JP2008071214A - Character recognition dictionary creation method and its device, character recognition method and its device, and storage medium in which program is stored - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a character recognition dictionary creation method and its device for discriminating similar characters with less computational complexity by selecting the small number of features making contribution to identification out of the original features, and a character recognition method and its device thereof, and a storage medium in which the character recognition dictionary creation program and the character recognition program are stored. <P>SOLUTION: By applying a genetic algorithm to the feature selection, and altering generation to the gene having a rate of the recognition beyond a fixed value with a degree of conformance in reduction ratio of the features, the small number of features with high recognition capability can be obtained. Moreover, the feature of the recognition dictionary is acquired by the feature number stored in the detailed recognition dictionary in the detailed recognition, so that the character recognition system with high accuracy and the small dictionary capacity can be obtained. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a character recognition dictionary creation method and apparatus, a character recognition method and apparatus, a character recognition dictionary creation program, and a storage medium storing a character recognition program.

  As a conventional character recognition method, a recognition system including preprocessing, feature extraction, and identification is often employed (see, for example, Non-Patent Document 1). The input character pattern is subjected to noise removal and normalization of the position and size of the character pattern in the preprocessing unit. The feature extraction unit extracts a feature representing the essence of the character pattern. This feature is predetermined by the type of character pattern. The identification unit prepares a standard pattern that is a typical pattern of a category to be recognized, uses a measure of the proximity between the input character pattern and the standard pattern, and recognizes the category corresponding to the closest measure as a result of recognition. The method of outputting as is known.

  In handwritten Kanji recognition, it has been proved that a high recognition rate can be obtained by using a character line direction feature as a feature and a statistical identification method as an identification method. However, even though the average recognition rate for all Kanji types is high, there are some that have low recognition rates for individual character types, many of which are seen in (question, darkness, jealousy, between) and (signature, fine). It is a similar character having a similar character shape as can be seen.

  Therefore, in order to identify similar characters, a method is adopted in which a detailed identification unit that recognizes similar characters exclusively is added to the recognition system. Detailed identification methods are roughly classified into a structural analysis method that directly handles local information of a character pattern and a statistical method that handles features extracted from the character pattern.

  As a structural analysis technique, a stroke that is a difference between similar characters is extracted and used for identification (see Non-Patent Document 2), and only a partial region such as a radical in which a difference between similar characters appears is used. And a method of identifying them (see Non-Patent Document 3). However, it is difficult to extract strokes and partial areas, which are local information, because contact and chipping occur in the strokes.

  Statistical methods are further divided into several methods, each with its own problems. The method of using only the difference feature having a large difference between the standard patterns between similar characters (see Non-Patent Document 4) has an advantage that the original feature can be used directly, but the difference feature has a large difference between similar characters. The stability is not examined, and the discriminating power is weakened when the characteristics change. Although the method using a subspace (see Non-Patent Document 5) can reduce the number of features, it only has a feature distribution of its own category, so there is no function to see the difference from other categories. The method using discriminant analysis (see Non-Patent Document 6) has a function of seeing a difference from other categories, but assumes that the feature distribution is a normal distribution. However, the distribution of the actual character pattern is often not a normal distribution, and when the number of categories to be identified is Q, only (Q-1) features can be obtained and the number of Q is small. Must be identified with a very small number of features, making identification difficult. In addition, since the subspace method and discriminant analysis use a feature obtained by performing linear transformation on the original feature as a new feature, the amount of calculation per feature increases.

"Pattern recognition", IEICE, 1988. “Handwritten Kanji Recognition by Stroke Extraction Method Using Outer Structure Information”, IEICE Transactions, (D), vol. J67-D, no. 9, pp. 1052-1059, 1984. “A Method for Detailed Identification of Handwritten Characters by Automatic Learning of Partially Matched Regions”, IEICE Transactions, D- (2), vol. J78-D- (2), no. 3, pp. 492-500, 1995. “A Study on Detailed Identification of Handwritten Kanji Using Quantification Theory”, General Conference of IEICE General Conference, no. 1620, 1984. "Handwritten similar Kanji identification by superposition method", IEICE Pattern Recognition / Understanding Technical Report, PRU 91-48, 1991. “High accuracy of handwritten character recognition using discriminant analysis”, NTT R & D, vol. 43, no. 8, pp. 873-880, 1994.

  As described above, it is difficult to extract strokes and partial areas in the structural analysis method, and the difference feature that directly uses the feature selected from the original feature in the statistical method is vulnerable to feature variation, and the number of features is reduced. However, subspace methods and discriminant analysis that identify with a small number of features are based on the assumption that the feature distribution is a normal distribution, so it is not always possible to select features that really contribute to discrimination, and the original features There is a problem that the amount of calculation per feature increases compared to the difference feature used directly.

  The present invention has been made in view of the above, and a character recognition dictionary creation method and apparatus for selecting similar characters with a small amount of calculation by selecting a small number of features that contribute to identification from among original features, and a character thereof It is an object of the present invention to provide a storage medium storing a recognition method and apparatus, a character recognition dictionary creation program, and a character recognition program.

In order to achieve the above object, the present invention uses learning data obtained by collecting a large number of character patterns per category, extracts the characteristics of the character patterns, and performs statistical processing to generate character patterns for the categories. In the dictionary creation method,
A plurality of genes consisting of bit strings in which the number of binary bits of “0” and “1” are connected for the number of features are prepared, and “0”, “ Create an initial gene population by assigning a value of 1 "
The feature of the standard pattern of the designated similar character and the feature of the character pattern of the learning data of the similar character using only the feature corresponding to the address where the bit takes a value of “1” in the individual of the initial gene population The process of calculating the discrimination scale between them and calculating the recognition rate by the individual is performed on all solids,
Create a set of only solids whose recognition rate in the solid is equal to or higher than a predetermined recognition rate as a parent gene candidate set,
For each individual of the parent gene candidate set, find the fitness set from the viewpoint of maintaining the recognition rate predetermined in the previous period with a small number of features,
Create a parent gene set by selecting a solid that will be the next generation parent gene using the fitness of each individual of the parent gene candidate set,
Two sets of solids are taken out from the parent gene set, two sets of solid genes are exchanged by crossover, and two sets of new solids are generated and returned to the parent gene set for a predetermined number of times,
Taking a solid from the parent gene set, causing a mutation that inverts the value of “0”, “1” to a part of the gene with a predetermined probability,
The above-mentioned series of recognition-fitness calculation-selection-crossover-mutation processing by an individual satisfies the predetermined convergence criteria or reaches a predetermined maximum number of generations. The generation change is repeated until the condition of whether or not is satisfied, and the solid with the maximum fitness in the finally obtained parent gene set is taken out,
Creating a detailed identification dictionary of the similar characters using the category name of the designated similar characters and the feature number corresponding to the address having a value of “1” from the object having the maximum fitness;
Is the most important feature.

  Further, the fitness is the total number of “1” that an individual has as a feature number, the maximum feature number among the objects of the generation is obtained, and the number of features that each individual can reduce from the maximum feature number is the maximum feature number. The feature number reduction ratio is defined by a value divided by a number.

In order to achieve the above-mentioned object, the present invention assigns categories to which the input character pattern belongs in ascending order of the scale using a measure of proximity obtained by calculating between the features of the input character pattern and the features of the standard pattern. In the character recognition method to output as a candidate string,
From the identification scale of the first candidate obtained by the identification and the identification scale of the second candidate, it is determined whether or not the first candidate is a correct category,
If it is determined that the reliability is high, the first candidate is output as a recognition result, and if it is not determined that the reliability is high, the candidate obtained by identification is sent as a recognition target to the subsequent detailed identification unit. And
The detailed identification unit searches the detailed identification dictionary to obtain a category set to which the first candidate belongs and a feature for detailed identification used for identification of the category set,
Read the feature specified by the feature number of the feature for detailed identification from the identification dictionary of the category set, perform the detailed identification using the feature and the feature for detailed identification in the feature obtained from the input character pattern, Outputting the obtained results as recognition results;
It is characterized by.

  The present invention has the following effects. The invention described in claims 1 and 4 is a method for performing detailed identification using a feature selected from the features used for identification, wherein the number of features is maintained while maintaining a constant recognition rate. The feature selected by the genetic algorithm is used to reduce it, so it is different from the feature selected only by the standard pattern information of similar characters like the difference feature, and the feature like the subspace method and discriminant analysis Unlike the features selected from the distribution alone, there is an advantage that a small number of features with high discrimination ability can be obtained.

  The invention described in claims 2 and 5 is the invention described in claims 1 and 4, in which the fitness is a reduction ratio of the number of features, and the value range of the fitness is large. There are advantages that it is easy to find a solution close to the optimum and that convergence is quick.

  In the invention described in claims 3 and 6, the detailed identification feature uses a part of the identification feature and the detailed identification feature is not separately set. There is an advantage in that it is not necessary to store the features, the dictionary capacity is reduced, and the recognition system configuration can be made consistent.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a block diagram of a character recognition dictionary creating apparatus according to an embodiment of the present invention. An input pattern memory unit 1, a preprocessing unit 2, a feature extracting unit 3, a feature selecting unit 4, a detailed identification dictionary creating unit 5, It consists of a detailed identification dictionary 6.

  The input pattern memory unit 1 stores character patterns captured by an input device such as a scanner or a TV camera, the preprocessing unit 2 performs normalization and noise removal, and the feature extraction unit 3 is a feature for use in recognition. Is extracted from the input character pattern, the feature selection unit 4 selects a feature necessary for detailed identification from the features, and the detailed identification dictionary creation unit 5 performs detailed identification of information obtained by adding a character code to the selected feature. Store in dictionary 6.

  Next, the operation of the feature selection unit 4 which is a main part of the present invention will be described with reference to FIG. FIG. 2 is a block configuration diagram showing an embodiment of the feature selection unit 4. An initial gene group creation circuit 41, an identification scale calculation circuit 42, a recognition rate calculation circuit 43, a parent gene candidate selection circuit 44, and a fitness calculation circuit. 45, a parent gene selection circuit 46, a crossover operation circuit 47, a mutation operation circuit 48, and a generation change continuation determination circuit 49.

  The initial gene group creation circuit 41 creates an initial state of genes. FIG. 3 is a diagram showing an initial state of a gene, and a gene solid is composed of a bit string in which d binary bits of “0” and “1” are connected in series, and a group of K pieces of these solids is gathered. Is shown. The address of each bit of the gene corresponds to the feature number. When the value of the bit is “0”, the feature corresponding to the address of the bit is not used for detailed identification, and is used when “1” is taken. Represents that The initial gene group creation circuit 41 sets the initial state by assigning the values “0” and “1” to each individual so that “1” occurs with the probability α (0 <α <1). By applying a genetic algorithm to the gene, the number of “1” is reduced, and a feature used for detailed identification is selected.

  The identification scale calculation circuit 42 corresponds to the address corresponding to an address that takes “1” in the bit string of the first solid (k = 1) among the K solids received from the initial gene population creation circuit 41 in the generation g = 1. Is used to calculate the identification scale between the feature of the standard pattern of the designated similar character and the feature of the character pattern of the learning data of the similar character. The feature of the input pattern is transmitted from the feature extraction unit 3, and the feature of the standard pattern of similar characters is read from the identification dictionary 10 and used for calculation of the identification scale. The calculation of the identification scale is performed for all character patterns of designated similar characters in the learning data. When the processing in the first solid is completed, the identification scale is calculated for all the individuals after the second individual.

The recognition rate calculation circuit 43 changes the recognition rate γ _k for the learning data of the k-th solid while K is changed to k = 1, 2,..., K using the similar character discrimination scale obtained by the discrimination scale calculation circuit 42. Calculate for all solids.

Parent gene candidate selection circuit 44 sets in generation g of only the collected solids as a predetermined recognition rate gamma ₀ or more with the recognition rate gamma _k of each individual obtained by the recognition rate calculating circuit 43 A (g ) And make these parent gene candidates.

  The fitness calculation circuit 45 calculates the fitness a (k) of the individual k belonging to the set A (g) obtained by the parent gene candidate selection circuit 44. When the feature number reduction ratio in the individual k is used as the fitness a (k)

It is represented by Here, D _max is the maximum number of features of the solids belonging to the set A (g), and D (k) is the number of features of the k-th solid. In equation (1), the feature number reduction ratio obtained by dividing the number of features that each individual can reduce from the maximum number of features by the maximum number of features is used as fitness, but the ratio of the number of individual features to the maximum number of features Any other scale may be used as long as it represents the relationship between the maximum feature number and the solid feature number.

  The parent gene selection circuit 46 causes the fitness a (k) obtained by the fitness calculation circuit 45 to act on the set A (g) obtained by the parent gene candidate selection circuit 44, and uses the fitness proportional strategy to generate the next generation parent gene. Is selected, and a parent gene set S (g) consisting of K individuals is created. Here, the fitness proportional strategy is used to create the parent gene, but an elite preservation strategy, an expected value strategy, or other methods may be used.

The crossover operation circuit 47 performs crossover on the parent gene set S (g) obtained by the parent gene selection circuit 46 to create a parent gene set S _c (g) composed of K individuals. One-point crossover is performed by taking any two individuals from S (g), cutting the two individuals with randomly generated addresses, and rearranging the bit strings after the cut point. In multipoint crossover, a plurality of crossover points are prepared, and genes are exchanged at the plurality of crossover points at a time. The crossover process may be one point, multiple points, or another method.

The mutation operation circuit 48 extracts an arbitrary individual from the parent gene set S _c (g) obtained by the crossover operation circuit 47 and performs mutation to create a parent gene set S _m (g) composed of K individuals. . Mutation is to reverse the contents of a bit string at a randomly generated address with a probability given in advance to an individual taken out.

The generational change continuation determination circuit 49 calculates a convergence measure of the genetic algorithm, and determines whether the convergence criterion is satisfied or whether g does not exceed a predetermined maximum value _gmax. If these two conditions are not satisfied, in the next generation g + 1, the parent gene set S _m (g) is replaced with S (g) and the above-described series of processing is repeated to obtain a new parent gene set S _m (g). . When any of the conditions is satisfied, an individual having the maximum fitness in the parent gene set S _m (g) is sent to the detailed identification dictionary creation unit 5. As a scale of convergence, there are a maximum value of fitness of a solid in a gene population, an average value of fitness of the whole solid in a gene population, etc., but other scales may be used.

  The gene in generation g is obtained by the above processing. Thereafter, a series of processes of recognition, fitness calculation, selection, crossover, and mutation by an individual is regarded as one generation process, and the above process is repeated until a termination condition is satisfied.

The detailed identification dictionary creation unit 5 creates a detailed identification dictionary based on the information of the parent gene set S _m (g) sent from the feature selection unit 4. FIG. 4 is a block diagram showing an embodiment of the detailed identification dictionary 6. The category set Ω _h , the target categories C _h1 , C _h2 ,..., C _{hL (h)} , the detailed identification features f _h1 , f _h2 , ..., f _{hL (h)} . The category set Ω _h (h = 1, 2,..., H) and the target categories C _h1 , C _h2 ,..., C _{hL (h)} that are elements thereof are created in a predetermined method. Feature selection unit 4 receives the parent gene set S _{m (g)} created by applying a genetic algorithm as an identification subject categories that form the elements of category set Omega _h. The address where “1” stands in the bit string of S _m (g) becomes the feature number, and the feature numbers f _h1 , f _{h2 ,.} Stored in position. The above processing is performed in order until the number of category sets H given in advance from Ω ₁ is reached, and the detailed identification dictionary 6 is completed.

2 will be described with reference to the flowchart of FIG. 5, which is an embodiment of the feature selection unit 4 which is a main part of the character recognition dictionary creation method according to the present invention. The initial gene group creation circuit 41 creates an initial gene group that is a group of K genes composed of d consecutive bits of “0” and “1”. Then, g = 1 is set as the generation number g, and k = 1 is set as the initial value for the individual number k (step 202). The identification scale calculation circuit 42 uses only the feature corresponding to the address that takes “1” in the k-th solid bit string in the generation g, and uses the standard pattern feature of the designated similar character and the learning data of the similar character. The recognition rate is calculated between the character pattern features, and the recognition rate calculation circuit 43 changes the recognition rate for the k-th solid learning data while changing k = 1, 2,... γ _k is calculated for all K solids (step 203). The parent gene candidate selection circuit 44 creates a set A (g) in the generation g in which only individuals whose recognition rate γ _{k of} each individual is equal to or higher than a predetermined recognition rate γ ₀ are collected, and these are set as parent gene candidates. (Step 204). The fitness calculation circuit 45 obtains the feature number D _max of the solid having the maximum number of features among the solids belonging to the set A (g) (step 205), and sets the feature number of the k-th solid as D (k). The fitness a (k) when the feature number reduction ratio of the solid k is defined as the fitness is calculated (step 206). The parent gene selection circuit 46 causes the fitness a (k) to act on the set A (g), selects a solid that will be the next generation parent gene by the fitness proportional strategy, and sets the parent gene set S composed of K individuals. (G) is created (step 207). The crossover operation circuit 47 performs crossover on S (g) to create a parent gene set S _c (g) composed of K individuals (step 208), and the mutation operation circuit 48 arbitrarily selects S _c (g). Are extracted and mutated to create a parent gene set S _m (g) consisting of K individuals (step 209). The generation change continuation determination circuit 49 calculates a convergence measure of the genetic algorithm (step 210), and whether the criterion of convergence is satisfied or whether g does not exceed a predetermined maximum value g _max A determination is made (step 211). If these two conditions are not satisfied, g = g + 1 is set (step 212), and S _m (g) of generation g is replaced with S (g) of generation g + 1. Processing is performed to obtain a new S _m (g). If any of the conditions is satisfied, the individual having the maximum fitness in the parent gene set S _m (g) is sent to the detailed identification dictionary creating unit 5 and the process ends (step 213).

  The gene in generation g is obtained by the above processing. Thereafter, a series of processes of recognition, fitness calculation, selection, crossover, and mutation by an individual is regarded as one generation process, and the above process is repeated until a termination condition is satisfied.

  FIG. 6 is a block diagram of a character recognition apparatus according to an embodiment of the present invention. The input pattern memory unit 1, the preprocessing unit 2, the feature extraction unit 3, the large classification unit 7, the large classification dictionary 8, the identification unit 9, It comprises an identification dictionary 10, a determination unit 11, a detailed identification unit 12, a detailed identification dictionary 6, a recognition result memory unit 13, and a control unit 14.

  The input pattern memory unit 1 captures a character pattern by an input device such as a scanner or a TV camera, the pre-processing unit 2 performs normalization, noise removal, and the like, and the feature extraction unit 3 inputs features to be used for recognition. The large classification unit 7 calculates, for each category, a measure of the closeness between the features of the input character pattern obtained from the feature extraction unit 3 and the features of the standard pattern stored in the large classification dictionary 8. The scales are arranged in ascending order and output to the identification unit 9 together with the candidates. The identification unit 9 calculates, for each category, a measure of the closeness between the features of the input character pattern and the features of the standard pattern stored in the identification dictionary 10 for the sent candidates, and the measures arranged in ascending order. Candidates corresponding to the scale are output to the determination unit 11 as identification results. The determination unit 11 determines the reliability of the result output from the identification unit 9 using a predetermined conditional expression. If the condition is satisfied, the identification result is transmitted to the recognition result memory unit 13 and stored. If the condition is not satisfied, the candidate obtained by the identification unit 9 is sent to the detailed identification unit 12. The detailed identification unit 12 calculates, for each category, a measure of the proximity between the input character pattern and the category belonging to the category set stored in the detailed identification dictionary 6 for the sent candidates, and the measures arranged in ascending order. And candidates corresponding to the scale are output to the recognition result memory unit 13 as identification results.

Next, the operation of the determination unit 11 will be described. First, the first candidate C _t and its identification scale d _t , the second candidate C _s and its identification scale d _s (where d _t ≦ d _s ) are extracted from the output from the identification unit 9. Next, determination threshold values θ _t1 and θ _t2 corresponding to C _t are read from the determination table, and when both the following expressions (2) and (3) are satisfied, it is determined that the identification result is highly reliable.
d _t ≦ θ _t1 (2)
d _s −d _t ≧ θ _t2 (3)
Equation (2) represents that the identification measure d _t is greater than the judgment threshold theta _t1 values of the first candidate C _t, Equation (3) is an identification measure d _t take the first of the candidate C _t This indicates that the difference Δd with respect to the discrimination scale d _s taken by the second candidate C _s is a value equal to or greater than the determination threshold θ _t2 . The determination threshold values θ _t1 and θ _t2 are set by analysis using a large number of learning patterns so that the misreading rate when the conditions of the expressions (2) and (3) are satisfied is not more than a predetermined value.
If it is determined that the reliability is high, the candidate string obtained by the identification and the identification scale of the candidate string are sent to the recognition result memory unit 13. If at least one of the conditions of Expression (2) and Expression (3) is not satisfied, the determination unit 11 does not determine that the reliability is high, and the candidate string obtained by the identification is transferred to the detailed identification unit 12, Candidate strings and identification scales obtained by the detailed identification unit 12 are output to the recognition result memory unit 13 as recognition results.

  Next, the operation of the detailed identification unit 12 will be described with reference to FIG. FIG. 7 is a functional block diagram of the detailed identification unit 12 showing an embodiment of the present invention, which comprises a category set readout circuit 121, a detailed identification feature readout circuit 122, a detailed identification scale calculation circuit 123, and a sort circuit 124.

Category set read circuit 121 takes out the first of the candidate C _t from the information transferred via which the determination unit 11 is sent from the identification unit 9, belongs to explore target category details identification dictionary 6 C _t Category detecting a set Omega _t, target category is an element of _{Ω t {C t1, C t2} ,. . . . , C _{tL (t)} }.

The detailed identification feature readout circuit 122 detects the detailed identification features {f _t1 , f _t2 , _... Corresponding to the Ω _t after the Ω _t is detected in the detailed identification dictionary 6. . . . , f _{tL (t)} }.

The detailed identification scale calculation circuit 123 performs the target category {C _t1 , C _t2,. . . . , C _{tL (t)} }, the detailed identification features {f _t1 , f _t2,. . . . , f _{tL (t)} }, the feature specified by the feature number is read out, and the detailed identification measure is calculated with the feature obtained from the input character pattern, and the detailed identification measure {p _t1 , p in each target category is calculated. _t2,. . . . , p _{tL (t)} }.

The sort circuit 124 includes target categories {C _t1 , C _t2,. . . . , C _{tL (t)} }, a detailed discriminant measure {p _t1 , p _t2,. . . . , p _{tL (t)} } are sorted in ascending order, and the target category and the detailed identification scale are paired and output to the recognition result memory unit 13.

  Control of each process of FIG. 6 described above is performed by a signal from the control unit 14.

The operation of FIG. 6 as an embodiment of the pattern recognition method according to the present invention will be described with reference to the flowchart of FIG. The character pattern is taken into the input pattern memory unit 1 by an input device such as a scanner or a TV camera (step 301), and preprocessing such as normalization and noise removal is performed by the preprocessing unit 2 (step 302). 3, features to be used for recognition are extracted from the input character pattern (step 303). The obtained features are sent to the major classification unit 7, where a measure of the closeness between the features of the input character pattern and the features of the standard pattern stored in the major classification dictionary 5 is calculated for each category, and the measures are in ascending order. Are output to the identification unit 9 together with the candidates (step 304). For the sent candidates, the identification unit 9 calculates a measure of the closeness between the features of the input character pattern and the features of the standard pattern stored in the identification dictionary 10 for each category, and the measures arranged in ascending order. Candidates corresponding to the scale are output to the determination unit 11 as identification results (step 305). The determination unit 11 determines the reliability of the identification unit 9 using a predetermined conditional expression (step 306). If the condition is satisfied, the identification result is transmitted to the recognition result memory unit 13 and stored (step 311), which is not satisfied. In this case, the candidates obtained by the identification unit 9 are sent to the detailed identification unit 12.
Category set read circuit 121 of the detailed identification unit 12 takes out the first of the candidate C _t from the information sent from the identification unit 9, to the genus of searching the target category that is stored in the detailed identification dictionary 6 C _t detecting a category set Omega _t (step 307), the elements of _{Ω t {C t1, C t2} ,. . . . , C _{tL (t)} } is a detailed identification target category (step 308). More discriminating feature reading circuit 122, the detailed identification features corresponding to _{Ω t {f t1, f t2} ,. . . . , f _{tL (t)} } is extracted (step 309). The detailed identification scale calculation circuit 123 performs the target category {C _t1 , C _t2,. . . . , C _{tL (t)} }, the detailed identification features {f _t1 , f _t2,. . . . , f _{tL (t)} }, the feature specified by the feature number is read out, and a detailed identification measure is calculated with the feature obtained from the input character pattern (step 310). The paired identification results are output to the recognition result memory unit 13 (step 311).

  As described above, in this embodiment, the genetic algorithm has a mechanism for reducing the characteristics used for identification by performing generational change of genes having a recognition rate of a certain value or more. Accordingly, the recognition performance can be improved as compared with methods such as differential features, principal component analysis, and discriminant analysis that do not use recognition.

  1 and 6 of the present invention is constructed as a program and stored in a portable storage medium such as a disk device, a flexible disk, or a CD-ROM, and the character recognition is performed when the character recognition is performed. By installing the program in a computer or a character recognition device to which the portable storage medium can be connected, the present invention can be easily realized.

  Although the present invention has been specifically described above based on the embodiments, it is needless to say that the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the invention.

It is a block diagram of one Example of the character recognition dictionary creation apparatus of this invention. It is a block diagram of the feature selection part 4 used for the character recognition dictionary creation apparatus shown in FIG. It is a figure which shows one Example of the gene group produced by the initial gene group production circuit 41 of FIG. It is a figure which shows one Example of the detailed identification dictionary 6 shown in FIG. It is a flowchart of one Example of the character recognition dictionary creation method of this invention. It is a block diagram of one Example of the character recognition apparatus of this invention. It is a block diagram of the detailed identification part 12 used for the character recognition dictionary creation apparatus shown in FIG. It is a flowchart of one Example of the character recognition method of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Input pattern memory part 2 Preprocessing part 3 Feature extraction part 4 Feature selection part 5 Detailed identification dictionary creation part 6 Detailed identification dictionary 7 Major classification part 8 Major classification dictionary 9 Identification part 10 Identification dictionary 11 Determination part 12 Detailed identification part 13 Recognition Result memory unit 14 Control unit 41 Initial gene group creation circuit 42 Identification scale calculation circuit 43 Recognition rate calculation circuit 44 Parent gene candidate selection circuit 45 Fitness calculation circuit 46 Parent gene selection circuit 47 Crossover operation circuit 48 Mutation operation circuit 49 Generation change Continuation determination circuit 121 Category set readout circuit 122 Detailed identification feature readout circuit 123 Detailed identification scale calculation circuit 124 Sort circuit

Claims (8)

In a character recognition dictionary creating method for creating a standard pattern of the category by using a learning data obtained by collecting a large number of character patterns per category, extracting characteristics of the character pattern and performing statistical processing,
A plurality of genes consisting of bit strings in which the number of binary bits of “0” and “1” are connected for the number of features are prepared, and “0”, “ Create an initial gene population by assigning a value of 1 "
The feature of the standard pattern of the designated similar character and the feature of the character pattern of the learning data of the similar character using only the feature corresponding to the address where the bit takes a value of “1” in the individual of the initial gene population The process of calculating the discrimination scale between them and calculating the recognition rate by the individual is performed on all solids,
Create a set of only solids whose recognition rate in the solid is equal to or higher than a predetermined recognition rate as a parent gene candidate set,
For each individual of the parent gene candidate set, find the fitness set from the viewpoint of maintaining the recognition rate predetermined in the previous period with a small number of features,
Create a parent gene set by selecting a solid that will be the next generation parent gene using the fitness of each individual of the parent gene candidate set,
Two sets of solids are taken out from the parent gene set, two sets of solid genes are exchanged by crossover, and two sets of new solids are generated and returned to the parent gene set for a predetermined number of times,
Taking a solid from the parent gene set, causing a mutation that inverts the value of “0”, “1” to a part of the gene with a predetermined probability,
The above-mentioned series of recognition-fitness calculation-selection-crossover-mutation processing by an individual satisfies the predetermined convergence criteria or reaches a predetermined maximum number of generations. The generation change is repeated until the condition of whether or not is satisfied, and the solid with the maximum fitness in the finally obtained parent gene set is taken out,
Creating a detailed identification dictionary of the similar characters using the category name of the designated similar characters and the feature number corresponding to the address having a value of “1” from the object having the maximum fitness;
Character recognition dictionary creation method characterized by   The fitness is the total number of “1” s that an individual has, and the maximum number of features is obtained among the solids of the generation, and the number of features that each solid can reduce from the maximum number of features is the maximum number of features. 2. The character recognition dictionary creation method according to claim 1, wherein the feature number reduction ratio is defined by the divided value. In the character recognition method of outputting the category to which the input character pattern belongs as a candidate string in ascending order of the scale using a measure of proximity obtained by calculating between the features of the input character pattern and the features of the standard dictionary,
From the identification scale of the first candidate obtained by the identification and the identification scale of the second candidate, it is determined whether or not the first candidate is a correct category,
If it is determined that the reliability is high, the first candidate is output as a recognition result, and if it is not determined that the reliability is high, the candidate obtained by identification is sent as a recognition target to the subsequent detailed identification unit. And
The detailed identification unit searches the detailed identification dictionary to obtain a category set to which the first candidate belongs and a feature for detailed identification used for identification of the category set,
Read the feature specified by the feature number of the feature for detailed identification from the identification dictionary of the category set, perform the detailed identification using the feature and the feature for detailed identification in the feature obtained from the input character pattern, Outputting the obtained results as recognition results;
Character recognition method characterized by In a character recognition dictionary creation device that creates a standard pattern of the category by using a learning data collected a large number of character patterns per category, extracting features of the character pattern and performing statistical processing,
A plurality of genes consisting of bit strings in which the number of binary bits of “0” and “1” are connected for the number of features are prepared, and “0”, “ Create an initial gene population by assigning a value of 1 "
The feature of the standard pattern of the designated similar character and the feature of the character pattern of the learning data of the similar character using only the feature corresponding to the address where the bit takes a value of “1” in the individual of the initial gene population The process of calculating the discrimination scale between them and calculating the recognition rate by the individual is performed on all solids,
Create a set of only solids whose recognition rate in the solid is equal to or higher than a predetermined recognition rate as a parent gene candidate set,
For each individual of the parent gene candidate set, find the fitness set from the viewpoint of maintaining the recognition rate predetermined in the previous period with a small number of features,
Create a parent gene set by selecting a solid that will be the next generation parent gene using the fitness of each individual of the parent gene candidate set,
Two sets of solids are taken out from the parent gene set, two sets of solid genes are exchanged by crossover, and two sets of new solids are generated and returned to the parent gene set for a predetermined number of times,
Taking a solid from the parent gene set, causing a mutation that inverts the value of “0”, “1” to a part of the gene with a predetermined probability,
The above-mentioned series of recognition-fitness calculation-selection-crossover-mutation processing by an individual satisfies the predetermined convergence criteria or reaches a predetermined maximum number of generations. A detailed identification dictionary creating means for repeating generation change until the condition of whether or not is satisfied, and extracting a solid with the maximum fitness in the finally obtained parent gene set,
Detailed identification dictionary means for creating a detailed identification dictionary of the similar character by using the category name of the designated similar character and the feature number corresponding to the address having a value of “1” from the object having the maximum fitness ,
A character recognition dictionary creating apparatus comprising:   The fitness is the total number of “1” s that an individual has, and the maximum number of features is obtained among the solids of the generation, and the number of features that each solid can reduce from the maximum number of features is the maximum number of features. The character recognition dictionary creation device according to claim 4, wherein a feature number reduction ratio defined by the divided value is used. In a character recognition device that outputs a category to which the input character pattern belongs as a candidate string in ascending order of the scale using a measure of proximity obtained by calculating between the features of the input character pattern and the features of the standard pattern,
From the identification scale of the first candidate obtained by the identification and the identification scale of the second candidate, it is determined whether or not the first candidate is a correct category,
If it is determined that the reliability is high, the first candidate is output as a recognition result, and if it is not determined that the reliability is high, the candidate obtained by identification is sent as a recognition target to the subsequent detailed identification unit. Determination means to perform,
Search the detailed identification dictionary to obtain the category set to which the first candidate belongs and the characteristics used for detailed identification of the category set,
Read the feature specified by the feature number of the feature for detailed identification from the identification dictionary of the category set, perform the detailed identification using the feature and the feature for detailed identification in the feature obtained from the input character pattern, Detailed identification means for outputting the obtained results as recognition results;
A character recognition device comprising:   A storage medium storing a program for causing a computer to execute the processing steps in the character recognition dictionary creating method according to claim 1.   5. A storage medium characterized in that a program for causing a computer to execute the processing steps in the character recognition method according to claim 4 is stored in a computer-readable medium.

Images