JP3537949B2 - Pattern recognition apparatus and dictionary correction method in the apparatus - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition apparatus suitable for character recognition, voice recognition, and the like, and a dictionary correction method in the apparatus.

[0002]

2. Description of the Related Art Conventionally, various pattern recognition techniques have been proposed. The conventionally proposed pattern recognition method is a method that uses the knowledge of identification rules that describe the differences between patterns, and a method that uses a large amount of sample data to perform statistical processing without human intervention. This is roughly classified into a statistical pattern recognition method of performing recognition using a recognition dictionary created by the above.

In the statistical pattern recognition method, various characteristic values are extracted from an input pattern, and the extracted characteristic values are arranged into n.
By treating the feature vector as a two-dimensional feature vector and statistically examining the distribution of the feature vector in the n-dimensional feature space, a recognition dictionary created for each category is used, and the feature vector extracted from the input pattern and The recognition result is output based on the evaluation value of the collation result with the dictionary of the category.

[0004] As typical examples of such a statistical pattern recognition method, a subspace method, a pseudo Bayes identification method, and the like are known (Transactions of the Institute of Electronics, Information and Communication Engineers, November 1995).
Vol.J78-D-II No.11 pp.1627-1638).

For example, in the subspace method, the distribution of n-dimensional feature vectors in each category is described in an m-dimensional subspace (m <n), and a dictionary (recognition dictionary) is formed using orthonormal base vectors in the subspace. In this method, a projection value of a vector onto each category subspace is used as an evaluation value and recognition results are output in descending order of the evaluation value. This method is widely applied in the field of pattern recognition such as character recognition and voice recognition because it can well describe pattern fluctuations by a statistical method and achieve high recognition performance.

[0006]

Recently, with the spread of high-performance personal computers, there has been an increasing demand for a technique capable of performing high-speed and high-performance pattern recognition with a small memory capacity by software processing. .

However, in the above-described conventional recognition method, both the memory capacity for the recognition dictionary and the amount of calculation for recognition processing are large. For example, consider the computational complexity when applying the subspace method to Japanese character recognition. First, it is assumed that the number of identification categories is approximately 3000 categories as JIS first-level characters, 256-dimensional features are used as features used for identification, and 16-dimensional is the number of dimensions of a subspace. In this case, to obtain the evaluation value of the input feature vector and the dictionary of one category, that is, the projection amount of the input feature vector onto the subspace, 1
Since it is necessary to calculate the value of the projection onto the 16 orthonormal basis vectors representing the 6-dimensional subspace, 256 × 1
Six product-sum operations are required. For identification, it is necessary to calculate the evaluation value with the dictionary of each category.
56 × 16 × 3000 = 122888000 product-sum operations are required.

Next, consider the recognition dictionary capacity in the above example. Here, since it is necessary to represent each orthonormal base vector of each category, when one element of the vector is represented by, for example, 1 byte, the recognition dictionary capacity is 256 × 16 × 3000 × 1 byte (byte) =
12888000 bytes (bytes) = 11.7 Mby
te (megabytes).

As described above, the conventional method has a problem that high-speed execution cannot be performed without dedicated hardware due to a large amount of calculation for recognition processing. In addition, there is also a problem that the memory capacity required for the recognition dictionary is huge, and the cost is high.

In order to solve such a problem, there is also known a method in which the number of feature dimensions is reduced by feature selection, and then a subspace method or a pseudo Bayes identification method is applied. Magazine, November 1995, Vol.J78-
D-II No.11 pp.1627-1638).

However, in such a case, there is a problem that the recognition performance is reduced as the calculation amount and the dictionary capacity are reduced. The present invention has been made in view of the above circumstances, and has as its object the purpose of requiring a small memory capacity for a recognition dictionary and the like, a small calculation amount for recognition, and achieving the same recognition performance as that of the related art. An object of the present invention is to provide a pattern recognition apparatus and a dictionary correction method in the apparatus.

[0012]

According to the present invention, an n-dimensional feature vector is extracted from an input pattern, and a feature selection dictionary (for example, a set of feature vectors of a large number of learning patterns is extracted from the extracted n-dimensional feature vector). Effective for recognition using a feature selection dictionary created by principal component analysis
Select a dimensional feature vector (m <n) and select the selected m
In a pattern recognition apparatus that performs a recognition process of calculating an evaluation value by comparing a dimensional feature vector with a recognition dictionary including a set of m-dimensional reference vectors and outputting recognition candidates in an order based on the calculation result, the mode, and sequentially inputting a plurality of training patterns, each time perform the above recognition process the learning pattern as the object, based on the result of the recognition process, learning path as a target of the recognition processing
Selected from the n-dimensional feature vectors extracted from the turns
M-dimensional feature vector and the reference vector of the correct category
Is the degree of coincidence with the reference vector of the incorrect category
The smaller the value is, the closer to the first boundary value is.
Loss approaching a second boundary value different from the first boundary value
Detecting the degree of misrecognition using the loss function
Of the misrecognition for both the book and the feature selection dictionary
A series of processes (competition learning process) for correcting by learning so as to reduce the matching are repeated a predetermined number of times.

[0013]

[0014]

In the present invention, the feature amount used for identification is reduced by selecting a feature, so that the capacity of the recognition dictionary and the amount of calculation for recognition can be kept low. By modifying the recognition dictionary and the feature selection dictionary by competitive learning based on the, it becomes possible to perform recognition processing using the corrected recognition dictionary and use the modified feature selection dictionary.
Since it is possible to select a feature effective for identification, it is possible to realize highly accurate recognition performance.

[0016]

[0017]

Embodiments of the present invention will be described below with reference to the drawings. [First Embodiment] FIG. 1 is a block diagram showing a schematic configuration of a pattern recognition apparatus according to a first embodiment of the present invention. The pattern recognition apparatus shown in FIG. 1 is realized using a personal computer or the like that executes software processing for pattern recognition, and includes a data input unit 11, a feature extraction unit 12, a feature selection unit 13, a recognition unit 14, It is composed of functional elements of a recognition result output unit 15, a feature selection dictionary 16, a recognition dictionary 17, and a recognition dictionary correction unit 18. Note that a control unit and the like for controlling the entire apparatus are omitted.

The data input section 11 inputs data (pattern data) to be recognized (character pattern, voice pattern, etc.). The feature extraction unit 12 extracts an n-dimensional feature vector from the data (input pattern) input by the data input unit 11.

The feature selecting unit 13 selects an m-dimensional feature vector (m <n) effective for recognition from the n-dimensional feature vector extracted by the feature extracting unit 12 using a feature selection dictionary 16.

The identification unit 14 calculates an evaluation value by comparing the m-dimensional feature vector selected by the feature selection unit 13 with the recognition dictionary 17, and calculates an evaluation value in the order based on the calculation result (here, the evaluation value Output recognition candidates (in descending order).

The recognition result output unit 15 displays the recognition candidates output from the identification unit 14 on, for example, a display device (not shown). The feature selection dictionary 16 calculates m from the n-dimensional feature vector
Used to select dimensional feature vectors (m <n).

The recognition dictionary 17 is composed of a set of m-dimensional reference vectors used for recognizing m-dimensional feature vectors. The recognition dictionary correction unit 18 detects the degree of erroneous recognition based on the recognition result of the identification unit 14, and corrects the recognition dictionary 17 so that the degree of erroneous recognition is reduced.

Next, the operation of the configuration shown in FIG. 1 will be described with reference to FIGS. In the present embodiment, a keyboard,
A recognition mode for executing a pattern recognition process and a learning process (recognition dictionary correction process) for learning (correcting) the recognition dictionary 17 by a mode designating unit (not shown) realized using input means such as a mouse and a switch. Can be selected and specified.

Hereinafter, (a) recognition processing in the recognition mode,
(B) The learning process (recognition dictionary correction process) in the dictionary learning mode will be described in order. (A) Recognition Processing in Recognition Mode First, the recognition processing when the apparatus in FIG. 1 is set to the recognition mode will be described with reference to the flowchart in FIG.

The data input unit 11 inputs a pattern to be recognized, such as a character pattern or a voice pattern. The feature extraction unit 12 extracts a feature from the pattern input by the data input unit 11 (Step S)
1). Taking character recognition as an example, for example, when a binary character pattern of 15 × 15 pixels is input as shown in FIG. 3, the white pixel is set to “0” and the black pixel is set to “1”, and the upper left corner is set. To the lower right corner (0,
, 0,..., 1, 1,.

The feature selecting unit 13 selects the n-dimensional feature extracted by the feature extracting unit 12 (in the above example of the character pattern,
From n = 15 × 15 = 225), an m-dimensional feature (m <n) required for identification is selected (step S2). In step S2, the name “select” is used. However, instead of selecting m elements from n feature vector elements, the following operation is performed using the feature selection dictionary 16. Is

That is, an input feature (n-dimensional feature extracted by the feature extracting unit 12) is converted into an n-dimensional vector X = (x1,
x2,..., xn) ^T (where T is a symbol representing transposition) and the feature selection dictionary 16 is represented by P using an m × n matrix, and in step S2, the following equation X ′ = PX. According to (1), an m-dimensional feature X ′ is selected from an n-dimensional feature X.

The feature selection dictionary 16 (= P) is designed, for example, by the following procedure using feature vectors extracted from a large number of learning patterns. First, the set of n-dimensional feature vectors used for creating a dictionary is {X1, X2,..., XN}. From this, an n × n matrix K is calculated according to the following equation (2).

[0029]

(Equation 1)

Next, the eigenvectors of the matrix K are defined as φ1, φ2,... In descending order of the corresponding eigenvalues, and an m × n matrix P defined by P = (φ1 φ2... Φm) ^T (3) is selected. The dictionary is assumed to be 16.
The eigenvectors of the matrix K express the first axis, the second axis,... Of the distribution of the feature vector set in descending order of the corresponding eigenvalues.

Here, from the above equation (1), X '= (x'
1, x'2, ..., x'n) ^T , x'i = (X, φi) ... (4) Since x'i is the inner product of X and φi, The feature selection dictionary 16 (=
The feature selection by P) has the meaning of projecting a set of learning feature vectors onto a principal component space. If the eigenvalue corresponding to the eigenvalue φi is λi, the m × n matrix P (feature selection dictionary 16) may be expressed by the following equation.

[0032]

(Equation 2)

This m × n matrix P (feature selection dictionary 16)
Is a normalized value of the variance of the projection value of the learning pattern on each principal axis. By the way, the feature selection unit 13 described above in (1)
M-dimensional feature (feature vector) X 'selected according to the equation
Is passed to the identification unit 14. The identification unit 14 performs an identification process of outputting a recognition candidate by collating the selected feature vector X ′ with the recognition dictionary 17 and classifying the feature vector X ′ into a category (step S3). The details of step S3 are as follows.

First, the recognition dictionary 17 stores, as a reference vector for each category, a vector representative of the category as one vector.
M-dimensional reference vector set {R1, R2,..., RM in all categories (M is the number of categories or more)
Consists of｝.

In the case of such a structure of the recognition dictionary 17, the identification unit 14 includes the m-dimensional feature vector X ′ selected by the feature selection unit 13 and each reference vector Ri (i = 1 to 1) constituting the recognition dictionary 17. The distance d (X ', Ri) to the distance M) is calculated according to the following equation.

D (X ′, Ri) = (X′−Ri) ² (6) Next, the discriminating unit 14 calculates the order of the distance calculated according to the above equation (6) in ascending order (that is, the vector X ′, Ri). The category to which the corresponding reference vector Ri belongs belongs to the recognition result output unit 15 as a recognition candidate.
Output to

The recognition result output unit 15 receives and outputs the recognition candidates output from the identification unit 14 to a display device (not shown). The above is the procedure of the recognition processing. As is apparent from this recognition processing, the capacity of the feature selection dictionary 16 is proportional to the number of dimensions of the selected feature vector X '. Since the number of dimensions of the reference vector Ri matches the number of dimensions of the selected feature vector X ', the capacity of the recognition dictionary 17 is also proportional to the number of dimensions of the selected feature vector X'.

On the other hand, the amount of calculation required for the recognition process (the number of product sum operations) is determined by the feature selecting unit 13 and the discriminating unit 14. The feature selecting unit 13: (the number of dimensions of the feature vector) × (the number of dimensions of the selected feature vector) ) Times discriminator 14: (number of total reference vectors) × (number of dimensions of selected feature vector) times, which is also proportional to the number of dimensions of the selected feature vector.

As described above, as the number of dimensions of the selected feature vector is reduced, the capacity of the recognition dictionary 17 (dictionary capacity) can be reduced, and the recognition process can be performed with a small amount of calculation (computation). Here, the actual number of dimensions of the selected feature vector is determined by the dictionary capacity required for the apparatus and the restriction on the amount of calculation.

As is apparent, the lower the number of dimensions of the selected feature vector is, the more advantageous in terms of dictionary capacity and calculation amount, but on the other hand, the amount of information in the recognition dictionary 17 is reduced, so that the recognition performance is reduced. Is expected.

Therefore, in the present apparatus, the recognition performance is improved by introducing competitive learning in which the recognition dictionary 17 described below is modified so that the learning patterns are actually recognized and erroneous recognition is reduced as much as possible. I'm trying. (B) Learning Process in Dictionary Learning Mode (Recognition Dictionary Correction Process) Hereinafter, the learning process when the apparatus in FIG. 1 is set to the dictionary learning mode will be described with reference to the flowchart in FIG.

First, as described above, a vector representing a category, which is called a reference vector for each category, is 1
A recognition dictionary 17 in an initial state is prepared, which includes at least one m-dimensional reference vector set {R1, R2,..., RM} in all categories. Each reference vector forming the recognition dictionary 17 in the initial state is designed, for example, as an average vector of selected feature vectors of a plurality of learning patterns belonging to each category.
The number of reference vectors in each category may be any number as long as it is one or more. In the case of a plurality of reference vectors, all initial values may be the same as the average vector.

When the dictionary learning mode is set in the apparatus shown in FIG. 1, the control unit sequentially inputs all learning patterns registered in advance in an external storage device such as a magnetic disk device through the data input unit 11. In each case, a series of operations of actually recognizing the learning pattern (in the procedure shown in the flowchart of FIG. 2) and correcting the recognition dictionary 17 based on the recognition result are performed according to the flowchart of FIG. The process is repeated for the target number of learning times (target learning number).

That is, after setting the counter value t for counting the number of times of learning to the initial value 0 (step S11), if the counter value t has not reached the target number of times of learning (the specified number of times of learning) (step S12). Then, a counter value i for counting a learning pattern to be recognized is set to an initial value 0 (step S13). Since the counter value i has not reached the predetermined (designated) number of learning patterns (step S14), the i-th learning pattern (i-th learning pattern) is given to the apparatus of FIG. (Step S1).
5) Based on the recognition result, the recognition dictionary correction unit 18 corrects the recognition dictionary 17 (step S16). Details of the correction processing in step S16 will be described later.

When the step S16 is completed, the counter value i is incremented by 1 (step S17). Thereafter, the processing after the step S14, that is, the recognition processing for the next learning pattern, and the recognition using the result of the recognition processing are performed. Dictionary 1
7 is performed.

Eventually, when the recognition process for the learning patterns for the predetermined number of learning patterns and the correction process of the recognition dictionary 17 using the result of the recognition process are all performed, the process proceeds from step S14 to step S18. , The counter value t is incremented by one. If the counter value t after +1 (the number of learnings t actually performed) has not reached the target number of learnings (step S12), the above-described step S1 is performed.
The processes after 3 are performed again.

Eventually, when the counter value t reaches the target number of times of learning, a series of learning processing according to the flowchart of FIG. 4 ends. Here, the recognition dictionary correction processing in step S16 by the recognition dictionary correction unit 18 is performed in step S15.
Is defined by defining a loss due to erroneous recognition when a learning pattern is recognized, and correcting the reference vector in a direction to reduce the loss. The details are as follows.

First, in the present embodiment, when the selected feature vector X'k of a certain learning pattern is recognized, the reference vector that is the highest candidate among the reference vectors of the same category (correct category) as the learning pattern is determined. Ri is the loss function h (X'k) when X'k is recognized, where Rj is the reference vector that was the highest candidate among the reference vectors of the category (incorrect answer category) different from the learning pattern. Is defined as follows.

H (X′k) = f (g (X′k)) (7) g (X′k) = d (X′k, Ri) −d (X′k, Rj) (8) ) F (x) = 1 / {1 + exp (-αx)} (α> 0) (9) where d (X′k, Ri), d (X′k, Rj)
Is a distance function defined by the above equation (6), and the former is the matching degree with the top candidate among the reference vectors of the same category as the learning pattern (the matching degree decreases as the distance increases).
And the latter indicates the degree of coincidence with the highest-order candidate among reference vectors of a category different from the learning pattern. The function f (x) is a sigmoid function as shown in FIG.

The loss function h (X 'defined as above)
k) has a smaller value (here, a value closer to 0) as the distance between the selected feature vector X'k of the learning pattern and the reference vector of the correct answer category is smaller than the distance from the incorrect answer category. Since it is a large value (a value close to 1 in this case), it is clear that the evaluation function indicates the degree of erroneous recognition.

Thus, the average L of the loss function h (X′k) for the selected feature vector {X′k | k = 1,..., N} of all the learning patterns is calculated by the following equation (10). , It can be said that the smaller the value L, the better the identification system.

[0052]

[Equation 3]

However, it is difficult to analytically find a reference vector that minimizes L (function L) shown in the above equation (10). Therefore, in the present embodiment, the reference vector is corrected (updated) little by little by using the well-known steepest gradient method (steepest descent method) in the recognition dictionary correction processing in the recognition dictionary correction unit 18, and the minimum solution is obtained. I want to ask.

Specifically, the selected feature vector X'k of the learning pattern is recognized by the identification unit 14, and the reference vector Ri and the highest candidate among the reference vectors of the same category (correct category) as the learning pattern are used. , The value obtained by differentiating the loss function h (X'k) with the reference vectors Ri and Rj from the reference vector Rj, which is the highest candidate among the reference vectors of the category (incorrect answer category) different from the learning pattern, that is, the reference vector Loss function h in space
Using the gradient of (X'k), the recognition dictionary correction unit 1 is directed in the direction in which the loss function h (X'k) decreases according to the following rule.
Correct (update) little by little at 8.

[0055]

(Equation 4)

Note that ε (t) in the above rule is for determining the learning speed, and is a counter value t which takes a positive value.
This is a decreasing function of (the number of times of learning pattern presentation t), for example, ε (t) = 1 / (t + 10).

By repeating the modification of the recognition dictionary 17 by the specified number of times according to such a rule, it is possible to correct the recognition dictionary 17 with a small loss due to erroneous recognition, that is, a recognition dictionary 17 with little erroneous recognition and good recognition performance.

A line graph showing the recognition performance (recognition rate) with respect to the amount of calculation (recognition dictionary capacity) when handwritten character recognition is performed using the recognition dictionary 17 modified as described above in the apparatus of FIG. Is shown in FIG. 6 in comparison with the subspace method which is a conventional method. The recognition target here is katakana characters.

In FIG. 6, the horizontal axis represents the ratio of the sum-of-products operation amount when the sum-of-products operation amount of the subspace method based on the 16-dimensional subspace of the 256-dimensional feature (conventionally known) is set to 1. And on a logarithmic scale. The ratio of the product-sum operation amount is the same even if it is considered as the ratio of the recognition dictionary capacity. On the other hand, the vertical axis represents the recognition rate.

A line graph indicated by reference numeral 61 in FIG. 6 represents the recognition performance of the apparatus according to the present embodiment.
When a dimensional feature, a 64-dimensional feature, and a 128-dimensional feature are selected, the recognition result (recognition rate) when one reference vector is recognized for each category is connected by a line.

Next, the line graph indicated by reference numeral 62 shows the recognition performance of a method in which a 256-dimensional feature for which no feature selection is performed is used as it is and the subspace method is applied to the 256-dimensional feature. And the subspace dimension is 1
The recognition results in the case of two-dimensional, two-dimensional, four-dimensional, eight-dimensional, and sixteen-dimensional are connected by a line.

Similarly, the line graphs denoted by reference numerals 63, 64, 65, and 66 respectively select a 16-dimensional feature, a 32-dimensional feature, a 64-dimensional feature, and a 128-dimensional feature, and apply a subspace to the selected feature. It shows the recognition performance of the method (feature selection + subspace method) to which the method is applied.
The recognition results in the case of six dimensions are connected by lines. In the examples of the graphs 63 to 66, although the amount of product-sum operation and the capacity of the recognition dictionary are reduced by performing feature selection as compared with the example of the graph 62, the recognition dictionary is modified as in the present embodiment. No recognition performance is obtained.

That is, comparing the conventionally known subspace method with the feature selection + subspace method, the feature selection +
It can be seen that, in the subspace method, for example, even if a 64-dimensional feature is selected and the product-sum operation amount is reduced to ４, the same recognition performance as the subspace method can be obtained. However, if the number of selected dimensions is further reduced, the recognition performance is significantly reduced.

Next, a comparison between the subspace method and the method according to the present embodiment shows that the method according to the present embodiment has the same degree of recognition as the subspace method even when the product-sum operation amount is about 1/30. It can be seen that performance is obtained.

As described above, from FIG. 6, the method according to the present embodiment reduces the amount of calculation required for the recognition processing by selecting the feature and corrects the recognition dictionary 17 by learning, thereby achieving high recognition accuracy. It can be read that it can be maintained, and the effect of the method in the present embodiment can be confirmed. [Second Embodiment] FIG. 7 is a block diagram showing a schematic configuration of a pattern recognition device according to a second embodiment of the present invention. The configuration of FIG. 7 is different from the configuration of FIG.
4, a feature selection dictionary correction unit 19 that detects the degree of false recognition based on the recognition result and corrects the feature selection dictionary 16 so as to reduce the degree of false recognition is added.
In the dictionary learning mode, the configuration differs from that of FIG. 1 in that not only the recognition dictionary 17 but also the feature selection dictionary 16 is corrected. That is, in the configuration of FIG. 1, the feature selection dictionary 16 created by the principal component analysis is used as it is, but in the configuration of FIG. 7, by modifying the feature selection dictionary 16 by learning, it is more advantageous for identification. Feature selection is possible, and the recognition performance is further improved. Note that the recognition process in the recognition mode is performed according to the flowchart of FIG. 2 as in the first embodiment.

The operation in the dictionary learning mode (learning processing) will be described below with reference to the flowchart of FIG. When the dictionary learning mode is set in the device of FIG. 7, the control unit causes the data input unit 11 to sequentially input all the learning patterns registered in advance in the external storage device, and each time the learning pattern is set as shown in FIG. The operation of actually recognizing and revising the recognition dictionary 17 and the feature selection dictionary 16 based on the recognition result is repeated by the target number of times of learning according to the flowchart of FIG. Steps S21 and S2
7). The difference from the first embodiment is that the modification processing step S16 corresponds to step S16 in the first embodiment.
26, in that not only the recognition dictionary 17 but also the feature selection dictionary 16 is modified.

The recognition dictionary 17 in this step S26
The correction of the feature selection dictionary 16 is performed by defining a loss function h (X'k) = h (PXk) similar to that of the first embodiment, and steepest gradient method (steepest descent method) in the direction of decreasing the loss function. )
Is performed as follows by modifying the feature selection dictionary 16. However, it is assumed that the feature selection dictionary 16 (= P) in the initial state applied in the present embodiment is created by principal component analysis as in the case of the first embodiment. Also. It is assumed that the recognition dictionary 17 in the initial state is created by principal component analysis using the feature selection dictionary 16 in the initial state as in the case of the first embodiment.

[0068]

(Equation 5)

Here, the value obtained by differentiating the loss function h (X'k) = h (PXk) with the feature selection dictionary (feature selection dictionary matrix) P, that is, the feature selection dictionary, is obtained by the equations (15) and (18). Loss function h (X'k) in reference parameter space
= H (PXk), the feature selection dictionary P, that is, the feature selection dictionary 16 is modified.

On the other hand, the correction of the reference vector Ri (Equation (1)
6), equation (19)) and correction of reference vector Rj (equation (1)
7) and Equation (20) are the correction of the reference vector Ri (Equations (11) and (13)) and the correction of the reference vector Rj ((Equations (12) and (14)) in the first embodiment. ).

By repeating the modification of the recognition dictionary 17 and the feature selection dictionary 16 by such a rule a specified number of times, it is possible to correct the recognition dictionary 17 and the feature selection dictionary 16 with a small loss due to erroneous recognition.

In the apparatus shown in FIG. 7 (the recognition method of this embodiment),
The recognition performance when similar characters are recognized (identified) using the recognition dictionary 17 and the feature selection dictionary 16 corrected as described above is evaluated by the apparatus of FIG. 1 (the recognition method of the first embodiment). FIG. 9 shows a comparison between the recognition performance of the conventional method and the recognition performance of the conventional method of feature selection + subspace method. In this example, it is assumed that a 64-dimensional feature is selected by feature selection. In the subspace method, a three-dimensional subspace is used. In the first embodiment and the present embodiment (second embodiment), the number of reference vectors is set to three for each category, and comparison is performed with the same amount of calculation. ing.

As is clear from FIG. 9, the recognition method of the second embodiment can realize higher recognition performance than the first embodiment, as well as the conventional method. Note that in the dictionary learning mode in the embodiment (first and second embodiments),
A plurality of learning patterns prepared in advance are input one by one, and each time the learning patterns are actually recognized, a dictionary (a recognition dictionary 17 or a recognition dictionary 17 and a feature selection dictionary) is created based on the recognition result. Although a series of operations of correcting 16) is repeated for the target number of times of learning, the invention is not limited to this. For example, for one learning pattern, when the learning pattern is input and actually recognized, and the operation of correcting the dictionary based on the recognition result is repeated by the target number of learning times, the learning pattern is switched to the next learning pattern. No problem. However, in this method,
If the learning process of the dictionary using one learning pattern is not repeated for the target learning number, the dictionary cannot be switched to the next learning pattern. Therefore, the dictionary after the series of learning processes (the recognition dictionary 17 or the recognition dictionary 17 and the In the selection dictionary 16),
There is a possibility that the learning result for (the category of) the learning pattern used in the early stage of the series of learning processing is not reflected. Therefore, the learning effect can be enhanced by performing the learning process according to the procedure applied in the embodiment.

[0074]

As described above in detail, according to the present invention, the feature amount used for identification is reduced by selecting a feature, so that the capacity of the recognition dictionary and the amount of calculation for recognition can be reduced, and learning can be performed. since so as to modify both the recognition dictionary and feature selection dictionary by competitive learning based on the recognition result for the pattern recognition processing can and Do Rutotomoni, modified feature selection using the modified recognition dictionary
It is now possible to select effective features for identification using a dictionary.
Thus, highly accurate recognition performance can be realized.

[0075]

[Brief description of the drawings]

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

FIG. 2 is a flowchart for explaining recognition processing when the apparatus in FIG. 1 is set to a recognition mode.

FIG. 3 is a view for explaining feature vector extraction from an input character pattern.

FIG. 4 is a flowchart for explaining a recognition dictionary learning process when the apparatus in FIG. 1 is set to a dictionary learning mode.

FIG. 5 is a diagram illustrating a sigmoid function used to define a loss function.

FIG. 6 is a line graph showing recognition performance (recognition rate) with respect to the amount of calculation (recognition dictionary capacity) when handwritten character recognition is performed using the recognition dictionary 17 corrected by the learning process in the apparatus of FIG. The figure shown in comparison with the subspace method which is a conventional method.

FIG. 7 is a block diagram showing a schematic configuration of a pattern recognition device according to a second embodiment of the present invention.

8 is a flowchart for explaining a learning process of a recognition dictionary and a feature selection dictionary when the apparatus of FIG. 7 is set to a dictionary learning mode.

9 shows the recognition performance when similar characters are recognized (identified) using the recognition dictionary 17 and the feature selection dictionary 16 corrected by the learning process in the apparatus of FIG.
FIG. 9 is a diagram showing the recognition performance in the recognition method according to the embodiment and the recognition performance in the feature selection + subspace method which is a conventional method.

[Explanation of symbols]

11 ... data input part, 12 ... feature extraction unit, 13 ... Feature selection unit, 14 ... identification part, 15 ... Recognition result output unit 16. Feature selection dictionary, 17 ... Recognition dictionary, 18. Recognition dictionary correction unit, 19: Feature selection dictionary correction unit.

Continuation of the front page (56) References JP-A-63-177285 (JP, A) JP-A-7-225836 (JP, A) JP-A-4-256087 (JP, A) JP-A-6-251158 (JP) , A) Yonezawa, Y., et al., Formulation of a context effect model by minimum classification error learning, IEICE Technical Report SP94-108-117, March 24, 1995, Vol. 94, No. 569, pp. 47-54 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06K 9/00-9/82 G10L 15/06 G06T 7/00-7/60

Claims (2)

(57) [Claims] 1. Inputting pattern data to be recognized
Inputting means, extracting an n-dimensional feature vector from the pattern data input by the inputting means , extracting an n-dimensional feature vector, and extracting an n-dimensional feature vector into an m-dimensional feature vector (m <n).
A feature selection dictionary used for selecting a feature vector; a recognition dictionary consisting of a set of m-dimensional reference vectors used for recognition of m-dimensional feature vectors; and the feature selection from the n-dimensional feature vectors extracted by the feature vector extracting means. M that is effective for recognition using a dictionary
A feature selection unit that selects a dimensional feature vector; an evaluation value is calculated by comparing the m-dimensional feature vector selected by the feature selection unit with the recognition dictionary; and recognition candidates are output in an order based on the calculation result. Identifying means and pattern data to be recognized in the dictionary learning mode
To input a learning pattern by the input means
And an n-dimensional feature vector extracted from the learning pattern based on the output result of the identification means in the dictionary learning mode.
M-dimensional feature vector selected from the
The degree of coincidence with the reference vector is
The closer to the first boundary value the greater the degree of agreement with the
On the contrary, the smaller the second boundary value is, the smaller the second boundary value is.
Detecting the degree of misrecognition using a loss function approaching the boundary value, for both the recognition dictionary and the feature selection dictionary
Dictionary correction processing by learning to reduce the degree of misrecognition
A pattern recognition device comprising: a dictionary correction unit that executes a process . 2. An n-dimensional feature vector from an input pattern
From the extracted n-dimensional feature vector.
M-dimensional feature vectors (m <
n) and select the selected m-dimensional feature vector and m-th order
Collation with a recognition dictionary consisting of a set of source reference vectors
The evaluation value is calculated with
Pattern recognition device that executes recognition processing to output knowledge candidates
In the dictionary learning mode in which a plurality of learning patterns are sequentially input.
Each time, the recognition process is performed for the learning pattern.
Performing recognition processing, and each time the recognition processing is performed, based on the result of the recognition processing.
Extracted from the learning pattern targeted for the recognition process
M-dimensional feature vector selected from the obtained n-dimensional feature vector
The match between the reference vector and the correct category reference vector is incorrect.
The greater the degree of coincidence with the reference vector of the solution category,
The first boundary value approaches the first boundary value, and conversely, the first boundary value
Error using a loss function approaching a second boundary value different from the
Detecting the degree of recognition and selecting the recognition dictionary and the feature selection
Learning to reduce the degree of misrecognition for both dictionaries
Performing a dictionary correction process by learning, performing the recognition process, and performing the dictionary correction process
The plurality of learning patterns
Is repeated a predetermined number of times.
Control step of controlling
Dictionary correction method in turn recognition device.