JP5319726B2 - PATTERN SELECTION DEVICE, PATTERN SELECTION METHOD, AND PROGRAM - Google Patents

Description

  The present embodiment relates to a pattern selection device, a pattern selection method, and a program.

  There is a pattern recognition technique for identifying an affiliation class of unknown patterns such as characters, images, and voices input to a recognition system, such as a character recognition process represented by an optical character recognition device (OCR). In this pattern recognition technique, an affiliation class of an unknown pattern is identified using a recognition dictionary. The class is a label attached when the recognition target is classified. For example, the character type is used for character recognition, and the name of the object is used for general object recognition.

  The recognition dictionary is prepared by preparing a pattern (hereinafter referred to as a dictionary pattern) to be used for the dictionary for each class in advance. When it is desired to increase the recognition accuracy in the pattern recognition technique, generally, the number of dictionary patterns prepared for each class is increased.

  However, this method has problems such as a large amount of dictionary patterns, which takes time to create a recognition dictionary, and the dictionary patterns may include patterns that are not so effective in improving recognition accuracy. . In addition, when a dictionary pattern such as the knn method or a method using a part of the dictionary pattern as a recognition dictionary is used, there is a problem that the capacity of the recognition dictionary tends to increase and the speed of the recognition processing decreases.

  As a method for selecting a dictionary pattern used for a recognition dictionary capable of realizing efficient pattern recognition, for example, there is a support vector machine (Support Vector Machine: SVM). In this method, only a dictionary pattern near the class boundary is selected to become a recognition dictionary, and other patterns are deleted. In other words, the number of necessary dictionary patterns is reduced by registering only the dictionary patterns necessary for expressing the class boundaries in the recognition dictionary.

  However, it is known that the above-described method does not take into account a portion away from the class boundary, and thus causes a problem that a high certainty factor is not necessarily obtained for a standard pattern. In addition, when applied to a problem with many variations such as handwritten character recognition, the reduction rate is often low because the ratio of patterns near the class boundary is high.

  Also known is a method of reducing the number of dictionary patterns by performing, for example, vector quantization and expressing each class with a small number of representative patterns.

  Although this method can greatly reduce the number of dictionary patterns, there is a problem that character recognition is vulnerable to such input because, for example, a small number of exceptional patterns such as characters and variants are easily reduced. .

  In addition, a pattern that seems to be ineffective from the dictionary pattern is also deleted manually, but especially when the number of dictionary patterns is enormous, the work time becomes long. The problem is that the dictionaries vary.

JP 2000-99723 A

  An object of the present invention is to provide a pattern selection device, a pattern selection method, and a pattern selection program that can efficiently select a pattern.

  The pattern selection apparatus according to the present embodiment refers to a storage unit that stores a first pattern group and a second pattern group used for evaluation of the first pattern, and refers to the second pattern. As the plurality of evaluation values, the first evaluation value, which is the degree of contribution when the first pattern recognizes the second pattern, and the first pattern are standard in the class to which the first pattern belongs. An evaluation value calculation unit for calculating a second evaluation value for evaluating whether the pattern is a simple pattern, and a third evaluation value for evaluating whether the first pattern is a pattern near the boundary of the class, and the calculated evaluation A score calculation unit that calculates a score for each first pattern using a value and weighting information for each evaluation value, and a pattern selection unit that selects a pattern from the first pattern based on the calculated score. Prepare.

1 is a block diagram showing the overall configuration of a pattern selection device according to a first embodiment. The figure which shows an example of the candidate pattern information which concerns on 1st Embodiment. 6 is a flowchart illustrating an example of an operation of the pattern selection device according to the first embodiment. The flowchart which shows an example of the calculation procedure of the evaluation value in the evaluation value calculation part which concerns on 1st Embodiment. The figure which shows an example of the pattern information for evaluation which concerns on 1st Embodiment. The figure which shows an example of the calculation formula of the evaluation value which concerns on 1st Embodiment. The flowchart which shows an example of the calculation procedure of the evaluation value in the evaluation value calculation part which concerns on 1st Embodiment. The flowchart which shows an example of the calculation procedure of the evaluation value in the evaluation value calculation part which concerns on 1st Embodiment. The figure which shows an example of the table of the evaluation value information storage part memorize | stored in the information storage part which concerns on 1st Embodiment. The flowchart which shows an example of the calculation procedure of the score in the score calculation part which concerns on 1st Embodiment. The figure which shows an example of the table of the weighting information storage part 45 for every evaluation value memorize | stored in the information storage part which concerns on 1st Embodiment. The figure which shows an example of the table of the score information storage part which concerns on 1st Embodiment. The figure which shows an example of the table of the pattern selection reference | standard information which concerns on 1st Embodiment. The block diagram which shows an example of the whole structure of the pattern selection apparatus which concerns on 2nd Embodiment. 9 is a flowchart showing an information update procedure in the pattern selection device according to the second embodiment. The figure which shows an example of the update conditions which concern on 2nd Embodiment. The figure which shows an example of the update information which concerns on 2nd Embodiment.

  Hereinafter, embodiments of a pattern selection device of the present invention will be described with reference to the drawings. In the present embodiment, as an example, a pattern selection device that selects a pattern used to create a recognition dictionary of a character recognition device will be described.

(First embodiment)
FIG. 1 is a block diagram showing the overall configuration of the pattern selection apparatus according to the first embodiment.

  As shown in FIG. 1, the pattern selection apparatus according to the present embodiment includes an evaluation value calculation unit 10, a score creation unit 20, a pattern selection unit 30, and an information storage unit 40.

  Note that the pattern selection apparatus according to the present embodiment includes a CPU, a main memory, an external storage device, and the like as hardware. For example, the above-described evaluation value calculation unit 10, score creation unit 20, and pattern selection unit 30 are realized by loading a pattern selection program stored in the external storage device onto the main memory by the CPU. In addition, these components may be realized by hardware configured by combining various electronic components instead of software.

The information storage unit 40 includes a candidate pattern storage unit 41, an evaluation pattern storage unit 42, an evaluation value information storage unit 43, a score information storage unit 44, a weighting information storage unit 45 for each evaluation value, and a pattern selection criterion. And an information storage unit 46.

The candidate pattern storage unit 41 stores the candidate pattern itself and information on the candidate pattern (hereinafter referred to as an information tag) as candidate pattern (first pattern) information that is a candidate for the dictionary pattern. In this embodiment, it is assumed that various patterns are prepared in advance as candidate patterns and stored in the candidate pattern storage unit 41 in association with information tags. From these candidate patterns, a pattern to be used for the recognition dictionary (hereinafter referred to as a dictionary pattern) is selected.

For example, in character recognition, a candidate pattern is character image data, and tag information of the candidate pattern includes a candidate pattern identifier, a class to which the candidate pattern belongs, a font, a tilt, an image size, and the like. The candidate pattern storage unit 41 may be input from the outside when performing pattern selection processing described later. Further, the identifier may be, for example, the registration order in the information storage unit 40 or the registration order for each class when the candidate patterns are classified into classes.

Here, with reference to FIG. 2, the candidate pattern information memorize | stored in the candidate pattern memory | storage part 41 of this embodiment is demonstrated.

  FIG. 2 is a diagram illustrating an example of candidate pattern information stored in the candidate pattern storage unit 41 according to the present embodiment. In character recognition, the candidate pattern has, for example, character image data. In addition, as a candidate pattern of the character recognition device, it is possible to target character patterns of all attributes such as kanji, numbers, hiragana, katakana, and alphabets, but in this embodiment, alphabets are applied as candidate patterns. .

  As shown in FIG. 2, the candidate pattern storage unit 41 of the present embodiment stores candidate patterns 1 to 3 belonging to the class “A” and candidate patterns 4 to 4 belonging to the class “B”. Candidate pattern 6 is stored. Each candidate pattern has image data of a candidate pattern (hereinafter referred to as a candidate pattern image) and an information tag corresponding to the candidate pattern image. The information tag includes a pattern ID 416 that is a pattern identifier, an affiliation class 417, a font 418, and a slope 419. The counterclockwise inclination is indicated by “−”, and the clockwise inclination is indicated by “+”. For example, if the character shape is inclined 10 degrees clockwise, the inclination information is +10, and if the character shape is inclined 20 degrees counterclockwise, the inclination information is -20.

  Candidate pattern 1 has candidate pattern image 410, pattern ID “A101”, belonging class “A”, font “font 1”, and inclination “± 0”. Candidate pattern 2 has candidate pattern image 411, pattern ID is “A102”, belonging class is “A”, font is “font 2”, and inclination is “+20”. Candidate pattern 3 has candidate pattern image 412, pattern ID “A103”, belonging class “A”, font “font 3”, and inclination “−30”.

  Candidate pattern 4 has candidate pattern image 413, pattern ID is “B101”, belonging class is “B”, font is “font 4”, and inclination is “± 0”. Candidate pattern 5 has candidate pattern image 414, pattern ID “B102”, belonging class “B”, font “font 5”, and inclination “+20”. Candidate pattern 6 has candidate pattern image 415, pattern ID is “B103”, belonging class is “B”, font is “font 2”, and inclination is “−30”.

  In this embodiment, in order to simplify the description, it is assumed that all of candidate patterns 1 to 6 shown in FIG. 2 are image data normalized in advance to a width of 100 pixels and a height of 100 pixels.

The class classifies patterns. When the target candidate pattern is alphabetic as described above, for example, classes “AZ” corresponding to the character types “AZ” included in the alphabet are prepared. Will be. That is, the candidate pattern “A” of the character type “A” belongs to the class “A”. In addition to the character type, the class may be classified for each font. Therefore, the candidate pattern may belong to a plurality of classes at the same time, such as belonging to the character type class and the font class at the same time.

  The evaluation pattern storage unit 42 stores information associating image data of an evaluation pattern (second pattern) used by the evaluation value calculation unit 10 described later for evaluating candidate patterns with tag information related to the evaluation pattern. Remember. In this embodiment, the evaluation pattern is selected in advance from the candidate patterns stored in the candidate pattern storage unit 41 and stored in the evaluation pattern storage unit 42.

  That is, in the present embodiment, a candidate pattern having the same affiliation class and the same font information as the target character to be subjected to character recognition is selected from the candidate patterns 1 to 6 stored in the candidate pattern storage unit 41 and evaluated. A pattern. In other words, the evaluation pattern is selected from the candidate pattern group based on the attribute of the pattern recognition target. The evaluation pattern storage unit 42 may be input from the outside when performing pattern selection processing to be described later, like the candidate pattern storage unit 41. The pattern selected as the evaluation pattern may be deleted from the candidate pattern group.

Here, with reference to FIG. 3, the pattern information for evaluation memorize | stored in the pattern storage part 42 for evaluation of this embodiment is demonstrated. Similar to the candidate pattern information, the evaluation pattern information includes the evaluation pattern itself and an information tag corresponding to the evaluation pattern.

  FIG. 3 is a diagram illustrating an example of the evaluation pattern storage unit 42 according to the present embodiment. As shown in FIG. 3, the evaluation pattern storage unit 42 of the present embodiment includes an evaluation pattern 1 belonging to the class “A” and an evaluation pattern 2 belonging to the class “B”. It is remembered. In this embodiment, for the sake of simplicity, all of the evaluation patterns 1 to 24 shown in FIG. 3 have a width of 100 pixels and a height of 100 pixels, and have the same size as the candidate pattern of FIG. Suppose there is. The class to which the pattern for evaluation of this embodiment belongs corresponds to the class to which the candidate pattern belongs. That is, the candidate pattern class stored in the candidate pattern storage unit 41 and the evaluation pattern class are in a correspondence relationship.

  The evaluation pattern 1 has a pattern ID “A201”, a belonging class “A”, a font “font 1”, and a slope “± 0”. In the evaluation pattern 2, the pattern ID is “B201”, the belonging class is “B”, the font is “font 5”, and the inclination is “+20”. That is, the evaluation pattern 1 is the candidate pattern 1 stored in the candidate pattern storage unit 41, and the evaluation pattern 2 is the candidate pattern 5 stored in the candidate pattern storage unit 41.

The evaluation value information storage unit 43 stores the evaluation value for each candidate pattern calculated by the evaluation value calculation unit 10.

The score information storage unit 44 stores a score for each candidate pattern calculated by the score creation unit 20.

The weighting information storage unit 45 for each evaluation value stores weighting information for performing weighting for each evaluation value used when the evaluation value calculation unit 10 to be described later calculates the evaluation value. The weighting information for each evaluation value is set in advance by a user or a system designer.

The pattern selection criterion information storage unit 46 is a selection criterion for selecting a dictionary pattern from the candidate pattern storage unit 41 during pattern selection processing by the pattern selection unit 30 described later. In this embodiment, the pattern selection reference information is set in advance by a system designer or the like.

Note that the information storage unit 40 may be configured by dividing an area into one storage unit by appropriately using one or a plurality of storage units, or may be stored in separate storage units. Or they may be combined. As the information storage unit 40, for example, a magnetic disk device, a semiconductor memory, an optical disk memory, or the like is used.

The evaluation value calculation unit 10 is connected to the score creation unit 20 and the information storage unit 40. The evaluation value calculation unit 10 calculates an evaluation value for each candidate pattern using the candidate pattern information stored in the candidate pattern storage unit 41 and the evaluation pattern information stored in the evaluation pattern storage unit 42. To do.

The evaluation values in the present embodiment are the contribution degree (first evaluation value), the shape of the character (second evaluation value), and the distance score (third evaluation value) for each candidate pattern. The calculation method of each of the contribution degree, the shape of the character shape, and the distance score will be described later. Each evaluation value calculated by the evaluation value calculation unit 10 is stored in the evaluation value information storage unit 43 for each candidate pattern.

The score creation unit 20 is connected to the evaluation value calculation unit 10, the pattern selection unit 30, and the information storage unit 40. The score creating unit 20 uses the evaluation value stored in the evaluation value information storage unit 43 and the weighting information for each evaluation value stored in the weighting information storage unit 45 for each evaluation value, and calculates a score for each candidate pattern. calculate. Note that the score calculation method of the present embodiment is calculated by a method of summing the product of the evaluation value and the weighting information for each evaluation value. However, a score obtained by quantizing the result into, for example, five levels may be used. The scores calculated by the score creation unit 20 are stored as score information in the score information storage unit 44 of the information storage unit 40 in the order of score.

The pattern selection unit 30 is connected to the score creation unit 20 and the information storage unit 40. Based on the score calculated by the score creation unit 20 described above, the pattern selection unit 30 stores, for example, a pattern used as a recognition dictionary of the pattern recognition device (hereinafter referred to as a dictionary pattern) stored in the candidate pattern storage unit 41. Select from patterns. Using the evaluation pattern storage unit 42 and the score information storage unit 44 of the information storage unit 40, a candidate pattern selection process described later is performed.

Hereinafter, an example of processing for selecting a dictionary pattern from the six candidate patterns shown in FIG. 2 in the pattern selection apparatus of the present embodiment will be described with reference to FIGS. 4 to 13.

FIG. 4 is a flowchart showing an example of the operation of the pattern selection apparatus of this embodiment. As shown in FIG. 4, in the pattern selection apparatus according to the present embodiment, first, the evaluation value calculation unit 10 performs an evaluation value calculation process in which the evaluation value of each candidate pattern included in the candidate pattern storage unit 41 is calculated.

Specifically, the evaluation value calculation unit first performs a contribution calculation process (step S1). Next, the evaluation value calculation unit 10 performs a character shape calculation process (step S2). Subsequently, the evaluation value calculation unit 10 performs a distance score calculation process (step S3). Each evaluation value calculated in steps S1 to S3 is stored in the evaluation value storage unit 43 for each candidate pattern by the evaluation value calculation unit 10.

The score creation unit 20 performs a score calculation process using the calculated evaluation value and the weighting information for each evaluation value stored in the weighting information storage unit 45 for each evaluation value (step S4). The pattern selection unit 30 selects a dictionary pattern from the candidate pattern storage unit 41 using the score calculated in step S4 and the pattern selection reference information stored in the pattern selection reference information storage unit 46 (step S5). As a result, the pattern selection apparatus ends the operation.

Here, with reference to FIG. 5 to FIG. 8, the evaluation value calculation processing in step S1 to step S3 of FIG. 4 will be described.

FIG. 5 is a flowchart showing an example of the procedure of the contribution calculation process in step S1 of FIG.

First, the evaluation value calculation unit 10 selects a specific evaluation pattern as a sample from the evaluation patterns stored in the evaluation pattern storage unit 42, and uses the k-nearest neighbor (k-nearest neighbor) using the candidate pattern of FIG. pattern recognition by algorithm (knnn method) is performed (step S11). That is, the selected evaluation pattern is recognized using all candidate patterns as a recognition dictionary. The knn method is a technique often used in pattern recognition.

The evaluation value calculation unit 10 calculates the number of appearances and the number of correct answers for each candidate pattern (step S12). The number of appearances is the number of times that the target candidate pattern appears in the vicinity of the evaluation pattern as a sample, and the number of correct answers is the number of times that the target candidate pattern appears in the vicinity of the evaluation pattern of the same class.

When pattern recognition is performed for all the evaluation patterns registered in the evaluation pattern storage unit 42 (Yes in step S13), the evaluation value calculation unit 10 calculates the accumulated number of the number of appearances and the number of correct answers for each candidate pattern. Calculate (step S14). That is, for the target candidate pattern, the number of appearances and the number of correct answers for each candidate pattern calculated in step S12 are summed.

The evaluation value calculation unit 10 calculates the contribution by the equation (A) in FIG. 6 using the cumulative value of the number of appearances and the cumulative value of the number of correct answers for each candidate pattern (step S15), and ends the contribution degree calculation process. To do.

When not all the patterns for evaluation are recognized, that is, when there is an evaluation pattern for which step S11 is not performed in the pattern information for evaluation registered in the pattern storage unit for evaluation 42 (No in step S13). ), The evaluation value calculation unit 10 uses the next evaluation pattern that has not been subjected to pattern recognition as a sample (step S16), and returns to step S11 to perform pattern recognition. However, the evaluation pattern used in the evaluation value calculation unit 10 may be limited to a part arbitrarily selected from the evaluation patterns stored in the evaluation pattern storage unit 42. In this case, the above-described processing is not performed for evaluation patterns that are not targeted.

As is clear from the equation (a), the contribution is a ratio in which the evaluation pattern and the candidate pattern are in the same class when the corresponding candidate pattern is selected as the vicinity of the evaluation pattern. In other words, it is the ratio of patterns in the same class as the candidate pattern among the evaluation patterns in the vicinity of the candidate pattern. That is, it can be considered that many of the same classes as the candidate pattern are distributed in the vicinity of the candidate pattern having a high contribution, and as a result, it is appropriate to identify the input pattern as the same class as the candidate pattern. Conceivable.

That is, a candidate pattern with a high contribution contributes to determining the correct class of the input pattern.

In this embodiment, the cumulative value of the number of appearances and the number of correct answers is used as the evaluation value calculation criterion. However, the closer the distance between the input pattern and the candidate pattern selected as the neighborhood, the more likely they are in the same class. In consideration of the high possibility, the contribution may be calculated based on the distance, such as the sum of the reciprocal of the distance between the candidate pattern and the input pattern. In addition, the distance here is a distance used in the Knn method.

The calculation in this case is performed as follows, for example. When a certain candidate pattern x is selected as the vicinity of the evaluation pattern y _i , the distance D _x, y _i between x and y _i is set to the pixel at the same position in the candidate pattern x and the evaluation pattern y _i. It is assumed that the square of the difference of values is obtained and summed, and the square root is taken.

Here, the above evaluation with respect to D ₁ , which is the sum of all the evaluation patterns y ₁ , y ₂ , y ₃ ,..., _{N n in} which the candidate pattern x appears in the vicinity of the reciprocal of the distance. use evaluation pattern _y 1 class is the same as the x of the _{patterns, y 2, y 3, ···} , the ratio of _{D 2} may be contribution is the total value only for _{y m.}

If D _{x, yi} is smaller than a predetermined threshold value Th _, the value of D _{x, yi} may be set to Th. In addition, when there is a candidate pattern whose distance from the evaluation pattern is 0, that is, when the evaluation pattern completely matches a certain candidate pattern, the evaluation pattern is excluded from the above calculation and the contribution is obtained. Alternatively, the contribution may be 100. According to the calculation as described above, the calculation of the contribution can be greatly influenced as the distance between the evaluation pattern and the candidate pattern is shorter. Next, with reference to FIG. 7, the character shape calculation processing in step S2 of FIG. 4 will be described. FIG. 7 is a flowchart illustrating an example of a procedure for calculating the shape of the character by the evaluation value calculation unit 10.

First, the evaluation value calculation unit 10 normalizes the image size and pixel gradation of the candidate pattern (step S21). These normalizations are performed according to predetermined criteria. In this embodiment, the pixel gradation is normalized to two gradations in which a pixel with characters is black and a pixel without characters is white.

Next, pixels at the same position in the normalized candidate pattern and the template pattern are compared, and the number of pixel positions where black pixels appear in both are calculated (step S22).

  The template of the present embodiment is a standard pattern in the class to which the candidate pattern of the present embodiment belongs, and is created from candidate pattern information in the candidate pattern storage unit 41 in advance. For example, a pattern having a slope of 0 is selected for each class or font from the candidate pattern information. The average of the selected candidate patterns is used as a template. When there are a plurality of selected patterns, one of them may be selected as a template. In that case, for example, a selected pattern having the shortest distance from the average pattern of all the selected patterns may be selected. That is, a candidate pattern determined to be the most standard target for the user to perform pattern recognition is set in the template.

  In addition to the creation method described above, the template may be set in advance from the evaluation pattern group shown in FIG. 3, or a pattern different from the evaluation pattern group may be registered in advance as a template. good. For example, one candidate pattern without inclination for each class or font is prepared in advance, a template tag is attached, and when a template is required, the candidate pattern with this template tag may be used as a template.

Here, among the candidate patterns shown in FIG. 2, a class A template is set as candidate pattern 1, and a class B template is set as candidate pattern 4.

Next, the evaluation value calculation unit 10 calculates the number of pixel positions at which black pixels appear in both of the normalized candidate pattern and the template pattern at the same position (step in FIG. 7). S22).

Next, the evaluation value calculation unit 10 calculates the number of pixel positions where a black pixel appears in at least one of the normalized candidate pattern and the pixel at the same position in the template pattern (FIG. 7). Step S23).

The evaluation value calculation unit 10 sets the number of pixel positions where black pixels appear at the same pixel position calculated in step S22 in FIG. 7 as X, and calculates black in either one calculated in step S23 in FIG. The number of pixel positions where the pixel appears is Y, and the shape of the character shape is obtained by the equation (b) in FIG. 6 (step S24). Thereby, the evaluation value calculation unit 10 ends the character shape calculation process.

That is, the shape of the character shape is an evaluation value indicating how similar to the pattern that the user has determined to be standard as an object for pattern recognition.

Next, a distance score calculation process by the evaluation value calculation unit 10 will be described with reference to FIG. FIG. 8 is a flowchart showing an example of the distance score calculation process.

First, the evaluation value calculation unit 10 normalizes the image size and pixel gradation of the candidate pattern (step S31). These normalizations are performed according to predetermined criteria. Note that the processing in step S31 is the same as the processing in step S21 in the character shape calculation processing. For this reason, the data after normalization in step S21 may be used without performing the process in step S31.

The evaluation value calculation unit 10 creates an average pattern in the class using the normalized candidate pattern (step S32). The average pattern is created by obtaining an average value of pixel values of candidate patterns in the same class at the same position.

Next, the evaluation value calculation unit 10 calculates the distance from the average pattern created in step S32 for each candidate pattern (step S33). The distance here is obtained by calculating and summing the square of the difference between the pixel values for each pixel at the same position in the candidate pattern and the average pattern, and taking the square root of the total value.

The evaluation value calculation unit 10 converts the calculated distance by log (step S34), normalizes the converted value to a range from 0 to 100 (step S35), and ends the distance score calculation process. This normalized value is used as a distance score in the same class.

The average pattern is the center of the distribution of the class in a given candidate pattern. In other words, it can be considered that the class is distributed around the average pattern. That is, a candidate pattern with a small distance score is a pattern that is close to the average pattern, that is, located near the center of the distribution of the class, in other words, a pattern that seems to be a class from the candidate pattern. Conversely, a candidate pattern with a large distance score is a candidate pattern that is located at a location outside the center of the distribution of the class, in other words, a candidate pattern that is considered exceptional in that class.

After completion of the distance score calculation process, the evaluation value calculation unit 10 registers these evaluation values in the evaluation value information storage unit 43 as an evaluation value table. FIG. 9 shows an example of the evaluation value table. As shown in FIG. 9, the evaluation value table has items of pattern ID 431, contribution degree 432, character shape adjustment 433, and distance score 434, and evaluation values for each candidate pattern are stored. The pattern ID 431 stores the ID of the candidate pattern shown in FIG.

The contribution degree, the shape of the character shape, and the distance score may be calculated sequentially or simultaneously. In addition to the method described above, any method may be used as a method for calculating the contribution degree, the shape of the character shape, and the distance score.

The score creation process in step S4 of FIG. 4 will be described with reference to FIGS. FIG. 10 is a flowchart illustrating an example of score creation processing by the score creation unit 20.

The score creation unit 20 calculates a score with reference to the evaluation value stored in the evaluation value information storage unit 43 and the weighting information storage unit 45 for each evaluation value (step S41). Here, the total value of the values obtained by multiplying each evaluation value and the weighting information for each evaluation value is treated as the score of the candidate pattern.

Here, the weighting information for each evaluation value will be described with reference to FIG. The weighting information for each evaluation value is set in advance by the user or the pattern selection device creator of the present embodiment, and stored in the weighting information storage unit 45 as a table. As shown in FIG. 11, the weighting information for each evaluation value according to the present embodiment is a value between 0 and 1 for each of the contribution degree, the shape of the character shape, and the distance score, and the total value is “ 1 "is set. By setting the weighting information for each evaluation value, the pattern selection device according to the present embodiment enables selection of candidate patterns according to the user's request.

In pattern recognition, since the recognition rate of patterns similar to the candidate pattern selected as the dictionary creation pattern is high, it is determined what pattern is recognized with high recognition rate by weighting when calculating this score To do. For example, if the contribution weight is increased, a candidate pattern with a small number of candidate patterns in other classes is given priority, and the recognition rate for a pattern that is likely to be that class, in other words, a pattern that is unlikely to be another class. Becomes higher. Increasing the shape of the character shape increases the percentage of patterns that are similar to the patterns that the user has determined to be standard in the class. The rate can be expected to improve. Increasing the weight of the distance score increases the percentage of exceptional dictionary patterns in the class, so it can be efficiently used when there are large variations in the class, such as handwritten characters, or when there is noise in the image. It can be expected to improve the recognition rate.

Returning to the description of FIG. The score creation unit 20 creates a score table by registering the scores obtained in step S41 in descending order of numerical values (step S42), and registers the created score table in the score information storage unit 44 (step S43).

For example, a case where the score of candidate pattern 1 in FIG. 2 is calculated will be described. As shown in FIG. 9, candidate pattern 1 in FIG. 2 has a contribution degree of 75, a letter shape of 100, and a distance score of 80. As shown in FIG. 11, the contribution weighting information is 0.5 for the contribution, 0.25 for the shape adjustment, and 0.25 for the distance score. Since the score is a total obtained by multiplying the respective scores, the score of the candidate pattern 411 is “82.5”. Scores are similarly obtained for other candidate patterns.

FIG. 12 shows an example of the score table according to the present embodiment. As shown in FIG. 12, the score table has items of pattern No., pattern ID, and score. For No, the registration order of scores is registered. The ID of the registered candidate pattern is registered as the pattern ID. As the score, the score calculated by the score creating unit 20 is registered in correspondence with the pattern ID.

  When the score creation unit 20 creates the score table, the pattern selection unit 30 of the present embodiment refers to the pattern selection criterion 46 and selects dictionary patterns in order from the top of the score table. FIG. 13 shows an example of pattern selection criteria.

  As shown in FIG. 13, the number for selecting a dictionary pattern is registered in the pattern selection criterion. Since it is “3” in the present embodiment, the pattern selection unit 30 selects candidate patterns No. 1 to No. 3 in the score table of FIG. 2 as dictionary patterns used to create a recognition dictionary.

  In the present embodiment, the dictionary is a set of selected dictionary creation patterns. As a dictionary creation method, a dictionary creation method adapted to a different recognition method such as a subspace method may be selected.

  Thus, according to this embodiment, a pattern can be efficiently selected by selecting a pattern from a candidate pattern based on the score calculated with respect to each candidate pattern. Furthermore, the criteria for selecting a pattern can be clarified.

  For example, according to the present embodiment, it is possible to select a pattern near the class boundary and a standard pattern that match the balance specified by the weighting information. Therefore, it is possible to create a recognition dictionary that includes a pattern near the boundary between classes and a standard pattern in the class in a balanced manner. Also according to this embodiment. It is possible to reduce the patterns used in the recognition dictionary without reducing the recognition accuracy of pattern recognition.

  Note that the evaluation pattern, the weighting information for each evaluation value, and the pattern selection criterion can be changed according to the object that the user wants to recognize.

(Second Embodiment)
Hereinafter, a pattern selection apparatus according to the second embodiment will be described with reference to FIGS. 14 to 17. In addition, the same code | symbol is attached | subjected to the structure same as 1st Embodiment, and description is abbreviate | omitted.

FIG. 14 is a block diagram illustrating the overall configuration of the pattern selection apparatus according to the second embodiment.

As shown in FIG. 14, the pattern selection device of the present embodiment has a reselection determination unit 31, a recognition dictionary creation unit 32, an update condition storage unit 47, and an update information storage unit 48 in the configuration of the pattern selection device shown in FIG. 1. The information input unit 50 and the information update unit 60 are further included. Note that the reselection determination unit 31 and the recognition dictionary creation unit 32 of the present embodiment are included in the pattern selection unit 30.

  The information input unit 50 is used by the user to input update information. The information input unit 50 is a keyboard or a mouse, for example.

  The update condition storage unit 47 stores an update condition of a pattern selection condition described later. This update condition is used in a pattern update process described later.

  The update information storage unit 48 stores update information to be described later. The update information includes the item name of the information to be updated and the changed information. In the present embodiment, the update information and the update condition are information for updating the pattern selection criterion in the pattern selection criterion information storage unit 46. In addition to the pattern selection reference information, selection condition information, the number and type of evaluation values, weighting information for each evaluation value, a score calculation method, and the like may be updated.

  The reselection determination unit 31 determines whether to perform pattern selection again based on the update condition.

  The recognition dictionary creation unit 32 creates a recognition dictionary used in pattern recognition by the pattern recognition device based on the pattern selected by the pattern selection unit 30.

  The information update unit 60 updates the pattern selection criterion based on the update information stored in the update information storage unit 48.

  Hereinafter, with reference to FIG. 15, a pattern selection reference update processing procedure when creating a recognition dictionary of the pattern recognition apparatus using the pattern selection apparatus of the present embodiment will be described.

  FIG. 15 is a flowchart showing an example of a pattern selection condition update procedure when creating a recognition dictionary according to the pattern selection device of this embodiment.

  First, the user uses the information input unit 50 to input pattern selection reference update information and update conditions (step S51).

  Here, FIGS. 16 and 17 show an example of the pattern selection update condition 471 and the pattern selection reference update information 481. As shown in FIG. 16, the pattern selection update condition 471 has a maximum number of patterns 472 and a minimum recognition rate 473. The maximum number of patterns 472 indicates the maximum number of patterns selected from the candidate pattern group by the pattern selection unit 30. The minimum recognition rate 473 defines the minimum recognition rate required when an evaluation pattern group is recognized using a recognition dictionary created based on the pattern selected by the pattern selection unit 30. . The recognition rate is, for example, the number of correct pattern recognitions with respect to the number of times the pattern is recognized.

  As shown in FIG. 17, the pattern selection reference update information 481 includes a pattern selection number change unit 482. The pattern selection number change unit 482 defines an increase / decrease value of the number of patterns when the pattern selection unit 30 selects a pattern again.

  The input pattern selection update condition 471 is stored in the selection condition information storage unit 47 of the information storage unit 40, and the update information 481 is stored in the update information storage unit 48 of the information storage unit 40 (step S52).

  Subsequently, the pattern selection apparatus of the present embodiment performs the pattern selection process shown in FIG. 4 and selects a dictionary pattern from the candidate pattern group (step S53). The recognition dictionary creation unit 32 of the pattern selection unit 30 creates a recognition dictionary using the selected dictionary pattern (step S54).

  When the pattern selection unit 30 creates a recognition dictionary, the reselection determination unit 31 performs pattern recognition of the evaluation pattern stored in the evaluation pattern storage unit 42 using the created recognition dictionary. This pattern recognition is performed using, for example, the knn method (step S55). Note that the evaluation pattern used in step S55 may be an evaluation pattern used in the pattern selection process in step S53, or may be registered in advance for the pattern selection condition update process. Alternatively, the reselection determination unit 31 may perform pattern recognition using an external pattern recognition device instead of performing pattern recognition, and the reselection determination unit 31 may use the recognition result.

  The reselection determination unit 31 calculates a recognition rate when pattern recognition is performed using the recognition dictionary created by the recognition dictionary creation unit 32 (step S56). When the calculated recognition rate satisfies the pattern selection update condition 471 input in step S51 (Yes in step S57), the reselection determination unit 31 ends the pattern selection reference update process without performing pattern reselection. That is, when the calculated recognition rate exceeds the minimum recognition rate 473 of the update condition 471, the pattern selection condition update process is terminated.

  When the calculated recognition rate does not satisfy the pattern selection update condition 471 input in step S51 (No in step S57), the reselection determination unit 31 determines to update the pattern selection reference, and based on the determination result, The pattern selection unit 30 updates the pattern selection reference stored in the pattern selection reference storage unit 46 using the update information 481 (step S58). Specifically, the pattern selection reference pattern selection number is added based on the pattern selection reference change unit 482 of the pattern selection reference update information 481. Thereafter, the pattern selection apparatus according to the present embodiment repeats the processing from step S53 using the updated pattern selection reference.

  As described above, according to the present embodiment, the pattern selected by the pattern selection unit 30 based on the pattern selection criterion is determined based on the update condition of the pattern selection criterion, and the pattern to be added based on the determination result. It is possible to select with reference to the calculated score. That is, when the recognition target is recognized as a pattern, the minimum number of dictionary creation patterns that can create a recognition dictionary that satisfies the recognition rate required by the user can be selected from the candidate patterns.

  As mentioned above, although some embodiment of this invention was described, these embodiment is shown as an example and is not intending limiting the range of invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

  In the first and second embodiments, the evaluation pattern is selected from the candidate pattern group. In this case, the evaluation pattern can be narrowed down using an information tag. For example, those having an inclination of ± 10 or more may be deleted. Alternatively, an upper limit may be set for the number of patterns for evaluation selected from the candidate pattern group, and the patterns may be deleted randomly for each class. This is to prevent the same pattern from being included in the dictionary pattern and the evaluation pattern in the pattern selection device of the second embodiment. In this case, selection may be made so that there is no bias in the number of patterns for each class or font.

  In the first and second embodiments, the pattern used for the character recognition device has been described as an example. However, in addition to characters, it can also be used for patterns such as images and sounds such as faces and fingerprints. .

  In the first and second embodiments, the creation of a recognition dictionary used for pattern recognition is given as an example. For example, which test pattern is important by calculating the test pattern score used for the test of the pattern recognition device You may do something like that.

In this example, for example, when there is a test pattern that has not been recognized, the score of the test pattern is calculated.

For example, when the weight of the distance score is increased in the calculation of the score, a large score means that it is a very exceptional pattern, and if such a pattern can be correctly recognized, until now Detrimental effects such as erroneous recognition of test patterns that could be recognized normally may occur. For this reason, when the score exceeds a preset threshold, the test pattern is deleted from the test pattern set. As a result, it is possible to avoid that the recognition rate is evaluated to be lower than actual because there is an extremely exceptional pattern.

Further, it is not necessary to limit to test patterns that have not been recognized, and it is also possible to obtain scores for all the test patterns and determine which test patterns are regarded as important according to the scores. In this case, for example, using the number of recognized patterns and the number of correct patterns when the recognition rate is set may be used as the sum of the scores of the recognized patterns and the sum of the correct patterns.

Note that, when calculating the test pattern score, the contribution cannot be calculated unlike the candidate pattern. Therefore, it is possible to substitute by using another evaluation value calculation method such as calculating the ratio of candidate patterns belonging to the same class as the test pattern in the 10 patterns near the test pattern. Alternatively, a prescribed value, for example, 0 may be used as the contribution degree.

DESCRIPTION OF SYMBOLS 1 ... Pattern selection apparatus, 10 ... Evaluation value calculation part, 20 ... Score preparation part, 30 ... Pattern selection part, 40 ... Information storage part, 41 ... Candidate pattern information, 42 ... Pattern information for evaluation, 43 ... Evaluation value information storage Unit 44 ... Score information storage unit 45 ... Weighting information, 46 ... Pattern selection reference information

Claims (6)

A storage unit storing a first pattern group and a second pattern group used for evaluation of the first pattern;
With reference to the second pattern, as a plurality of evaluation values of the first pattern, a first evaluation value that is a contribution degree when the first pattern recognizes the second pattern, A second evaluation value for evaluating whether the first pattern is a standard pattern in a class to which the first pattern belongs, and whether the first pattern is a pattern near the boundary between classes. An evaluation value calculation unit for calculating a third evaluation value to be evaluated;
A score calculation unit that calculates a score for each of the first patterns using the calculated evaluation value and weighting information for each evaluation value;
A pattern selection unit that selects a pattern from the first pattern based on the calculated score;
A pattern selection device comprising:   The pattern selection apparatus according to claim 1, wherein the second pattern group is a pattern selected based on a pattern recognition target from the first pattern group.   The pattern selection device according to claim 1, wherein the pattern selection unit selects a prescribed number of the first patterns in descending order of the scores calculated by the score calculation unit.   The pattern selection apparatus according to claim 1, further comprising a recognition dictionary creation unit that creates a pattern recognition dictionary from the pattern selected by the pattern selection unit. In a pattern selection device comprising a storage unit storing a first pattern group and a second pattern group used for evaluation of the first pattern,
With reference to the second pattern, as a plurality of evaluation values of the first pattern, a first evaluation value that is a contribution degree when the first pattern recognizes the second pattern; A second evaluation value for evaluating whether the first pattern is a standard pattern in a class to which the first pattern belongs; and whether the first pattern is a pattern near the class boundary. Calculating a third evaluation value to be evaluated;
Calculating a score for each of the first patterns using the calculated evaluation value and weighting information for each of the evaluation values;
Selecting a pattern from the first pattern based on the calculated score;
A pattern selection method comprising: A program in a pattern selection device including a storage unit storing a first pattern group and a second pattern group used for evaluation of the first pattern,
On the computer,
A plurality of evaluation values of the first pattern with reference to the second pattern, and a first evaluation value that is a contribution degree when the first pattern recognizes the second pattern; A second evaluation value for evaluating whether the first pattern is a standard pattern in a class to which the first pattern belongs, and whether the first pattern is a pattern near the boundary between classes. A function of calculating a third evaluation value for evaluating
A function for calculating a score for each of the first patterns using the calculated evaluation value and weighting information for each of the evaluation values;
A function of selecting a pattern from the first pattern based on the calculated score;
A program that realizes

Images