tried, so E1 is replaced with E2 (step 40) and the process returns to step 34. E2> E1
If so, since the misrecognition has started to increase, the learning using the misrecognized pattern of interest is terminated. That is, the subspace of the focused category is replaced with the previous subspace Rn (m-1) (step 41), and 1 is added to the learning number counter k (step 42), and whether or not k reaches a predetermined learning number. (Step 43), and if not, step 3
Return to 1. If it has, the process proceeds to step 44 to step 59 to learn the misrecognition pattern by moving the standard feature vector.

First, the initial standard feature vector μn, μ
n _1, μn _2, ..., and μns, subspace Ln which has finished learning by _{rotation, Ln 1, Ln 2, ...} , performs the confirmation process of learning the pattern of the competition category as the target category using a Lns (step 44 ), The number of misrecognized patterns is stored in the variable E1 (step 45). Subsequently, the learning number counter k is set to 0 (step 46), and the misrecognition pattern is learned by moving the standard feature vector from steps 47 to 59.

A pattern erroneously recognized as the focused category Cn and a pattern belonging to the focused category Cn erroneously recognized as another category are selected (step 47). Next, the movement amount reference value τ ° is calculated (step 48), the iteration number counter m is set to 0 (step 49), and the iterations of steps 50 to 56 using the selected misrecognition pattern are repeated. Perform learning.

After adding 1 to m (step 50), the movement amount τ: τ ← τ ° * m / M '(M': constant) is calculated (step 51), and the standard feature vector μn of the target category is τ. Then, νn (m) that has been moved is created (step 52). Then, the standard feature vector νn
(M), μn ₁ , μn ₂ , ..., μns and subspace Ln,
Ln ₁ , Ln ₂ , ..., Lns are used to recognize the learning patterns of the wearing category and its competing categories (step 53), and the number of misrecognized patterns is stored in the variable E 2 (stitch 54).

Number of erroneous recognition patterns E1, before and after the subspace rotation
E2 is compared (step 155), and if E2 ≦ E1,
In order to try a larger movement amount, E1 is replaced with E2 (step 56), and the process returns to step 50. If E2> E1, the misrecognition has started to increase, so the learning using the misrecognized pattern of interest is terminated. That is, the standard feature vector of the focused category is set to the previous standard feature vector νn (m-1).
(Step 57), after adding 1 to the learning number counter k (step 58), it is determined whether k is equal to a predetermined learning number (step 59).
Return to step 47. If they are equal, then it is determined whether or not the focused category number n is equal to the number of categories N (step 60), and if they are equal, the process is terminated. If they are not equal, the learning of the next category is executed, and the process returns to step 27.

[0059]

As described above, according to the present invention, by introducing the parallel movement of the affine subspace of the target category, it is possible to correct the misrecognition pattern which cannot be handled by the rotation of the affine subspace of the target category. Therefore, it is not necessary to rotate the affine subspace of the competitive category with respect to the target category, and the processing amount is reduced. As a result, it is possible to optimize the affine subspace by rotating and translating the affine subspace of the target category and repeating the evaluation of the number of erroneous recognition patterns by learning pattern recognition, and to easily apply it to a pattern recognition dictionary of multiple categories. A recognition dictionary creation device can be realized.

[Brief description of drawings]

FIG. 1 is a block diagram of a pattern recognition dictionary creating apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart showing an operation of the pattern recognition dictionary creating device of FIG.

FIG. 3 is a flowchart showing an operation of the pattern recognition dictionary creating device of FIG.

FIG. 4 is a flowchart showing an operation of the pattern recognition dictionary creating device of FIG.

FIG. 5 is an explanatory diagram of the present invention.

FIG. 6 is an explanatory diagram of a projection distance method.

FIG. 7 is an explanatory diagram of a learning subspace method.

[Explanation of symbols]

 1 learning pattern feature vector storage unit 2 initial standard feature vector creation unit 3 standard feature vector storage unit 4 initial subspace basis vector creation unit 5 subspace basis vector storage unit 6 learning pattern recognition unit 7 learning pattern recognition result storage unit 8 competition category Set determination unit 9 Competitive category set storage unit 10 Category limited learning pattern recognition unit 11 Category limited learning pattern recognition result storage unit 12 Subspace basis vector updating unit 13 Standard feature vector updating unit 21 to 60 steps

Claims (4)

[Claims] 1. A pattern recognition dictionary creating apparatus for determining an affine subspace for each category using a set of feature vectors extracted from a learning pattern, wherein the feature vector of the learning pattern and the affine subspace for each category are Based on the learning pattern recognition means for recognizing the learning pattern by using the distance and the recognition result of the learning pattern set recognized by the learning pattern recognition means, a set of categories with high misrecognition frequency is focused. The competitive category set determining means for determining each category, the category of the affine subspace used for recognition, and the category of the learning pattern to be recognized are limited to the focused category and the competitive category set for the focused category. The category-limited learning pattern recognition means for recognizing, and the category For the learning pattern of the competitive category that has been erroneously recognized as above, the affine subspace is rotated in the direction in which the distance between the feature vector of the learning pattern and the affine subspace of the target category increases, and the category-limited learning pattern recognition means With respect to the learning pattern of the target category that is erroneously recognized as a competitive category by the affine subspace rotating means for rotating the affine subspace in the direction in which the distance between the feature vector of the learning pattern and the affine subspace of the target category becomes smaller. And, regarding the learning pattern of the competitive category that is erroneously recognized as the focused category by the category-limited learning pattern recognition means, the affine part in the direction in which the distance between the feature vector of the learning pattern and the affine subspace of the focused category increases. The space is translated, and the category-specific learning pattern For the learning pattern of the target category that is erroneously recognized as a competitive category by the recognition means, an affine part that translates the affine subspace in a direction in which the distance between the feature vector of the learning pattern and the affine subspace of the target category decreases. A pattern recognition dictionary creating apparatus comprising: a space parallel moving means. 2. The category limited learning pattern recognition means executes the category limited recognition on the focused category and its competing category by using the affine subspace rotated by the affine subspace rotating means, and 2. The pattern recognition dictionary creating apparatus according to claim 1, wherein the affine subspace translated by the affine subspace translation means is used to perform category-limited recognition of the target category and its competitive category. 3. As long as the number of erroneous recognition patterns continues to decrease, the rotation of the affine subspace of the target category by the affine subspace rotating means and the category limited recognition by the category limited learning pattern recognizing means are repeated to obtain the erroneous recognition pattern. 3. The pattern recognition dictionary creating device according to claim 1, wherein when the number starts to increase, the rotation of the affine subspace and the category limited recognition are repeated for the next erroneous recognition pattern. 4. As long as the number of erroneous recognition patterns continues to decrease, the parallel movement of the affine subspace of the target category by the affine subspace parallel movement means and the category limited recognition by the category limited learning pattern recognition means are repeated,
3. The pattern recognition dictionary creating apparatus according to claim 1, wherein when the number of misrecognized patterns starts to increase, the parallel movement of the affine subspace and the category limited recognition are repeated for the next misrecognized pattern.