JPH0632087B2 - Pattern recognition device - Google Patents

Description

The present invention relates to a pattern recognition device.

[Background of the Invention]

In character recognition, an input pattern is compared with a standard pattern registered in a memory in advance for recognition. This set of standard patterns is called a dictionary. As shown in FIG. 2, when recognizing handwritten Chinese characters in the recognition unit 20 using the common dictionary (a set of standard patterns for recognizing character patterns written by an unspecified writer) 21, the number of target character categories is set. There are many thousands and the glyph is complicated, so
It is possible to read a polite typeface, but if you try to read a daily typeface, you will not get a sufficient recognition rate, or the dictionary capacity (that is, cost) and the recognition time that will try to obtain a sufficient recognition rate are enormous and impractical. There was a problem such as a thing.
As a method for solving these problems, the third conventional method is used.
The scheme shown in the figure has been proposed. Fig. 4 shows an example of a well-designed regular typeface and everyday typeface.

The system shown in FIG. 3 uses a personal dictionary (set of standard patterns dedicated to an individual) 31 as a dictionary. The personal dictionary 31 is made up only of characters written by the individual. With this method, the recognition rate and recognition speed can be increased and the dictionary capacity can be reduced as compared with the case of using a shared dictionary, but there are the following drawbacks.

At the time when the number of learned characters was small, the personal dictionary still had few categories used by individuals. Therefore, personal dictionaries are useless.

Even if the usage category is in the personal dictionary, the recognition rate is low if the number of learning samples for that category is small. See, for example, the learning curve for the recognition rate when using the personal dictionary in FIG. This figure shows that the recognition rate in the vicinity of the learning sample numbers 1 and 2 is lower than that in the shared dictionary, and the personal dictionary is not useful.

The conventional method is discussed in the following document.

IEICE Technical Report, PRL 81-94, Seiichiro Naito, Basic Study on Personal Handwritten Kanji Recognition, IEICE Transactions, Vol.J66-D, No.4, P.454
~ 455, Mitsu Yoshimura, Regarding Effectiveness of Personal Template in Handwritten Character Recognition [Object of the Invention] An object of the present invention is to provide a pattern recognition apparatus capable of solving the above-mentioned drawbacks of the prior art.

[Outline of Invention]

In order to achieve the above object, the present invention configures a pattern recognition device having the configuration shown in FIG.

In FIG. 1, the dictionary 1 is a dictionary for recognition. Dictionary 2
Is a part for storing the statistical value for each category for the characteristic pattern entered by the specific writer. When the number of characters entered by the specific writer is small or there is no character, the common dictionary is set in the dictionary 1.

In addition, the dictionary replacement unit 8 increases the number of characters written by the specific writer, and as the learning progresses, the dictionary 2 (created only from the character samples written by the specific writer) is expanded, and the entire dictionary 1 is expanded. Alternatively, the individual standard patterns belonging to the dictionary 1 are replaced with the entire dictionary 2 or the standard patterns belonging to the dictionary 2.

In this device, the drawbacks of the conventional invention were solved and the following effects could be obtained.

As the number of characters entered by the specific writer increases, the dictionary 1 can be gradually replaced from the common dictionary to the personal dictionary. Moreover, it can be done automatically without human intervention. The recognition performance improves as the number of shared standard patterns in the dictionary 1 decreases. Also, the recognition speed is improved.

For the categories that individuals have not filled in even once, or the number of filled samples is small even if they are filled in, the conventional standard OCR for unspecified writers (Fig. 2 The same level of recognition performance can be obtained. In other words, it is possible to read a neat typeface.

As for the categories written by a large number of individuals, everyday typefaces can be read with high recognition performance, similar to the conventional OCR for specified writers (method of FIG. 3).

The personal standard patterns in the dictionary 1 and the dictionary 2 are always pure ones made only from the samples written by the individual. In other words, the sample filled in by others is not used for making.

Example of Invention

Next, the first embodiment of the present invention will be described using the configuration example of FIG. The configuration example includes a recognition unit 3, an answer category memory unit 4, a feature memory unit 5, a correction unit 6, a learning unit 7, a dictionary 1, a dictionary 2, a dictionary replacement unit 8, and a control unit 9.

First, a character pattern string is input to the recognition unit 3 as an input pattern string. The recognition unit 3 converts the input pattern sequence into an electric signal by photoelectric conversion, and then performs preprocessing, cutout, feature extraction, and the like, and outputs it as a sequence of feature patterns cut out for each character.

Further, the recognition unit 3 evaluates the similarity between each of these characteristic patterns and all the standard patterns belonging to the dictionary 1, and then outputs the answer category based on the evaluation value. The sequence of feature patterns and the sequence of answer categories output by the recognition unit 3 are stored in the feature memory unit 5 and the answer category memory unit 4, respectively. In the correction unit 6, the user of this apparatus confirms and corrects the answer category in the answer category memory unit 4. The learning unit 7 inputs the characteristic pattern belonging to the characteristic memory unit 5 and the corrected answer category belonging to the answer category memory unit 4, calculates the average value of the inputted characteristic pattern for each category, and outputs it to the dictionary 2. The dictionary replacement unit 8 replaces the standard pattern belonging to the dictionary 1 with each standard pattern belonging to the dictionary 2 based on the evaluation value of the standard pattern in the dictionary 2.

A feature of the present invention is that a dictionary 2 is provided in addition to the dictionary 1, and
The dictionary replacement unit 8 is provided. The dictionary 2 is a part for storing the statistical value for each category of the characteristic patterns extracted from the character patterns written by the specific user of the apparatus. For example, in the learning unit 7, the characteristic patterns f ₁ , f ₂ , ..., Fn ₁ whose character category is “sea”, the characteristic patterns g ₁ , ..., Gn ₂ whose character category is “mountain”, and the character category “ When the characteristic pattern h ₁ , h ₂ , ..., H _n3 that is “river” is input, the learning unit 7 sets the average values f, g, and h for each of the character categories “sea”, “mountain”, and “river”. Output as a standard pattern of. Here, the characters inside the brackets (“,”) represent character categories, and f, g, and h are respectively Shall be represented. The dictionary replacement unit 8 is a unit that determines whether or not each standard pattern belonging to the dictionary 2 can be sufficiently used in the recognition dictionary 1, and if it is determined to be sufficient, replaces it in standard pattern units. Specifically, the dictionary replacement unit 8 executes the following procedure (referred to as dictionary replacement method 1).

(Dictionary Replacement Method 1) (1) The following evaluation values for each standard pattern belonging to the dictionary 2, (a) the category of the standard pattern is in the dictionary 1, or (b) if there is, the standard pattern in the dictionary 1 Calculate common standard pattern or personal standard pattern, (c) number of learning samples, (d) other.

(2) The standard pattern belonging to the dictionary 2 is added to or updated in the dictionary 1 based on all or any combination of the evaluation values (a) to (d) obtained in the above (1).

In the drawing, the description table 2B of the dictionary 2 lists the evaluation values (a) to (d) calculated in the procedure (1) of the dictionary replacement method 1 for each standard pattern.

7 and 8 show a specific example of the dictionary replacement method 1. Both were calculated and evaluated based on the dictionary 2 main body and the dictionary 2 of FIG. 7 (b) and FIG. 8 (b) based on the dictionary 1 of FIG. 7 (a) and FIG. 8 (a). After that, FIG. 7 (c) and FIG. 8 (c)
The dictionary 1 of is output. 7 (c) and 8 (c)
As shown in the note 1 below, Fig. 7 (a)
(B) (c) and FIGS. 8 (a) (b) (c),
Both represent common standard patterns, and individual represent individual standard patterns. Further, as shown in Note 2, θ ₁ represents the learning sample number evaluation threshold for the common standard pattern, and θ ₂ represents the learning sample number evaluation threshold for the individual standard pattern. Specifically, if θ ₁ is a shared standard pattern (A ₁ ) for a category in dictionary 1 and it is replaced with a personal standard pattern (A ₂ ) for the same category in dictionary 2, (1 It is used to request that the number of learning samples in ( ₁ ) is greater than θ ₁ . θ ₂ is the personal standard pattern (A
_{If 1} ) already exists, it is used to request that the number of learning samples of (A ₂ ) is θ ₂ or more in order to replace it with the standard pattern (A ₂ ) for the same category of dictionary 2.

Next, the second specific example will be described. In this example, the dictionary replacement unit executes the following procedure (referred to as dictionary replacement method 2).

(Dictionary Replacement Method 2) (1) It is determined whether the entire dictionary 2 can be sufficiently used as a personal dictionary for a specific user of this device. Specifically, whether (a) the number of categories n _c is sufficiently large, (b) the average number of learning samples per standard pattern n _s is sufficiently large, or (c) the total value n _t of all standard patterns If there are enough, evaluate (d) other. Based on all of (a) to (d) or an appropriate arbitrary combination, it is determined whether the dictionary 2 can be sufficiently used. For example, if n _c θ _c and n _s θ _s , it is determined that the dictionary 2 can be used sufficiently. Here, θ _c and θ _s are positive constants.

(2) If it is determined in the above (1) that the dictionary 2 can be used sufficiently,
The entire dictionary 1 is replaced with the dictionary 2 (special user only).

The dictionary replacement method 2 could also be realized by the configuration example of FIG. However, in this case, the following 2 is different from the case of the dictionary replacement method 1.
The points are different.

(i) Calculate the above evaluation values (a) to (d) of the dictionary 2 as a whole,
Record them in the description table of the dictionary 2.

(ii) It is not necessary to input from the dictionary 1 to the dictionary replacement unit.

Next, a third specific example will be described. In this example, the dictionary replacement unit performs a process that combines dictionary replacement methods 1 and 2. Specifically, as shown in FIG. 9, the dictionary replacement method 2 is first applied. If it is determined that the dictionary 2 can be used enough, the dictionary 2
The entire dictionary 1 is replaced as a whole, and the dictionary replacement method 2 is completed. If it is determined that the dictionary 2 is insufficient, the dictionary replacement method 1 is applied.

Next, a fourth specific example will be described. As a result of the dictionary replacement unit 2 executing the dictionary replacement method 2 in the specific example 2 or 3, the dictionary 1
It is assumed that the entire dictionary is replaced with the entire dictionary 2 and the dictionary 1 is a dictionary dedicated to a specific user (not including a shared standard pattern at all). In this example, after this state, the dictionary replacement unit executes the following dictionary replacement method 3.

(Dictionary Replacement Method 3) (1) The amount of the portion of the dictionary 2 that has changed due to learning (that is, the added or updated standard pattern) is evaluated as the amount of change. The evaluation of changes is as follows: (a) increase in the number of categories, (b) increase in the average number of learning samples per standard pattern, (c) increase in the total number of learning samples for all standard patterns Minutes, (d) etc., and based on all of them or any suitable combination.

(2) If the change evaluated in (1) above is large enough, dictionary 2
The above-mentioned changed part of is substituted into the dictionary 1. Here, substitution means
This is to add or update the standard pattern added or updated in the dictionary 2 to the dictionary 1.

The dictionary replacement method 3 could also be realized by the configuration example of FIG. However, in this case, the following two points are different from the case of the dictionary replacement method 1.

(i) Calculate the above (a) to (d), which is the change of the entire dictionary 2,
Record them in the description table of the dictionary 2.

(ii) It is not necessary to input from the dictionary 1 to the dictionary replacement unit.

The above two embodiments can be realized by using a microprocessor, a memory, a scanner, a microphone, a display and the like.

The present invention can also be applied to the following cases.

(1) The above embodiment is for the case where the dictionary consists of a set of standard patterns and thresholds per category. The recognition performance can be further improved by using a plurality of sets of standard patterns per category. The present invention can be applied to this case as well.

(2) In the above embodiment, all feature patterns obtained from the input patterns are used as learning samples. A dictionary can be created in the same manner by using the following as a learning sample.

A learning sample is obtained by excluding the characteristic patterns determined by the user not to be learned from all the characteristic patterns.

A feature pattern that is an error or rejected by recognizing the input pattern is used as a learning sample.

A learning sample is obtained by excluding the characteristic pattern determined by the user not to be learned from the characteristic pattern that has become an error or rejected.

The learning sample selection and determination processing described above can be performed by the correction unit in FIG.

(3) In the above embodiment, the character pattern is the recognition target. The present invention can be applied to recognition of patterns other than characters such as voice patterns and graphic patterns.

The present invention can also be applied to each logically appropriate case that is a combination of (1) to (3).

〔The invention's effect〕

According to the present invention, the following effects (1) to (4) can be achieved.

(1) As the number of characters written by the specified scribe increased, it was possible to gradually replace the dictionary 1 from the common dictionary to the personal dictionary. Moreover, it could be done automatically without human intervention. The recognition performance improved as the number of shared standard patterns decreased from within the dictionary 1. The recognition speed has also improved.

(2) For categories that individuals have not filled in even once, or who have filled in a small number of samples even though they have filled in, by using the common standard pattern, the conventional OCR for unspecified writers ( The same recognition performance as that of the method shown in FIG. 2) was obtained. In other words, I was able to read the polite typeface.

(3) As for the category in which a large number of individuals were filled in, everyday typefaces could be read with high recognition performance, similar to the conventional OCR for specified writers (method of FIG. 3).

(4) The standard patterns for personal use in the dictionary 1 and the dictionary 2 were always pure ones made only from the samples entered by the individual. In other words, the sample written by another person is not used for making.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention, FIGS. 2 and 3 are block diagrams for explaining the prior art,
FIG. 4 is an explanatory diagram showing an example of a polite typeface and everyday typeface, FIG. 5 is a graph showing an example of a learning curve for the recognition rate when a personal dictionary is used, and FIG. 6 is a dictionary replacement method. FIG. 7 is a block diagram for explaining, FIG. 7 and FIG. 8 are diagrams for explaining a concrete example of the dictionary replacement method, and FIG. 9 is a flow chart for explaining another concrete example. 1 ... Dictionary 1, 2 ... Dictionary 2,3 ... Recognition unit, 4 ... Answer category memory unit, 5 ... Feature memory unit, 6 ... Correction unit, 7 ... Learning unit, 8 ... Dictionary replacement Department, 9 ... control department.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-59-106085 (JP, A)

Claims (5)

[Claims] 1. A first dictionary, which is a set of first standard patterns representing each category, and a similarity between a feature pattern cut out from an input pattern and each standard pattern belonging to the first dictionary are evaluated. , A recognition section that calculates and outputs an answer category corresponding to the input pattern based on the evaluation value of those similarities, and also outputs a feature pattern cut out from the input pattern, and an answer that is an output of the recognition section. A correction unit that corrects the category, the corrected answer category, and the feature pattern output from the recognition unit are input, and the average value of the feature patterns corresponding to the answer category is output as a second standard pattern for each category. The learning unit, the second dictionary to which the second standard pattern is input, added or updated, and the evaluation value of the contents of the second dictionary have a certain threshold value. Pattern recognition apparatus the contents When e is replaced with the corresponding portion of the first dictionary, or characterized by comprising a dictionary replacement unit to be added. 2. The pattern recognition apparatus according to claim 1, wherein the dictionary replacement unit replaces the first standard pattern with the second standard pattern in the category of the second standard pattern. For each of the above, whether the category of the second standard pattern is in the first dictionary, and if any, the standard pattern in the first dictionary is the first standard pattern or the second standard pattern. The pattern recognition apparatus is characterized in that it has been replaced, added, or updated by the standard pattern, or is based on at least one evaluation value among the evaluation values consisting of the number of learning samples. 3. The pattern recognition apparatus according to claim 1, wherein the dictionary replacement unit determines the sufficiency of the second dictionary as a personal dedicated dictionary for a specific user, per number of categories and standard patterns. Of the second dictionary as a whole after the judgment based on at least one evaluation value among the evaluation values consisting of the average learning sample number and the total value of all the standard patterns of the learning sample Is replaced with the entire first dictionary. 4. The pattern recognition apparatus according to claim 3, wherein when it is determined that the sufficiency of the second dictionary as a personal dictionary is unsatisfactory, the dictionary replacement unit causes the second dictionary. Replacement of each category constituting the standard pattern with the first standard pattern is performed for each category constituting the second standard pattern, in which the category of the second standard pattern to be the first standard pattern is the first category. 1 dictionary, if there is one,
Pattern recognition device based on at least one of the evaluation values consisting of the number of learning samples, or the second standard pattern. 5. The pattern recognition apparatus according to claim 3 or 4, wherein the second dictionary satisfies the sufficiency as a personalized dictionary, and the entire first dictionary is After replacing with the entire second dictionary, the amount of the second standard pattern changed by learning in the second dictionary is
It is evaluated as a change based on at least one of the evaluation values consisting of the increase in the number of categories, the increase in the average number of learning samples per standard pattern, and the increase in the total value of the learning samples for all standard patterns. A pattern recognition apparatus, wherein the changed part of the dictionary 2 is added to or updated in the first dictionary if the change is sufficiently large.