JP2994533B2 - Pattern recognition device - 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 device, and more particularly, to an LVQ (Learning Vector Quantizer).
option: learning vector quantization), and based on the recognition result of the sample pattern, invalidate, add,
The present invention relates to a pattern recognition device that creates a dictionary pattern that is nearly optimal by controlling the validation and enables online learning.

[0002]

2. Description of the Related Art Finnish Kohonen (Teuvo Koho)
nen) 's LVQ (LearningVector Quantizati)
According to a pattern learning method called “on: learning vector quantization”, one or more dictionary patterns are prepared for each category to be recognized, and all dictionary patterns are input pattern x belonging to category C _x. The closest dictionary pattern m ₁ and its category C ₁ are selected. When C _x ≠ C ₁ m ₁ ← m ₁ −α (x−m ₁ ) (1) When C _x = C ₁ m ₁ ← m ₁ + α (x−m ₁ ) The dictionary pattern m ₁ is updated based on (2). Normally, m ₁ is further normalized, and the norm is set to 1. Here, α is a coefficient that satisfies 0 <α <1, and usually decreases as the learning progresses.

[0003] Kohonen describes L, an improved version of LVQ.
VQ2 has published, according to the category C with respect to the input pattern x belonging to _x, every dictionary pattern in the nearest dictionary pattern m ₁ and the category C _1, m ₁ to followed close dictionary pattern m ₂ And its category C _2, and “x” is a “window” provided near the bisector of m ₁ and m ₂
, And C _x ≠ C ₁ and C _x =
When the _{C 2, m 1 ← m 1} -α (x-m 1) ... (3) m 2 ← m 2 + α (x-m 2) ... based on (4), the dictionary pattern m _1, m ₂ Perform an update. Normally, m ₁ and m ₂ are normalized, and the norm is 1 for each.
Keep it. Here, α is a coefficient similar to Expressions (1) and (2). The determination as to whether or not x is located within the “window” is made by, for example, setting the distance between m ₁ and x to d ₁ and the distance between m ₂ and x to d ₂ .

[0004]

(Equation 1)

[0005] At this time, it may be performed assuming that it is located in the "window". Here, s is an appropriately determined threshold value.

Further, Kohonen describes LVQ, LVQ
Has published a variant of 2 LVQ3, according to this, the input pattern x belonging to the category C _x, all the dictionary pattern in closest dictionary pattern m ₁ and its category C _1, m ₁ then picked out and close the dictionary pattern m ₂ and the category C _2, when x is located in the "window" which defines as in the LVQ2, when · C _x ≠ C ₁ and C _x = C ₂ m ₁ ← m ₁ −α (x−m ₁ ) (6) m ₂ ← m ₂ + α (x−m ₂ ) (7) When C _x = C ₁ and C _x ≠ C ₂ , m ₁ ← m ₁ + Α (x−m ₁ ) (8) m ₂ ← m ₂ −α (x−m ₂ ) (9) When C _x = C ₁ and C _x = C ₂ , m ₁ ← m ₁ + εα (x −m ₁ ) (10) m ₂ ← m ₂ + εα (x−m ₂ ) (11) The dictionary patterns m ₁ and m ₂ are updated based on (11). Normally, m ₁ and m ₂ are normalized, and the norm is set to 1. Here, α is a coefficient similar to Expressions (1) and (2). ε is a coefficient that is appropriately determined, and may take a value of about 0.1 to 0.5.

[0007] LVQ, LVQ2, LV as described above
According to Q3, it is known that a pattern recognition device having a high recognition rate can be obtained. Also, "Multi-font recognition by hierarchical dictionary arrangement" (Kato, Takahashi, IEICE Transactions, Vol.J74-D-II, No.1, p8-18)
Dictionary patterns are added and dictionary patterns are exchanged between the main storage and the secondary storage to optimize the number of dictionary patterns in the main storage. The pattern on the secondary storage is not used during operation as a recognition device. At the time of learning, first, erroneous recognition patterns are selected by recognizing the sample data using all dictionary patterns including the main storage and the secondary storage. An averaging operation of the misrecognition patterns creates an additional pattern, and determines whether or not to actually add this to the dictionary based on the distance between the pattern and an existing dictionary pattern. Further, for all dictionary patterns including the main storage and the secondary storage including these additional patterns, all sample data are recognized, and based on the recognition result,
The main memory and the secondary memory are exchanged.

[0008]

As described above, in the conventional pattern recognition apparatus, for example, in the case of character recognition,
Character categories with many similar character categories,
It is known that the optimal number of dictionary patterns differs from the other character categories. However, in the above-described LVQ, LVQ2, and LVQ3, an operation for increasing or decreasing the number of dictionary patterns in each category is not performed.
The number had to be set separately. Also, since additional registration of the dictionary by the user or automatic writer adaptation or speaker adaptation is not taken into account, dictionary pattern adjustment is repeatedly performed when these are performed, and the dictionary pattern converges. Up to a large number of iterations. In the method of the above-mentioned literature, it is necessary to recognize all sample patterns at the time of learning, and it is difficult to perform frequent learning because the main memory and the secondary memory are exchanged. A storage device for storing all sample data is essential. Furthermore, since the additional pattern is created by the averaging operation, it is considered that the effectiveness in improving the recognition rate is inferior to the method based on the pattern adjustment based on the recognition result such as LVQ.

The present invention has been made in view of such circumstances. Since an operation of increasing or decreasing the number of dictionary patterns in each category is performed, a nearly optimal number of dictionary patterns can be stored for each category. And also
When registering additional dictionaries by the user or performing automatic writer adaptation and speaker adaptation, for new patterns,
Depending on the validity or validity flag, it is possible to handle the pattern differently from the already registered patterns so that the dictionary can be imported without repeating the dictionary pattern adjustment. Pattern recognition to improve the recognition rate compared to the case of averaging operation, and to make the processing more efficient when using the valid flag compared to replacing the dictionary pattern with realization It is intended to provide a device.

[0010]

In order to achieve the above object, the present invention provides (1) a pattern input means for inputting a pattern such as character and voice data, and a pattern distance for calculating a distance between the patterns. Calculation means, dictionary pattern storage means for storing a variable number of dictionary patterns for each category to be recognized, effectiveness storage means for storing the effectiveness of each dictionary pattern, in a set of dictionary patterns, Proximity dictionary pattern selection means for selecting a dictionary pattern and a category in which the distance to a certain pattern is ordered in a certain order based on the distance between the patterns, and for the input pattern having a known category, the proximity dictionary pattern selection means A dictionary pattern is selected based on the input pattern, and the category to which the selected dictionary pattern belongs belongs to the category of the input pattern. A validity adjusting means for increasing or decreasing the validity of the dictionary pattern based on whether the dictionary pattern is the same as a category, and a dictionary pattern based on the proximity dictionary pattern selecting means for an input pattern with a known category. A dictionary pattern deleting unit for selecting and selecting the dictionary pattern based on the validity of the selected dictionary pattern; or (2) a pattern input unit for inputting a pattern such as character / voice data; Inter-pattern distance calculating means for calculating the distance between the patterns, dictionary pattern storing means for storing a variable number of dictionary patterns for each category to be recognized, and validity storage for storing the validity of each dictionary pattern Means and, in a set of dictionary patterns, ordering a certain distance to a certain pattern based on a distance between the patterns. A proximity dictionary pattern selecting means for selecting a dictionary pattern and a category, and a dictionary pattern based on the input pattern having a known category, based on the proximity dictionary pattern selecting means, and a category to which the selected dictionary pattern belongs. A validity adjusting means for increasing or decreasing the validity of the dictionary pattern based on whether or not the category belongs to the input pattern; and a proximity dictionary pattern selecting means for an input pattern with a known category. Based on whether the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs, based on the input pattern or a pattern created based on the input pattern, A dictionary pattern adding means for adding a dictionary pattern. Or (3) a pattern input unit for inputting a pattern such as character / voice data, an inter-pattern distance calculating unit for calculating a distance between the patterns, and a variable number of dictionary patterns for each category to be recognized. Dictionary pattern storing means for storing, validity storing means for storing the validity of each dictionary pattern, valid flag storing means for storing a valid flag that can take an on / off state for each dictionary pattern, dictionary Within a set of patterns, the elements of the set can be selected based on the valid flag, and based on the distance between the patterns, a dictionary pattern such that the distance to a certain pattern is ordered in a certain order. A proximity dictionary pattern selecting means for selecting a category, and a proximity dictionary pattern for an input pattern having a known category. A dictionary pattern is selected based on the selection unit, and the validity adjustment for increasing / decreasing the validity of the dictionary pattern based on whether the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs. Means, and dictionary pattern invalidating means for turning off the validity flag of the dictionary pattern based on the validity of the dictionary pattern.
In the above (3), there is provided a dictionary pattern validating means for turning on a valid flag of the dictionary pattern based on the validity of the certain dictionary pattern. Pattern inputting means for inputting, inter-pattern distance calculating means for calculating the distance between the patterns, dictionary pattern storing means for storing a variable number of dictionary patterns for each category to be recognized, and validity of each dictionary pattern Validity storing means for storing the validity flag, validity flag storing means for storing a validity flag that can be turned on / off for each dictionary pattern, and in the set of dictionary patterns, the elements of the set are the validity flag. It is possible to select based on
A proximity dictionary pattern selecting means for selecting a dictionary pattern and a category in which the distance to a certain pattern is ordered in a certain order based on the distance between the patterns, and selecting the proximity dictionary pattern for an input pattern having a known category. Means for selecting a dictionary pattern based on the means, and increasing or decreasing the effectiveness of the dictionary pattern based on whether or not the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs. And dictionary pattern validating means for turning on a valid flag of the dictionary pattern based on the validity of the dictionary pattern. (6) The method of (1), (3),
In (4) or (5), a dictionary pattern is selected based on the proximity dictionary pattern selection means for a known input pattern of the category, and the category to which the selected dictionary pattern belongs belongs to the input pattern. A dictionary pattern adding means for adding the input pattern or a pattern created based on the input pattern as a dictionary pattern based on whether or not the category is the same as the category.

[0011]

According to the pattern recognition apparatus of the present invention, a nearly optimal number of dictionary patterns is stored for each category for the creation and adjustment of the pattern dictionary, and the addition of a new pattern is realized while suppressing a reduction in recognition rate. . In addition to dictionary pattern adjusting means based on LVQ or the like, the system includes dictionary pattern invalidating means, dictionary pattern adding means, and dictionary pattern validating means. These control invalidation, validation and addition of the dictionary pattern based on the recognition result of the sample pattern. As described above, since the operation of increasing or decreasing the number of dictionary patterns in each category is performed, a nearly optimal number of dictionary patterns can be stored for each category. A new pattern can be taken into a dictionary without repeatedly performing dictionary pattern adjustment, and the present invention can be applied to additional registration of a dictionary by a user or automatic writer adaptation and speaker adaptation.

[0012]

Embodiments will be described below with reference to the drawings. FIG. 1 is a block diagram for explaining an embodiment of a pattern recognition apparatus according to the present invention. In the figure, 1 is a pattern input unit, 2 is a feature extraction unit, 3 is an inter-pattern distance calculation unit, and 4 is a dictionary pattern. Storage means, 5 is an effectiveness storage means,
6 is a proximity dictionary pattern selecting means, 7 is a recognition result creating means, 8 is a recognition result presenting means, 9 is a correct answer category selecting means, 10 is a dictionary pattern adjusting means, 11 is a validity adjusting means, 12 is a dictionary pattern deleting means, Reference numeral 13 denotes a dictionary pattern adding unit.

A feature required for recognition is extracted by a feature extracting means 2 from a pattern input by the pattern input means 1. The inter-pattern distance calculation means 3 calculates the distance between the patterns, and the dictionary pattern storage means 4 stores a variable number of dictionary patterns for each category to be recognized. The validity storage means 5 stores the validity of each dictionary pattern, and the proximity dictionary pattern selecting means 6 sets the order of the distance to a certain pattern based on the distance between the patterns in a set of dictionary patterns. Select dictionary patterns and categories that can be ordered. The recognition result creating means 7 creates a recognition result to be finally presented to the user from the recognition result obtained by the proximity dictionary pattern selecting means 6. The recognition result presenting means 8 presents the recognition result by the recognition result creating means 7 to the user. The correct answer category selecting means 9 determines the category to which the input pattern should belong based on the operation of the user based on the operation of the user based on the recognition result presented by the recognition result creating means 7.

The dictionary pattern adjusting means 10 selects a dictionary pattern for the input pattern of a known category based on the proximity dictionary pattern selecting means 6, and changes the dictionary pattern based on the input pattern. The validity adjusting unit 11 selects a dictionary pattern based on the proximity dictionary pattern selecting unit 6 for an input pattern whose category is known, and the category to which these dictionary patterns belong is the same as the category to which the input pattern belongs. The validity of the dictionary pattern is increased or decreased based on whether or not the dictionary pattern exists.

The dictionary pattern deletion means 12 selects a dictionary pattern from the input pattern of a known category based on the proximity dictionary pattern selection means 6, and deletes the dictionary pattern based on the validity of these dictionary patterns. delete. The dictionary pattern adding means 13 selects a dictionary pattern for the input pattern of a known category based on the proximity dictionary pattern selecting means 6, and the category to which these dictionary patterns belong is the same as the category to which the input pattern belongs. Based on whether or not there is a pattern, a pattern created based on the input pattern is added as a dictionary pattern.

In the following, in the online character recognition of the in-frame writing, each part will be exemplified by performing online learning of a stroke pattern (time locus of xy coordinates on a tablet plane of a pen tip corresponding to each stroke of a character). A method for realizing is described. The pattern input means 1 detects a time trajectory of xy coordinates on a tablet plane of a pen tip corresponding to each stroke of a character using a tablet-integrated liquid crystal display and a stylus pen. On the tablet-integrated liquid crystal display, a frame for character writing is provided, and only when one character is written in each frame, as a recognition device,
Do the right thing. After performing a smoothing process on the time trajectory, the pattern feature extraction unit 2 extracts a time series of a quantization direction in which a direction between adjacent points is quantized and a time series of a distance between adjacent points. With reference to these two time series, the quantization directions are integrated so that the sum of the corresponding distances is made as equal as possible, and a quantization direction vector having a predetermined dimension is created as a feature vector.

The pattern-to-pattern distance calculation means 3 uses a window-based DP matching (Dynamic Programming Matc) based on a direction difference.
hing) to calculate the distance between the feature vectors. The recognition result creating means 7 includes a character dictionary describing a character category corresponding to the stroke time series. From the stroke candidates for each of the strokes, of the direct product thereof, for each stroke time series candidate in which the corresponding character exists, The distance as a character determined as the sum of the distances for each stroke is calculated, and character candidates are obtained based on the calculated distance.

The recognition result presenting means 8 displays the character candidates on the tablet-integrated liquid crystal display. The correct answer category selecting means 9 receives the user's selection of the presented character candidate by the pen, detects which character is correct, and reads out a stroke time series corresponding to this character from the character dictionary. Then, the correct stroke category corresponding to each stroke is obtained. Here, if there is no correct character in the character candidates, this is detected by not selecting.

FIG. 2 is a flowchart for explaining the operation of the pattern recognition apparatus according to the present invention, in which a valid flag is used. Hereinafter, the steps will be sequentially described. Initial patterns are set in advance in the dictionary pattern storage means, and the validity is a positive value V
The init and valid flags are set to, for example, ON. These initial values depend on the nature of learning, for example, whether to create a dictionary from a so-called blank state, or whether to fine-tune a dictionary designed to some extent generally to suit the user, etc. Select and set an appropriate value (S
1).

When character / voice data or the like is input by the pattern input means (S2), the characteristic extracting means creates a characteristic pattern from these data (S3). Proximity dictionary pattern selecting means, the category C with respect to the input pattern x belonging to _x, every dictionary pattern in the nearest dictionary pattern m ₁ and the category C _1, m ₁ followed close dictionary pattern in m ₂ and its Kadekori C ₂ , C _x , the closest pattern m _x is selected. In the case where the valid flag is used, the dictionary pattern m ′ ₁ closest to x among the dictionary patterns for which the valid flag is on.
And its category C ′ ₁ are selected and set as the first recognition result (S5).

At this time, the validity adjusting means sets the validity of m ₁ as V _{1. When} C _x ≠ C ₁ (S 6) If V ₁ > Vmin holds (S 7) V ₁ ← V ₁ -1 (S8) ... (12) if V _x <Vmax is Naritate _{(S11) V x ← V x} +1 (S12) ... (13) · when C _x = C ₁ and _{V 1 <Vmax (S6, S16} ) V 1 ← V ₁ +1 (S17) (14)

Here, Vmax and Vmin are the upper and lower limits of the effectiveness, which are appropriately determined. In this case, if V ₁ is becomes a negative value (S9), the dictionary pattern deleting unit, remove the dictionary pattern, or the case of using a valid flag, the dictionary pattern invalidating means (S1
0), the corresponding valid flag is turned off. Also, if V ₁ or V _x becomes positive value (S13, S18), if is used valid flag, the dictionary pattern enabling means (S14, S19), and turns on the corresponding valid flag . Here, it is assumed that the dictionary pattern is invalidated / validated using V ₁ = 0 as a reference value. However, the value of V ₁ is arbitrary, and other values may be used as reference values. The same effect may be obtained even if the width of adjustment is different. The dictionary pattern adjusting means may include the LV
According to Q, the dictionary pattern m ₁ is adjusted in accordance with the above equations (1) and (2). Here, α in the equation is determined, for example, as α = 1 / (max (V ₁ +1, ₁ )) (15).

Further, when C _x ≠ C ₁ and the adjusted V ₁ is positive or 0, or when C _x ≠ C ₁ and C _x ≠
In the case of C ₂ is the dictionary pattern adding means adds the x to the dictionary (S21). However, because of the limited capacity of the dictionary can not have this additional, for example, if it exceeds the maximum value Ndicmax number dictionary pattern for each previously provided category (S22), among the patterns of the dictionary that belong to the category C _x Valid The dictionary pattern with the smallest degree is deleted (S23), and the above-described addition is performed. The validity of the new dictionary pattern is a predetermined value Vadd, for example, 0, and the validity flag is turned on when it is set (S
24).

Here, by turning on the valid flag, if the user again inputs a pattern similar to this,
Normally, this can be correctly recognized immediately, and by setting the validity to 0, if the pattern has a greater adverse effect on recognition, the dictionary pattern is changed when it is found. Can be disabled.
Alternatively, for example, when it is desired to validate the dictionary pattern after an input pattern similar to the dictionary pattern is frequently found, Vadd may be set to a negative value and the initial setting of the valid flag may be turned off.

The feature extracting means is not always necessary depending on the type of pattern to be recognized. Dictionary pattern adjustment means is not always necessary,
Alternatively, it is conceivable to execute only one of the expressions (1) and (2) and omit the other. When the number of dictionary patterns per category is sufficiently large, it is considered that an effect equivalent to dictionary pattern adjustment can be obtained by enabling / disabling the dictionary according to the increase / decrease of the effectiveness. Also, instead of LVQ, LVQ2, LVQ3, or another pattern adjustment method may be used.

However, the learning method of the present invention is characterized in that the dictionary pattern is adjusted or increased or decreased based on the influence on the recognition result. Therefore, it is desirable to perform the pattern adjustment based on the recognition result. Further, the effectiveness reduction when the recognition result is not correct and the invalidation of the dictionary pattern based on the result are performed only for the first candidate. For example, when several candidates are obtained in the order of the closest distance. It is conceivable that the same process is performed for all the candidates higher than the top candidate among the candidates belonging to the correct answer category.

The condition for adding a dictionary pattern when the recognition result is not correct is not necessarily the same as described above. The dictionary pattern adding unit or the dictionary pattern validating unit is not always necessary, and can be generally input only by deleting the dictionary pattern or disabling the dictionary pattern, but invalidating a pattern that is not input by an individual user to be adapted. By doing so, the recognition rate may be improved. Further, instead of adding the input pattern x as it is, the dictionary pattern adding means may add, for example, x '= x + β (mx- _x ) (16). Here, β is an appropriately determined coefficient.

[0028] In addition, depending on the operation method, sometimes C _x to the correct category is not necessarily found, in this case,
It is conceivable that no learning is performed on this input pattern, or only the effectiveness of the first candidate is reduced and the dictionary pattern is invalidated as a result.
Further, the pattern recognition apparatus of the present invention is characterized in that it can perform writer adaptation, speaker adaptation, and the like by learning the tendency of a pattern that is simultaneously input while operating as a recognition apparatus. May be created and adjusted in advance, and the dictionary contents may not be changed during operation as a recognition device.

Also, the dictionary pattern storage means is divided into two types, the contents of the first dictionary pattern storage means are not changed during operation as a recognition device, and the second dictionary pattern storage means is used for learning during operation. May be used. In this case, the contents of the first dictionary pattern storage means are also adjusted by adjusting the corresponding validity or validity flag.
A learning effect can be obtained. Alternatively, the pattern recognition device of the present invention may be used only for creating and adjusting dictionaries, and when operating as a recognition device, only the stored dictionary patterns may be stored and used in another device.

FIG. 3 is a flow chart for explaining another operation of the pattern recognition apparatus according to the present invention, and is a flow chart in the case of performing online learning of a stroke pattern in online character recognition. Hereinafter, the steps will be sequentially described. First, perform initial settings (S
1) Stroke recognition is performed for the first stroke (S2, S3). Next, it is determined whether or not there is any remaining stroke (S4). If there is any remaining stroke, the process proceeds to the next stroke (S5). If no stroke remains, character candidates are created (S6). Next, character candidates are presented (S7), the correct character is selected by the user (S8), and stroke learning is performed for the first stroke (S9, S10).
Next, it is determined whether or not any strokes remain (S1).
1) If remaining, the process moves to the next stroke (S1).
2) If not remaining, the process is terminated.

For the dictionary pattern, an initial pattern is created in advance and stored in the dictionary pattern storage means 4. this is,
For example, the sample data may be created by clustering using a technique such as the k-means method. Further, the validity corresponding to the dictionary pattern existing at this time is set to Vinit. Vinit is a positive value, for example, about 5. Learning of a dictionary pattern, that is, adjustment, addition, and deletion is performed in the following procedure for each character input. When a pattern is input by the pattern input unit 1, the feature extracting unit 2 creates a feature pattern x from this data. Proximity dictionary pattern selection means 6 using the inter-pattern distance calculating unit 3 and the dictionary pattern storage unit 4, all near the dictionary pattern in the 1 ... Ns-th characteristic pattern x dictionary pattern m _i
And each category C _i (i = 1... Ns) is selected. Here, Ns is an appropriately determined number of stroke candidates for each stroke.

The above procedure is performed for each stroke in a character, and stroke candidates for each stroke are obtained. The recognition result creating means 7 creates a character candidate by the above method, and the recognition result presenting means 8 presents this to the user. The correct answer category selecting means 9 obtains a correct answer stroke category corresponding to each stroke from the correct answer character selected by the user by the above-described method. However, if there is no correct answer in the candidates, the correct stroke category is not known, and therefore, learning regarding this input pattern is not performed. Thereafter, the following procedure is executed for each stroke in the character. Here, the correct category referred to as C _x on the stroke of being processed.

First, the minimum i that satisfies C _i = C _x (i = 1... Ns) is selected. The m _i corresponding to the i or less m _x. The effectiveness adjusting means 11 calculates the values of the expressions (12) and (1).
3) Based on Equation (14), adjust the validity V ₁ and V _x corresponding to m ₁ and _mx , respectively. Here, when the validity of V ₁ becomes a negative value, the dictionary pattern deletion unit 12 deletes the dictionary pattern m ₁ from the dictionary pattern storage unit 4. The dictionary pattern adjusting means may include the LV
According to Q, the dictionary pattern m ₁ is adjusted in accordance with the above equations (1), (2) and (15). Further, C _x ≠
If C ₁ and the adjusted V ₁ are positive or 0, or if C _x ≠ C ₁ and C _x ≠ C ₂ , x is added to the dictionary of category C _x by the dictionary pattern adding means. Add to However, if this additional to the limitation of the capacity of the dictionary is not possible, remove the smallest dictionary pattern of effectiveness of the dictionary patterns belonging to the category C _x, then perform additional said. The validity of the new dictionary pattern is a predetermined value Vadd. Vadd may be set to 0, for example.

One of the pattern adding means and the pattern deleting means may be omitted. Further, when the number of dictionary patterns per category is sufficiently large, it is considered that an effect equivalent to dictionary pattern adjustment can be obtained by enabling / disabling the dictionary according to the increase / decrease of the degree of effectiveness. Is not always necessary. If there is no correct answer in the character candidates, the learning is not performed. However, it is conceivable that only the effectiveness reduction for the first candidate and the dictionary pattern deletion based on the result are performed. Alternatively, monitor user retries or alternate keyboard input, or
By outputting a message and asking for the correct answer,
It is also conceivable to know the correct answer category.

Further, in the above description, online learning of a stroke pattern in online character recognition has been described as an example, but audio data, OCR (Optical Character
Readers may learn other types of patterns, such as character data in an optical character reading device, or learn offline. In these cases, the recognition result creation means,
The recognition result presenting means and the correct answer category selecting means are not always necessary.

FIG. 4 is a diagram showing another embodiment of the pattern recognition apparatus according to the present invention. In the drawing, reference numeral 21 denotes a pattern input unit, 22 denotes a feature extraction unit, 23 denotes a pattern distance calculation unit, and 24 denotes a dictionary pattern. Storage means, 25 is a validity storing means, 26 is a proximity dictionary pattern selecting means, 27 is a recognition result creating means, 28 is a recognition result presenting means, 29 is a correct answer category selecting means, 30 is a dictionary pattern adjusting means, and 31 is a validity degree. Adjusting means, 32 is a dictionary pattern invalidating means, 33 is a dictionary pattern adding means, 34 is a dictionary pattern validating means,
Reference numeral 35 denotes a validity flag storage unit.

The features required for recognition are extracted by the feature extracting means 22 from the pattern input by the pattern input means 21. The inter-pattern distance calculation means 23 calculates the distance between the patterns, and the dictionary pattern storage means 24
A variable number of dictionary patterns are stored for each category to be recognized. The validity flag storage means 35 stores the validity degree of each dictionary pattern,
An on / off state can be set for each dictionary pattern. The proximity dictionary pattern selecting means 26 selects a dictionary pattern and a category in a set of dictionary patterns, based on the distance between the patterns, such that the distance to a certain pattern is ordered in a certain order.

The recognition result creating means 27 creates a recognition result to be finally presented to the user from the recognition result obtained by the proximity dictionary pattern selecting means 26. The recognition result presenting means 28 presents the recognition result to the user by the recognition result creating means 27. Correct answer category selection means 29
Determines the category to which the input pattern should belong from the user's operation, based on the user's operation based on the recognition result presented by the recognition result creating means 27. The dictionary pattern adjusting means 30 selects a dictionary pattern for the input pattern of a known category based on the proximity dictionary pattern selecting means, and changes based on the dictionary pattern.

The validity adjusting means 31 selects a dictionary pattern for the input pattern of a known category based on the close dictionary pattern selecting means, and the category to which these dictionary patterns belong is the same as the category to which the input pattern belongs. The validity of the dictionary pattern is increased or decreased based on whether the dictionary patterns are the same. The dictionary pattern validating means 34
Dictionary pattern invalidating means for turning off the valid flag of the dictionary pattern based on the validity of the dictionary pattern;
The validity flag of a dictionary pattern is turned on based on the validity of the dictionary pattern. Dictionary pattern adding means 33
Selects a dictionary pattern for an input pattern of a known category based on the proximity dictionary pattern selecting means, and determines whether the category to which these dictionary patterns belong is the same as the category to which the input pattern belongs. Then, a pattern created based on the input pattern is added as a dictionary pattern. An example of a method of realizing each part is the same as in the first embodiment of the present invention.

Next, the operation of the device will be described.
The initial settings of the dictionary pattern and the validity are the same as in the first embodiment of the present invention. Further, the valid flag corresponding to the dictionary pattern existing at this time is set to ON. Learning of a dictionary pattern, that is, adjustment, addition, and deletion is performed in the following procedure for each character input.

From the stroke recognition for each stroke to the creation of character candidates and obtaining the correct answer category,
This is the same as described above. However, the procedure of creating a stroke candidate for learning a stroke pattern is the same as that of the first embodiment of the present invention. However, when obtaining a recognition result to be presented to the user, a dictionary pattern whose valid flag is ON is obtained. Dictionary pattern m close to 1...
_i ′ and respective categories C _i ′ (i = 1... Ns ′) are selected, and character candidates are created using these. Also,
If there is no correct answer in the candidate, the correct stroke category is not determined, and therefore, learning on this input pattern is not performed. Thereafter, the following procedure is executed for each stroke in the character. Here, the correct category referred to as C _x on the stroke of being processed.

First, the smallest i that satisfies C _i = C _x (i = 1... Ns) is selected. The m _i corresponding to the i or less m _x. The effectiveness adjusting means 31 calculates the values of the equations (12) and (1).
3) Based on Equation (14), the validity corresponding to m ₁ and m _x adjusts V ₁ and V _x , respectively. Here, when the validity of V ₁ becomes a negative value, the valid flag of the dictionary pattern m ₁ is turned off. When the validity of V ₁ or V _x becomes a positive value, the dictionary patterns m ₁ ,
_{Turn on} the valid flag for mx. Dictionary pattern adjustment means, by the of LVQ, the formula (1), equation (2), according to equation (15), the adjustment of the dictionary pattern m _1. Further, when C _x ≠ C ₁ and the adjusted V ₁ is positive or zero, or when C _x ≠ C ₁ and C _x ≠ C ₂ ,
The dictionary pattern adding means adds _{x to} the dictionary of category Cx. However, if this additional from the limitation of the capacity of the dictionary is not possible, to remove the smallest dictionary pattern of the effectiveness of the dictionary patterns belonging to the category C _x, then,
The above addition is performed. The validity of the new dictionary pattern is
A predetermined value Vadd is set, and the valid flag is turned on. Vadd may be set to 0, for example.

As in the case of the first embodiment, one of the pattern adding means and the pattern deleting means may be omitted, and the dictionary pattern adjusting means is not necessarily required. Further, even when there is no correct answer in the character candidates, learning may be performed by the above-described means. Further, in the above description, online learning of a stroke pattern in online character recognition has been described as an example, but learning of other types of patterns or offline learning may be performed. In these cases, the recognition result creating means, the recognition result presenting means, and the correct answer category selecting means are not always necessary.

[0044]

As is apparent from the above description, according to the present invention, since the operation of increasing or decreasing the number of dictionary patterns in each category is performed, a nearly optimal number of dictionary patterns can be stored for each category. The new pattern is handled differently from the already registered pattern depending on the validity or validity flag, so it can be imported into the dictionary without repeating the dictionary pattern adjustment, and the user can add the dictionary It can be applied to registration or automatic writer and speaker adaptation.

[Brief description of the drawings]

FIG. 1 is a configuration diagram for explaining an embodiment of a pattern recognition device according to the present invention.

FIG. 2 is a flowchart for explaining the operation of the pattern recognition device according to the present invention.

FIG. 3 is a flowchart for explaining another operation of the pattern recognition device according to the present invention.

FIG. 4 is a diagram showing another embodiment of the pattern recognition device according to the present invention.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Pattern input means, 2 ... Feature extraction means, 3 ... Pattern distance calculation means, 4 ... Dictionary pattern storage means, 5 ... Effectiveness storage means, 6 ... Proximity dictionary pattern selection means, 7 ... Recognition result creation means, 8 ... Recognition result presenting means, 9 ... Correct answer category selecting means, 10 ... Dictionary pattern adjusting means, 11 ... Effectiveness adjusting means, 12 ... Dictionary pattern deleting means, 13 ... Dictionary pattern adding means.

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-57-66498 (JP, A) JP-A-7-28955 (JP, A) JP-A-4-326471 (JP, A) (58) Field (Int.Cl. ⁶ , DB name) G06T 7/00 G06K 9/68 G10L 3/00 533

Claims (6)

(57) [Claims] 1. A pattern input means for inputting a pattern such as character / voice data, an inter-pattern distance calculating means for calculating a distance between the patterns, and a variable number of dictionary patterns for each category to be recognized. Dictionary pattern storage means, and effectiveness storage means for storing the validity of each dictionary pattern, and within a set of dictionary patterns, a distance between a certain pattern based on the distance between the patterns is ordered in a certain order. Proximity dictionary pattern selecting means for selecting a suitable dictionary pattern and category, and for an input pattern with a known category, selecting a dictionary pattern based on the close dictionary pattern selecting means, and the category to which the selected dictionary pattern belongs, Based on whether the category is the same as the category to which the input pattern belongs, A validity adjusting means for increasing / decreasing validity, a dictionary pattern is selected based on the proximity dictionary pattern selecting means for an input pattern of a known category, and the dictionary pattern is selected based on the validity of the selected dictionary pattern. A pattern recognition device comprising: a dictionary pattern deletion unit for deleting a pattern. 2. A pattern input means for inputting a pattern such as character / voice data, an inter-pattern distance calculating means for calculating a distance between the patterns, and a variable number of dictionary patterns for each category to be recognized are stored. Dictionary pattern storage means, and effectiveness storage means for storing the validity of each dictionary pattern, and within a set of dictionary patterns, a distance between a certain pattern based on the distance between the patterns is ordered in a certain order. Proximity dictionary pattern selecting means for selecting a suitable dictionary pattern and category, and for an input pattern with a known category, selecting a dictionary pattern based on the close dictionary pattern selecting means, and the category to which the selected dictionary pattern belongs, Based on whether the category is the same as the category to which the input pattern belongs, A validity adjusting means for increasing / decreasing validity, and a dictionary pattern is selected based on the proximity dictionary pattern selecting means for an input pattern with a known category, and the category to which the selected dictionary pattern belongs is the category of the input pattern. A pattern recognition device comprising: a dictionary pattern adding unit that adds, as a dictionary pattern, the input pattern or a pattern created based on the input pattern based on whether the category belongs to the category. 3. A pattern input means for inputting a pattern such as character / voice data, an inter-pattern distance calculating means for calculating a distance between the patterns, and a variable number of dictionary patterns for each category to be recognized are stored. Dictionary pattern storing means, validity storing means for storing the validity of each dictionary pattern, valid flag storing means for storing a valid flag that can take an on / off state for each dictionary pattern, and dictionary pattern Within the set, the elements of the set can be selected based on the valid flag, and based on the distance between the patterns,
Proximity dictionary pattern selection means for selecting a dictionary pattern and a category such that the distance to a certain pattern is ordered in a certain order, and for a known input pattern of a category, select a dictionary pattern based on the proximity dictionary pattern selection means, Based on whether the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs, based on whether or not the validity of the dictionary pattern is increased or decreased; And a dictionary pattern invalidating unit for turning off a valid flag of the dictionary pattern based on the dictionary pattern. 4. The pattern recognition apparatus according to claim 3, further comprising dictionary pattern validating means for turning on a valid flag of the dictionary pattern based on the validity of the certain dictionary pattern. 5. A pattern input means for inputting a pattern such as character / voice data, an inter-pattern distance calculating means for calculating a distance between the patterns, and a variable number of dictionary patterns for each category to be recognized are stored. Dictionary pattern storing means, validity storing means for storing the validity of each dictionary pattern, valid flag storing means for storing a valid flag that can take an on / off state for each dictionary pattern, and dictionary pattern Within the set, the elements of the set can be selected based on the valid flag, and based on the distance between the patterns,
Proximity dictionary pattern selection means for selecting a dictionary pattern and a category such that the distance to a certain pattern is ordered in a certain order, and for a known input pattern of a category, select a dictionary pattern based on the proximity dictionary pattern selection means, Based on whether the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs, based on whether or not the validity of the dictionary pattern is increased or decreased; A pattern recognition device for turning on a valid flag of the dictionary pattern based on the dictionary pattern. 6. A dictionary pattern is selected based on the proximity dictionary pattern selecting means for an input pattern whose category is known, and the category to which the selected dictionary pattern belongs is the same as the category to which the input pattern belongs. 6. The dictionary pattern adding device according to claim 1, further comprising: a dictionary pattern adding unit that adds the input pattern or a pattern created based on the input pattern as a dictionary pattern based on whether or not the input pattern exists. Pattern recognition device.