JPH08272904A - Pattern recognition device - Google Patents

Abstract

PURPOSE: To perform recognition which is highly reliable for various candidate characters by providing a specific dictionary generating means and a specific pattern recognizing means. CONSTITUTION: An image input part reads out an image to be recognized first (S101) and segments characters out of the read image (S102). Then, a character feature extraction part performs feature extraction (S103) and collates the input character feature quantities with a rough classification dictionary (S104). On the basis of the collation results, a feature extracting method matching with each input character is selected out of a detailed classification dictionary (S105), and feature extraction for detailed classification is performed by the feature extracting means (S106). And, the extracted input character feature quantities are collated with the detailed classification dictionary (S107) and the results are outputted (S110). Consequently, even when an object to be recognized and a character feature extracting method are added, recognition and evaluation are only performed by the dictionaries and feature extracting methods after the addition and the maintenance of the dictionary on a dictionary generation side is facilitated to easily perform high-precision recognition.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition apparatus, and more particularly to automatic construction of a detailed classification dictionary which maintains a character recognition rate and reduces a misreading rate.

[0002]

2. Description of the Related Art Heretofore, a character recognition device has been proposed which uses a single character feature amount in detailed classification of similar recognition candidate characters and extracts and collates a characteristic component suitable for classification based on the variance of each characteristic component. (JP-A-1-233678).

[0003]

SUMMARY OF THE INVENTION In such a device,
The feature amount suitable for classification often differs between candidate characters, and there is a problem that sufficient recognition accuracy cannot be obtained with a single character feature amount.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a recognition device which can recognize various candidate characters at high speed and with high reliability.

[0005]

The features of the device of the present invention are:
For the character pattern, in the detailed classification after the major classification, a plurality of feature extraction methods are set, the classification determination threshold value is calculated for each feature extraction method, and the classification accuracy is evaluated. A feature extraction method having the smallest evaluation value, a dictionary creating means for storing a determination threshold value and an evaluation value of classification accuracy as a dictionary, and a feature extraction method suitable for classification between similar candidates in the large classification result. Read the threshold value from the dictionary,
It has a pattern recognition means for performing detailed classification.

That is, the present invention is designed to automatically construct a dictionary capable of selecting a characteristic amount and a characteristic component suitable for detailed classification for each candidate character. The character transformation pattern is automatically generated and prepared in various types. The character feature amount is recognized and evaluated, and as a result, one character feature amount having the highest character recognition rate and the lowest misreading rate is selected for each candidate character, and the character feature extraction method name and the determination threshold value are selected. And the evaluation value of the classification accuracy is stored as a detailed classification dictionary.

Then, using this dictionary, a feature amount and a feature component suitable for each candidate character are selected to perform detailed classification.

[0008]

With this configuration, even if a new recognition target or character feature extraction method is added, if a dictionary and feature extraction method after addition are recognized and evaluated, the new detailed classification dictionary is updated as a result. It is possible to maintain the dictionary easily, and it is easy to add a new character feature extraction method for detailed classification.

Further, since the classification accuracy based on the recognition result is applied as a means for judging the feature extraction method suitable for recognition, the simplification and the automatic registration are possible.

Further, since the input pattern is deformed and the classification accuracy is evaluated in consideration of the deformation, the recognition accuracy can be improved even for the character deformed pattern and the unknown pattern. Furthermore, by determining the feature extraction method and the determination threshold value between the patterns of the two categories, high classification accuracy can be realized.

Instead of selecting one from a plurality of character feature extraction methods, it is also possible to classify by a plurality of character feature extraction methods and use the respective output values (distance values) for classification determination.

[0012]

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

This pattern recognition device first automatically generates a character deformation pattern, recognizes and evaluates it with a variety of prepared character feature amounts, and as a result, has a highest character recognition rate for each candidate character and a lowest misreading rate. A dictionary creating means for selecting one possible character feature amount and storing the character feature extracting method name, the judgment threshold value, and the evaluation value of the classification accuracy as a detailed classification dictionary, and this dictionary are used. A pattern recognition means for selecting a feature quantity and a feature component suitable for each candidate character and performing detailed classification is provided.

Next, a recognition process using this pattern recognition device will be described with reference to the flow chart shown in FIG.

First, the image to be recognized by the image input unit is read (step 101).

Next, in the character cutting section, characters are cut out from the read image (step 102).

After this, the character feature extraction section performs feature extraction (step 103).

Then, the input character feature amount is compared with the large classification dictionary (step 104). It is assumed that this large classification dictionary has been created in step 108 in advance.

Similarly, it is assumed that the detailed classification dictionary is also created in advance in step 109.

Based on the comparison result with the large classification dictionary, a feature extraction method suitable for this input character is selected from the detailed classification dictionary (step 105), and the feature extraction for detailed classification is performed based on this feature extraction method. (Step 106).

Then, the input character feature amount which has been subjected to feature extraction by an appropriate method is compared with the detailed classification dictionary (step 107), and the result is output (step 110).

As shown in the following table, the feature extraction method has two types of feature values, that is, the density gradient feature has no intensity weighting, and the feature vector has partial features, full features, three types of vertical symmetry features, and bilateral symmetry. 20 types of feature extraction methods of 2 × 10 × 1 are used with 10 types of 3 degree features, 2 types of diagonal symmetry features, and 1 type of threshold setting.

[0023] This characteristic value will be described. When the input image has 256 gradations, the density gradient strength value is 1024 gradations, which is a value obtained by normalizing the density gradient strength value to a binary value or a multi-value (8 levels) (step 304 in FIG. 10). ) The feature vector is shown below.

[0024] Next, the creation of the large classification dictionary will be described with reference to the flowchart of FIG.

This flowchart corresponds to the large classification dictionary creating step 108 in the flowchart of FIG.

First, as shown in FIG. 3, character data of categories 0 to 9 and symbols for the fonts A to N is input (step 201).

Then, as shown in the example of FIG. 4, six types of modified patterns are generated for the reference pattern (step 202).

Then, character feature extraction (step 20)
3) is performed and the character feature amount is stored (step 20).
4).

Then, it is judged whether or not the generation of all the deformation patterns is completed (step 205), and if completed, the average of the character feature amount and the standard deviation are calculated for each feature dimension (step 206).

On the other hand, if not completed, the process returns to the modified pattern generation step 202.

The font of the input character data, the average character feature amount of the category, and the standard deviation for each feature dimension are stored in the large classification dictionary (step 207).

Then, it is judged whether or not the processing is completed for all fonts and categories (step 208), and if it is completed, this flow is ended, and if the processing is not completed, the first character data input step 201 is executed. Return to.

Next, the creation of the detailed classification dictionary will be described with reference to FIG. First, major classification is performed. As described with reference to FIG. 2, character data of categories 0 to 9 and symbols for fonts A to N is input (step 40).
1).

Then, a deformed pattern is generated with respect to the reference pattern (step 402). Then, character feature extraction (step 403) is further performed, and collation with the large classification dictionary is performed (step 404).

Then, classification is performed by using the feature extraction methods _{1 to} M (steps 4051 to 405 _m ), and the feature extraction method is classified into the first rank, the correct answer is classified according to the combination of the second rank, the error distance value and the distance value. Summarize the frequency distribution of the differences (Step 4
06). FIG. 6 shows the relationship between the absolute distance value of the first detailed classification and the relative distance value (difference in absolute distance between the first and second detailed classifications) when one kind of feature extraction method is used. ,
It is plotted by error.

[0036] Then, it is judged whether or not these processes have been completed for all the deformation patterns (step 407), and if not completed, the process returns to the deformation pattern generation step 402, but if completed, the processes are completed for all fonts and categories. It is determined whether or not (step 408).

If it is judged in the judgment step 408 that the processing is not completed, the procedure returns to the beginning, and if it is judged that the processing is completed, the threshold function is calculated (step 409). Detailed classification by feature extraction method Here, correct answers for each combination of first and second positions, error distance value frequency distribution (Fig. 6)
Then, the threshold function is calculated using the Mahalanobis distance shown below.

[0038] For the threshold function, the locus of points having the same Mahalanobis distance between the correct answer group and the error group, that is, the boundary between both groups is obtained, and the result of temporary regression of the point sequence is used as the threshold function. Threshold function (relative distance) -a * (absolute distance) -b = 0 When the value of the threshold function thus obtained is 0 or less, it is rejected.

FIG. 7 is an explanatory diagram for determining the threshold function.
Then, the evaluation value V of classification accuracy is calculated, and the feature extraction method is determined based on this value. (Step 410) When determining the feature extraction method, first, the evaluation value V of the classification accuracy is calculated by the following formula between the reference table of the first rank in the large classification and the reference table of the second rank, and the value is minimized. Decide on the method.

Here, the evaluation value V = α × E + R (E: error rate, R: reject rate, α: constant) The final detailed classification dictionary is a reference table of major classification first-ranked candidates and letters of second-ranked candidates. One kind of feature extraction method with high recognition accuracy is stored for each kind combination (number of large classification reference tables × number of character types) (step 411).

Here, the value of α is 10, and the method of calculating the error rate and the reject rate will be described. FIG. 8 shows an example in which the major classification first-ranked candidate is a combination of font A, category 0 and the major classification second-ranked candidate is category 8, and the feature extraction method uses feature values without weighting and full-features as feature vectors. .

Here, the threshold function calculated using the Mahalanobis distance from the group of correct answers and errors collected according to FIG. 5 (relative distance)-(0.24604) × (absolute distance)-(158.64) Substitute the relative absolute distance value for each input pattern into = 0. The frequency of the input pattern with a negative value is the reject rate R, and the frequency with which the value of the threshold function is positive in the error group (the input pattern in which the first rank of the detailed classification is an error) is the error rate E.
And

Next, an example of processing using this method will be described. Here is an example of the processing when the large category first place candidate and the category different from the first place are second place candidates. First, as shown in FIG. 9, the character patterns are roughly classified, and candidates for the upper M ranks, for example, the 2nd ranks are selected (step 50).
1).

Then, the detailed classification dictionary is referred to, and the feature extraction method, the feature amount, and the threshold function for the combination of the large classification first rank candidate C1 and the second rank candidate C2 are input from the detailed classification dictionary (step 502).

Then, the detailed classification feature of the input character is extracted (step 503). This will be described in detail later.

Thereafter, distance calculation is performed between the detailed classification feature amounts of the first and second major classifications and the detailed classification feature amount of the input character, and the minimum distance value D (absolute distance) and the first distance value are calculated. The difference B (relative distance) between the second place and the second place is calculated (step 50).
4).

Then, the minimum distance value D (relative distance) and the difference B (relative distance) between the distance values are substituted into the above-mentioned threshold function (step 505), and the output value of the threshold function is larger than 0. It is determined whether there is any (step 506), and when it is greater than 0, it is accepted (step 507). On the other hand, if it is smaller than 0, reject (step 50
8).

Here, the process of extracting the character feature of the input character will be described. First, as shown in FIG. 10, the image to be recognized by the image input unit is read (step 301).

Next, after smoothing the read image (step 302), the edge extraction operator 303 is applied to detect the density gradient strength of the pixel of interest, and the density of the pixel of interest is determined in the direction of maximum intensity among the eight directions. The gradient direction is set, and the intensity value is set as the concentration gradient intensity value. Then, a threshold value for determining the feature amount from the density gradient intensity distribution is set (step 304), the pattern is divided into M × N regions (step 305), and the feature values are totaled for each region (step 3).
06), the feature amount is calculated for the case without weight and the case with weight (steps 307 and 308).

In the processing example shown in FIG. 9, the processing between the large-class first-ranked candidate and the second-ranked candidate (a category different from the first-ranked candidate) has been described, but a case where the range of the candidate to be applied is expanded will be described. .

First, the detailed classification process 1 (FIG. 9) is applied to all combinations of candidate characters in the same category as the first major classification and category candidate characters different from the first major classification. However,
Judgment by the threshold function, that is, rejection is not set.

If there is a candidate character that wins all detailed classification processing with a candidate character of a different category, that is, a candidate character having a distance value smaller than all the candidates to be compared, the category of the candidate character is output, and if there is no reject character, it is rejected. To do.

For example, an example is shown in which the processing is applied to candidates up to the rank of three types of candidate categories in the large classification. As shown in the table below, "Font B," which is a candidate character in the same category as "Category 0 of font A", which is ranked first in the major categories.
Category 0, font D category 0, font F category 0 "and" C category 6 of font C, category 6 of font E, category 8 of font G "which are candidate characters of categories different from the first place in the major classification. The detailed classification process shown in FIG. 9 is performed for all combinations of.

[0054] In this table, if the distance value in the detailed classification feature is smaller than the candidate character to be compared, it is expressed as "winning ○", and if it is larger,
It is expressed as "Loss x". "Font B", which is the second largest category
“Category 0” indicates that the distance value in the detailed classification feature is smaller than all the different category candidate characters and that all the wins have been achieved. Therefore, the final candidate character is output in the font 0 category 0.

Another processing example will be described. First, brute force detailed classification processing between candidate characters up to the nth major classification 1
(Figure 9) applies. However, the judgment by the threshold function, that is, the rejection is not set.

Similar to the above example, if there is a candidate character that wins all the detailed classification processing with candidate characters of different categories, that is, a candidate character having a distance value smaller than all the candidates to be compared, the category of the candidate character is output. If not, reject it. Reject if multiple categories that win all games occur.

For example, an example is shown in which the processing is applied to the top 5 candidate characters in the major classification. In this table, if the distance value in the detailed classification feature is smaller than the candidate character to be compared, it is expressed as "winning ○", and if it is larger,
It is expressed as "Loss x". "Font B", which is the second largest category
“Category 0” indicates that the distance value in the detailed classification feature is smaller than all the different category candidate characters and that all the wins have been achieved. Therefore, the final candidate character is output in the font 0 category 0.

According to this method, since it is possible to make a determination for each combination of large classification results using a feature extraction method suitable for classification, it is possible to improve the overall recognition accuracy.

For example, when 0 having a partial stain is input as shown in FIG. 11A, when character features without feature value weighting are used as shown in FIG. The distance value is "8"<"0" and the output result is an error.
The distance value in the detailed classification using the partial features with feature value weighting as in 1 (c) is “8”> “0”, and the output result is the correct answer or the reject.

The first major category is category 8 (font i)
The following table shows the result of the detailed classification dictionary creation process (FIG. 5) when the second rank is the category 0 combination.

[0061] The feature extraction method in which the evaluation value V is the minimum, that is, the reject and error rates are small, is the feature value weighting and the partial feature. The detailed classification dictionary stores: -feature extraction method and feature amount-threshold function and evaluation value V = 5.0.

Actually, for the combination of the category 8 (font i) in the first major category and the category 0 in the second major category, the contents of the input patterns that are likely to cause rejects and errors in the output results of the major categories are as shown in FIG. There was a tendency that the center of 0 was partially soiled or there was a background pattern. It is considered that this tendency can be classified if stains and unstable information of the background picture can be represented by ambiguity.

In this case, since the feature extraction method in which unstable feature values such as faintness and dirt are expressed in multivalues and the ambiguity is used for classification determination is automatically selected, the recognition system is improved.

As shown in FIG. 12 (a), when the input category is 3, the result of the first major classification is category 8 of font j, and the major classification result is an error, but the correct answer is recognized by the detailed classification. The accuracy can be improved.

That is, at this time, as shown in FIG. 12B, the distance value in the large classification using the character features without feature value weighting is "8"<"3", and the output result is an error. 12
As shown in (c), in the left-right symmetry feature 2 with feature value weighting, the distance value in the detailed classification is “8”> “3”, and the output result is the correct answer.

The following table shows the result of the detailed classification dictionary creation processing (FIG. 5) in the case where the first major classification is category 8 (font j) and the second largest classification is a combination of categories 3.

[0067] The feature extraction method in which the evaluation value V is the minimum, that is, the reject and the error rate are small is the feature value weighting and the symmetry degree feature 2. The detailed classification dictionary includes: -feature extraction method (feature value weighting, left / right symmetry feature 2)
And feature amount-The threshold value function and the evaluation value V = 18.2 are stored.

Actually, for the combination of category 8 (font j) in the first major category and category 3 in the second major category, the output result of the major category confirms the contents of the input pattern that is likely to cause rejection or error. Like 12, there was a tendency for 3 to be similar to 8.

When the result that the first major category is category 8 (font j) and the second major category is category 3 is to focus on the feature of the character left center part where the difference between candidate categories 8 and 3 is large. However, it is clarified that the feature extraction method that expresses the left-right symmetry of the central part of the character is more suitable for improving the recognition accuracy by the method that considers the variation such as the shift.

When the first major classification is category 5 of font k, the input category is 6 as shown in FIG. 13 (a).
If there is (blurring or misalignment of the character center), the result of the large classification is an error, but a correct answer is obtained by the detailed classification, and the recognition accuracy can be improved.

That is, at this time, as shown in FIG. 13 (b), the distance value in the large classification using the character features without feature value weighting is "5"<"6", and the output result is an error. Thirteen
As shown in (c), diagonal symmetry feature 1 with feature value weighting
Then, the distance value in the detailed classification becomes “6”> “5”, and the output result becomes the correct answer or the reject.

The following table shows the result of the detailed classification dictionary creation process (FIG. 5) in the case where the first major classification is category 5 (font k) and the second largest classification is a combination of category 6.

[0073] The feature extracting method with the smallest evaluation value V, that is, the reject and the error rate being small, was the feature 1 with the feature value weighting. The detailed classification dictionary includes: -Feature extraction method (feature value weighting, diagonal symmetry feature 1)
And feature amount-The threshold value function and the evaluation value V = 23.2 are stored.

Actually, for the combination of category 5 (font k) in the first major category and category 6 in the second major category, when the contents of the input pattern in which rejection and error are likely to occur are confirmed in the output result of the major category, FIG. As shown in FIG. 6, the left part of 6 was faint, and the center of the character was displaced.

When the result that the first major category is category 5 (font k) and the second major category is category 6 is, the features of the left part of the character with a large difference between the candidate categories 5 and 6 (no weight,
Comparing with (partial feature), the input pattern has a result that the upper left part is similar to category 6 and the lower left part is similar to category 5, and it is impossible to determine. However, using the upper left and lower right diagonal symmetry features 1 (FIG. 13) that are automatically selected by the method of the present invention, the characteristic is that category 5 has no symmetry and category 6 has symmetry. It will be used and can be judged.

It is also possible to classify by a plurality of character feature extraction methods and use the respective output values for classification determination.

[0077]

As described above, according to the present invention, even if a recognition target or a character feature extraction method is added, recognition and evaluation can be performed using the dictionary and feature extraction method after the addition. The maintenance of the dictionary is easy, and it is possible to easily perform highly accurate recognition.

[Brief description of drawings]

FIG. 1 is an explanatory diagram showing a method of the present invention.

FIG. 2 is a flowchart of creating a large classification dictionary according to the present invention.

FIG. 3 is a diagram showing an example of input character data.

FIG. 4 shows an example of generation of a deformation pattern.

FIG. 5 is a flowchart of creating a detailed classification dictionary according to the present invention.

FIG. 6 is a diagram showing a distance value frequency distribution of a correct answer group and an error group.

FIG. 7 is an explanatory diagram of determination of a threshold function.

FIG. 8 is a diagram showing an example of determination of a threshold function.

FIG. 9 is a diagram showing a flowchart showing an example of detailed classification processing.

FIG. 10 is a diagram showing a flowchart of character feature extraction.

FIG. 11 is a diagram showing a recognition process using the device of the present invention.

FIG. 12 is a diagram showing a recognition process using the device of the present invention.

FIG. 13 is a diagram showing a recognition process using the device of the present invention.

Claims (1)

[Claims] 1. A plurality of feature extraction methods are set in the detailed classification after the major classification for a character pattern, and a classification determination threshold value is calculated for each feature extraction method to evaluate the classification accuracy. A feature extraction method having the smallest evaluation value of the classification accuracy, a dictionary creating means for storing the judgment threshold value and the evaluation value of the classification accuracy as a dictionary, and a feature suitable for classification between similar candidates in the large classification result. A pattern recognition device comprising: a pattern recognition means for performing detailed classification by reading an extraction method and a judgment threshold value from a dictionary.