JPH0773276A - Character recognition device - Google Patents

Abstract

PURPOSE:To provide a character recognition device which can shorten the recognition time and can assure the highly accurate recognition performance. CONSTITUTION:The different feature vectors are acquired from the character patterns respectively. Plural single feature recognizing parts 13 recognize individually the feature vectors. A candidate character selecting part 14 selects plural candidate character categories to the character patterns based on all acquired resemblance, and a recognition post-processing part 15 decides a recognizing character category. Under such conditions, a recognizing part 13 calculates the reversion degree of each category group to the character pattern through a rough sorting part and calculates the intra- group resemblance of each character category through plural fine sorting parts. Then the part 13 calculates the resemblance of each character category to the character pattern based on the group reversion degree and the intra-group resemblance. At the part 14, a converting part 15 converts the resemblance of each character category received from each part 13 into the integrated resemblance. Then a 1st selecting part 16 selects plural candidate characters based on the integrated resemblance, the size information on the character pattern, etc., and a 2nd selecting part 17 calculates the final resemblance of each character category based on the integrated resemblance of selected candidate characters to select the final candidate character category.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device for recognizing patterns of handwritten and printed characters on a document, and in particular, character patterns are roughly classified and then finely classified. The present invention relates to an apparatus for recognizing a character pattern by performing such hierarchical recognition.

[0002]

2. Description of the Related Art Conventionally, an input character pattern is first roughly classified, and a character category group to which the input character pattern belongs (here, the character category group is a character having a similar feature vector to the character pattern). , Which is a set of patterns), and an example of a character recognition device for recognizing a character pattern by finely classifying the selected character category group is, for example, IEICE Transactions D. -II Vol.J75-D-II No.3 pp545-553 "Large-scale neural network" CombNET-II "".

FIG. 12 shows a block diagram of this conventional character recognition apparatus. Reference numeral 110 denotes an image input unit such as an image scanner for inputting a document image to be recognized. Reference numeral 111 denotes a character cutout unit, which cuts out a character pattern for each character area from the document image input by the image input unit 110. A feature extraction unit 112 extracts a feature vector used to identify a character category from the character pattern cut out by the character cutout unit 110. Reference numeral 113 denotes a large classification unit, which uses the feature vector extracted by the feature extraction unit 112 to roughly classify the character patterns into respective character category groups. A sub-classification unit 114 sub-classifies the character pattern in each character category group using the feature vector. A group selection unit 115 selects a plurality of character category groups from the output value (hereinafter referred to as the goodness of fit) of the large classification unit 113. 116 is a subclassification unit input signal selection unit, which is a group selection unit 115
The sub-classification unit 114 for inputting the feature vector is selected on the basis of the group selection information obtained in. Reference numeral 117 denotes an identification unit, which identifies the character pattern based on the goodness of fit of the character category group selected by the group selection unit 115 and the output value of the fine classification unit 114.

In the general classification unit 113, 118 is an input unit for inputting the feature vector of the character pattern extracted by the feature extraction unit 112. A multi-input / single-output signal processing unit 119 calculates the degree of conformity of each character category group to the character pattern. In the subclassification unit 114, 120 is an input unit for inputting the feature vector output from the subclassification unit input signal selection unit 116. 121
Is a multi-input / single-output signal processing unit, which is connected to the lower input unit 120 or multi-input / single-output signal processing unit 121.
Is multiplied by a weighting factor, which is the degree of connection, and the sum is summed to output a threshold value. Here, the plurality of multi-input one-output signal processing units 121 are formed into a layered structure, and are connected to the network so that signals are propagated only to an upper layer without mutual coupling in each layer, so that the characters corresponding to the character pattern are formed. The degree of similarity to each character category in the category group is obtained. 12
Reference numeral 2 denotes a maximum value selection unit that selects the maximum value from the output values of the plurality of multi-input / output signal processing units in the uppermost layer.

In the identification unit 117, 123 is a similarity calculation unit, and each character is selected from the conformity of the character category group selected by the group selection unit 115 and the output value of the subclassification unit 114 corresponding to the character category group. This is to calculate the similarity of categories. A category identifying unit 124 identifies the character category of the input character pattern by obtaining the maximum value of the similarity of each character category obtained from the similarity calculating unit 123.

The operation of the conventional character recognition device configured as described above will be described below.

From the document image input by the image input unit 110, the character cutting unit 111 extracts character patterns one by one. The feature extraction unit 112 uses the feature vector X consisting of n feature data for the character pattern cut out by the character cutout unit 111.

[0008]

## EQU1 ## X = (x ₁ , x ₂ , ..., X _n ) is calculated. Here, the feature data is obtained by the density mesh method. In the light and shade mesh method, the input character pattern is divided into n small areas, and the area of the character part (the number of black pixels contained in the small area) is normalized by the area of the small area for each small area. Is the characteristic data.

The feature vector X thus extracted by the feature extraction unit 112 is then input to the input unit 118 of the large classification unit 113. There are n input units 118, which are equal in number to the feature data of the character pattern, and each feature data x _i is input to the corresponding input unit 118. Each multi-input one-output signal processing unit 119 in the general classification unit 113 has an input x _j of the input unit 118 connected thereto and a weighting factor v that is the degree of connection thereof.
_ij (1 ≦ i ≦ m _r ; m _r is the number of character category groups, 1 ≦ j ≦
n) and the sum of those multiplied by n are calculated, and then this is calculated as a feature vector X and a weighting coefficient vector V _{i of} each multi-input one-output signal processing unit 119.

[0010]

## EQU2 ## The output is obtained by dividing the norm | X | and | V _i | of V _i = (v _i1 , v _i2 , ..., V _in ). That is, the output value sim multiple input first output signal processing section 119 with the weighting coefficient vector V _i shown in FIG. 12 (X, V _i)
Can be expressed as in equation (3).

[0011]

[Equation 3]

As for the weighting coefficient vector V _i ,
The multi-input / single-output signal processing unit 119 is designed in advance so as to generate the maximum output for a set of character patterns having similar feature vectors X.

According to the conventional example, these weighting coefficient vectors V _i are designed by the following method. First, in the first step, to each input of the feature vector X of a character pattern for weighting coefficient vector design, most sim (X, V _i) determined a large V _C (in this case, X is optimally matched to the V _C ,), And bring V _C closer to X. Further, when the number of character patterns optimally matching one weighting coefficient vector exceeds a certain number, the area covered by the weighting coefficient vector is divided into two, and another weighting coefficient vector is generated.
In the second step, the optimum matching V _i is obtained for all character patterns for designing the weighting coefficient vector, and it is checked whether or not it has changed from the previous time. Then, if there is a change, the V _i is corrected. At this time, the weighting coefficient vector is also generated as in the first step. The above operation is repeated until the weight coefficient vector is no longer modified or generated.

By designing the weight coefficient vector in this manner, each weight coefficient vector V _i can be divided and quantized in the feature vector space of the character pattern. That is, the input character pattern is classified into a set of a plurality of character patterns having similar feature vectors, that is, a plurality of character category groups by each weighting coefficient vector V _i . And each multi-input one-output signal processing unit
The output value of 119 is output to the group selection unit 115 as the suitability of each character category group for the character pattern.

The group selection unit 115 selects an arbitrary number of character category groups in descending order of goodness of fit obtained by the large classification unit 113, and selects grouped character information indicating which character category group has been selected and the corresponding goodness of fit. Is output.

Based on the group selection information obtained from the group selection unit 115, the subclassification input signal selection unit 116 selects the subclassification unit 114 for inputting the feature vector X of the input character pattern, and sets X as these. It outputs to the subclassification unit 114.

In each subclassification section 114 corresponding to the character category group selected by the group selection section 115 (that is, the subclassification section 114 to which the feature vector X of the character pattern is input from the subclassification input signal selection section 116), First, the feature vector X is input to the input unit 120. There are n input units 120, which are equal in number to the feature data of the character patterns, and each feature data x _i is input to the corresponding input unit 120. Each multi-input / single-output signal processing unit 121 of the subclassification unit 114 multiplies the output of the lower-layer input unit 120 or the multi-input / single output signal processing unit 121 connected to the weighting coefficient, which is the degree of the connection. After converting the sum of all things with a threshold function, the value is output to the upper layer. Here, the highest-level multi-input / single-output signal processing section 121 of each subclassification section 114 is set to the same number as the number of character categories of the character pattern included in each character category group. Input / output signal processing unit 121
Corresponds to each of these character categories. The maximum value selection unit 122 selects the maximum output value among the output values of each multi-input / output signal processing unit 121 in the uppermost layer, the character category corresponding to this multi-input / output signal processing unit 121, and its maximum output. Output the value.

The weighting factor of each multi-input one-output signal processing section 121 is such that for the feature vector X of a character pattern having each character category in the character category group, the weighting coefficient of the highest layer corresponding to each character category is set. Multi-input / single-output signal processing unit 121
Are pre-learned to generate maximum output.

Specifically, such a weighting coefficient learning method is performed by a learning algorithm called an error back propagation method. The error backpropagation method is described, for example, by DE Rumelhart, GE Hinton and RJ Williams in "Learning Reproducible by Learning Learning Representations by Back".
-Propagating Errors), "Nature, vol.323, pp.5
33-536, Oct. 9, 1986.

The outline of the error back propagation method will be described below.

First, the feature vector X of the character pattern for weighting coefficient learning is input to the input unit 120 of the fine classification unit 114.
As described above, each multi-input one-output signal processing unit 121 has a lower-layer input unit 120 connected thereto, or an output of the multi-input one-output signal processing unit 121 and a weighting factor which is the degree of connection thereof. After the total sum of those multiplied by is converted by the threshold function, the value is output to the upper layer. here,
Outputs o _{k of} all the top-level multi-input one-output signal processing units 121
The error E between the desired output t _k (referred to as a teacher signal) and the desired output t _k is obtained by the equation (4).

[0022]

## EQU4 ## E = 0.5 Σ _p Σ _k (t _k −o _k ) ² where Σ _p is the sum of the number of character patterns of the teacher signal. The purpose of learning is to determine the value of the weighting coefficient that minimizes the error E.
The change amount Δw _ij of the weight coefficient 121 is calculated based on the equation (5).

[0023]

[Expression 5] Δw _ij = −ε∂E / ∂w _ij where ε is a positive constant called a learning rate. The error E can be reduced by repeating such updating of the weighting coefficient based on (Equation 5) each time the feature vector X of the learning character pattern is input. When the error E becomes sufficiently small, the learning is terminated assuming that the output signal has become sufficiently close to the desired value.

By such a weighting coefficient learning method,
For a character pattern having each character category in the character category group, the highest input multi-input single-output signal processing unit 121 corresponding to each character category can generate the maximum output. Therefore, among the plurality of multi-input one-output signal processing units 121 in the uppermost layer, the one that produces the maximum output is selected by the maximum value selection unit 122, so that each character category group, that is, each sub-classification is performed. It is possible to identify the character category of the character pattern input in the section. In the identification unit 117, first, in the similarity calculation unit 123, from the suitability of the character category group selected by the group selection unit 115 and the output value of the subclassification unit 114 corresponding to the character category group, the expression (6) is obtained. Is used to calculate the similarity of each character category obtained by the subclassification unit 114, and the similarity is calculated by the category identification unit 12
Output to 4.

[0025]

(Similarity) = (Fitness) ^a (Output value) ^b where a and b are real constants.

Finally, the category identifying unit 124 compares the similarities of the character categories obtained from the similarity calculating unit 123 and outputs the character category corresponding to the maximum similarity among them as the identification result. .

[0027]

In character recognition, a feature that is effective in identifying a character pattern is extracted as a feature vector, but it is generally difficult to realize sufficient recognition ability with only a single feature vector. By using a plurality of types of feature vectors, a high identification function can be realized. That is, by using a plurality of types of features, even a character pattern that can only be ambiguous or misidentified by a certain feature vector may be accurately identified by different feature vectors. At this time, accurate identification can be performed. This is because there is a possibility that the identification performance can be improved by placing more importance on this.

However, in the character recognition device for hierarchically recognizing a character pattern having a large number of character categories, as in the conventional example, in a structure having only one feature extraction section and recognition section, When a plurality of types of feature vectors are used, they must be input together and recognized. In this case, compared to the case of using only a single feature vector, it is possible to achieve a certain degree of high discrimination performance, but the advantages obtained by using different feature vectors as described above cannot be fully utilized effectively. There is a problem. In other words, when multiple types of feature vectors are used together, the recognition performance can be improved, but the recognition result is used to identify characters that are erroneously recognized by a certain feature vector as described above. It is very difficult to correctly recognize a pattern based on the identification result of different feature vectors.

Further, when a plurality of types of feature vectors are used together, there is a problem in that recognition requires a lot of calculation time as the number of dimensions of the feature vector increases.

In view of the problems of the conventional character recognition device, the present invention recognizes with high accuracy by sufficiently effectively utilizing the advantage of using a plurality of types of feature vectors when using a plurality of types of feature vectors in the recognition device. It is an object of the present invention to provide a character recognition device capable of realizing the above and further shortening the time required for recognition.

[0031]

According to a first aspect of the present invention, an image input section for inputting a document image, and a character for cutting out a character from the document image input by the image input section for each character area to obtain a character pattern. A cutout unit, a plurality of feature extraction units that obtain different feature vectors from the character patterns, and a plurality of single feature recognition units that obtain the degree of similarity of the character patterns belonging to each character category from each feature vector,
And a character selection recognition post-processing unit that integrates the similarity of each character category obtained from a plurality of single feature recognition units for each same character category and determines the recognized character category using the result, The single feature recognizing unit stores a group dictionary in which a plurality of group reference feature vectors representing a character category group consisting of a set of character patterns having similar feature vectors are stored, a group reference feature vector, and a feature of a character pattern. A fuzzy major classification unit that calculates the degree of group membership, which is the degree to which a character pattern belongs to each character category group using a vector, and a character pattern for each character category included in the character category group, using the feature vector of the character pattern. A plurality of sub-classification units that obtain the degree of belonging within a group, a group selection unit that selects a plurality of character category groups from the group membership degree, and a group selection unit that is obtained from the group selection unit. A subclassification unit that selects the subclassification unit that inputs the feature vector of the character pattern based on the information Input signal selection unit, the group membership degree of the character category group selected by the group selection unit, and the group obtained by the subclassification unit The single feature similarity calculator calculates the similarity of each character category to the character pattern using the internal similarity, and the single feature similarity calculator calculates the group membership of the character category group selected by the group selector. And a plurality of multipliers for multiplying all the in-group similarities obtained from the fine classification unit, which are input with the feature vector of the character pattern from the input signal selection unit, and the output value of the multiplier for each character category Is a character recognition device including a category similarity calculation unit that selects a plurality of large output values and calculates the sum of these output values.

According to the second aspect of the present invention, an image input unit for inputting a document image, and a character pattern and a character are obtained by cutting out a character for each character region from the document image input by the image input unit using a plurality of parameter values. A character segmenting unit that obtains pattern size information and position information, a plurality of feature extraction units that obtain different feature vectors from the character patterns, and a plurality of feature vectors that determine the degree of similarity that character patterns belong to each character category. A single feature recognition unit, a candidate character selection unit that obtains a plurality of secondary candidate character categories for a character pattern using the similarity of each character category obtained from the plurality of single feature recognition units, and a character pattern A post-recognition processing unit that determines a recognition character category from a plurality of secondary candidate character categories using size information and position information is provided, and The recognition unit uses a group dictionary in which a plurality of group reference feature vectors representing a character category group composed of a set of character patterns having similar feature vectors are stored, and a group reference feature vector and a character pattern feature vector. A fuzzy major classification part that calculates the degree of group membership, which is the degree to which a character pattern belongs to each character category group, and the degree to which the character pattern belongs to each character category included in the character category group using the feature vector of the character pattern. Input a character pattern feature vector based on the group selection information obtained from the group selection unit that selects a plurality of character category groups from the group membership degree Select a subclassification section to be selected Subclassification section Input signal selection section, group membership degree of character category group selected in group selection section and similarity within group obtained in subclassification section Equipped with a single feature similarity calculator that finds the similarity of each character category to a character pattern, the single feature similarity calculator calculates the group membership and subclassification of the character category group selected by the group selector. Multiple multipliers for multiplying all the intra-group similarities obtained from the sub-classification unit, which are input with the feature vector of the character pattern from the input signal selection unit, and the output value of the multiplier for each character category is large And a category similarity calculation unit that calculates the sum of these output values.The candidate character selection unit determines the similarity of each character category obtained from the single feature recognition unit corresponding to each feature vector. A plurality of integrated similarity conversion units that convert to the integrated similarity by normalizing with the maximum value among those similarities, and using the integrated similarity and the size information and position information of the character pattern of all character categories During ~ By adding a plurality of primary candidate character selection units for selecting a plurality of primary candidate character categories from each of the selected primary candidate character categories and the integrated similarities of the selected primary candidate character categories for each same category. The character recognition device is provided with a secondary candidate character selection unit that obtains the final similarity of the primary candidate character category and selects a plurality of those having a large final similarity value as the secondary candidate character category.

[0033]

According to the present invention, according to the configuration, first, the character cutout unit extracts a character pattern for each character region from the document image input by the image input unit, and the character pattern is extracted by each feature extraction unit and also by the character pattern. The size information and the position information are output to the primary candidate character selection unit and the post-recognition processing unit in the candidate character selection unit. Each feature extraction unit extracts a different feature vector from the character pattern cut out for each character region and outputs it to a plurality of corresponding single feature recognition units.

In each single feature recognition unit, first, the feature vector is input to the fuzzy major classification unit. The fuzzy major classification unit reads all the group-referenced feature vectors stored in the group dictionary, and performs major classification on the feature vector by ambiguously defining the boundaries of each character category group in the feature vector space, A group membership degree, which is the degree to which a character pattern belongs to each character category group, is obtained. Using these group membership degrees, the group selection unit selects a plurality of character category groups, and outputs the group selection information to the subclassification unit input signal selection unit and the corresponding group membership degree to the single feature similarity calculation unit. To do. The fine classification unit input signal selection unit outputs the feature vector to the fine classification unit corresponding to the group selection information, and each fine classification unit uses the feature vector to generate a character pattern for each character included in each character category group. The within-group similarity, which is the degree of similarity to the category, is calculated and output to the single feature similarity calculation unit. The single feature similarity calculation unit calculates the group membership of the character category group selected by the group selection unit and all in-group similarity obtained from the subclassification unit to which the feature vector is input from the subclassification unit input signal selection unit. Is multiplied by a multiplier, and this is output to the category similarity calculation unit. In the category similarity calculation unit, a plurality of output values of the multiplier are selected for each character category, and the sum of these output values is calculated.
This is the similarity of each character category to the character pattern. Thus, in each single feature recognition unit, from each feature vector input from each corresponding feature extraction unit, the similarity of each character category with respect to the input character pattern is obtained, and this is sent to the candidate character selection unit. Output.

In the candidate character selection unit, each integrated similarity conversion unit first normalizes the similarity of each category obtained from the single feature recognition unit corresponding to each feature vector by the maximum value of those similarities. It is converted into an integrated similarity by converting it to Each of the primary candidate character selection units uses the integrated similarity obtained from the corresponding integrated similarity conversion unit and the size information and position information of the character pattern obtained from the character cutout unit to select a plurality of character categories from all character categories. The candidate character category of is selected as the primary candidate character category. The secondary candidate character selection unit adds up the integrated similarities of the primary candidate character categories obtained from each of the primary candidate character selection units for each same category, so that all the primary candidate character categories are added. After the final similarity is calculated, a plurality of ones having a large final similarity value are selected and the second similarity is selected.
The next candidate character category is output to the post-recognition processing unit. As described above, the candidate character selection unit uses the similarity of each character category obtained from the plurality of single feature recognition units and the size information and position information of the character pattern obtained from the character cutout unit to obtain a plurality of candidate characters. A character category is selected and output to the post-recognition processing unit.

Finally, the post-recognition processing unit uses the size information and the position information of the character pattern obtained from the character cutting unit to finally select from the plurality of secondary candidate character categories obtained from the candidate character selection unit. Character categories.

As described above, the present character recognition apparatus performs the discrimination processing of a plurality of types of feature vectors by the separate single feature recognition units to obtain the similarity of each character category to each feature vector, and then the candidate characters. The selection unit integrates the identification results obtained from the single feature recognition units using all of these similarities, and recognizes the input character pattern. Therefore,
Even a character pattern that is erroneously recognized by a certain feature vector can be correctly recognized by the identification result by a different feature vector, and highly accurate recognition can be realized. Further, in the conventional method, the time required for recognition is increased by using a plurality of types of feature vectors, but in the present character recognition device, since each single feature recognition unit performs the identification processing in parallel, the time required for recognition is increased. Can be shortened as compared with the conventional example.

[0038]

Embodiments of the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 1 is a block diagram of a character recognition apparatus according to a first embodiment of the present invention.

In FIG. 1, reference numeral 10 denotes an image input unit such as an image scanner for inputting a document image to be recognized. Reference numeral 11 denotes a character cutout unit, which cuts out character patterns from the document image input by the image input unit 10 for each character area using a plurality of parameter values, and extracts size information and position information of the character patterns. A feature extracting unit 12 extracts a feature vector used for identifying a character category from the character pattern cut out by the character cutting unit 11. However, each of the plurality of feature extraction units 12 extracts a different feature vector. Reference numeral 13 denotes a single feature recognition unit, which is prepared in the same number as that of the feature extraction unit 12, and which inputs the corresponding feature vectors to obtain the similarity, which is the degree to which the character pattern belongs to each character category. Reference numeral 14 denotes a candidate character selection unit, which obtains a plurality of candidate character categories for a character pattern by using the similarity of each character category for each feature vector of the character pattern obtained from each single feature recognition unit 13. . A post-recognition processing unit 15 uses the size information and the position information of the character pattern obtained from the character cutting unit,
The final character category is determined from a plurality of candidate character categories obtained from the candidate character selection unit 14.

In the candidate character selection unit 14, 16 is an integrated similarity conversion unit, the same number as the single feature recognition unit 13 is prepared, and each character category obtained from the corresponding single feature recognition unit 13 Is converted into an integrated similarity. Reference numeral 17 denotes a primary candidate character selection unit, which is provided in the same number as the single feature recognition unit 13 like the integrated similarity conversion unit 16, and the integrated similarity obtained from the corresponding integrated similarity conversion unit 16 respectively. And character pattern size information and position information obtained from the character cutout unit 11, and standard size information and standard position information of each character category stored in the character information storage unit 50, which will be described later, are used for all character categories. A plurality of candidate character categories (this is called a primary candidate character category) is selected from among them. Reference numeral 18 denotes a secondary candidate character selection unit, which is a first candidate character obtained from each primary candidate character selection unit 17.
Using the integrated similarity of the next candidate character category, all first
A plurality of candidate categories (which are called secondary candidate categories) are selected from the next candidate character categories and output to the post-recognition processing unit 15. Further, in FIG. 1, the configuration of the secondary candidate character selection unit 18 is specifically shown, where 181 is an adder, and a plurality of primary candidate character selection units 17 are provided.
The final similarity of each character category is obtained by adding the integrated similarity of the primary candidate character categories obtained from the above for each same category. A final candidate character selection unit 182 selects a plurality of secondary candidate character categories from the final similarity of each character category.

In the post-recognition processing unit 15, a character information storage unit 50 stores standard size information and standard position information of character patterns for each character category.
A post-processing target character set storage unit 51 stores a combination of character categories that require post-recognition processing (post-processing target character category set). 52 is a post-processing target character set determination unit, which is the second character obtained from the candidate character selection unit 14.
When the first candidate character category of the next candidate character category is the post-processing target character category set, the character information storage unit 50
Using the standard size information of each character category corresponding to the post-processing target character category set stored in, the standard position information, the size information of the input character pattern, and the position information,
Any one of the character categories in the post-processing target character category set is determined as a recognized character category.

FIG. 2 is a block diagram specifically showing the configuration of the single feature recognition unit 13. In FIG. 2, reference numeral 20 denotes a group dictionary in which a plurality of group reference characteristic vectors representing a character category group composed of a set of character patterns having similar characteristic vectors are stored. Reference numeral 21 is a fuzzy major classification unit, which uses the group reference feature vector stored in the group dictionary 20 and the feature vector obtained from the feature extraction unit 12 to extract the character pattern extracted from the character extraction unit 11 into each character category. The degree of group membership, which is the degree of belonging to a group, is calculated. A sub-classification unit 22 calculates the intra-group similarity, which is the degree to which the character pattern belongs to each character category included in the character category group, using the feature vector. A group selection unit 23 selects a plurality of character category groups from the degree of group membership. A subclassification unit input signal selection unit 24 selects the subclassification unit 22 to which the feature vector of the character pattern is input, based on the group selection information obtained from the group selection unit 23. Reference numeral 25 denotes a single feature similarity calculation unit, which determines the similarity of each character category to the character pattern from the group membership of the character category group selected by the group selection unit 23 and the intra-group similarity obtained by the subclassification unit 22. It is to seek the degree.

In the single feature similarity calculation unit 25, 26 is a multiplier, and the character belonging to the character category group selected by the group selection unit 23 and the feature of the character pattern from the subclassification unit input signal selection unit 24. The vector is multiplied by the in-group similarity obtained from the sub-classification unit 22. A category similarity calculation unit 27 selects a plurality of output values of the multiplier 26 for each character category and calculates the sum of these output values.

FIG. 3 is a block diagram specifically showing the configuration of the fuzzy major classification unit 21. In FIG. 3, reference numeral 30 is an input unit for inputting the feature vector obtained from the feature extraction unit 12. Reference numeral 31 denotes a distance calculation unit, which calculates the distance between all the group reference feature vectors of the group dictionary 20 and the feature vectors. Reference numeral 32 is a divider that calculates the reciprocal of the output of the distance calculation unit 31. Reference numeral 33 denotes an adder, which calculates the total sum of the outputs of the respective dividers 32. 34 is a multiplier, and the output of the adder 33 and the distance calculation unit 31
Is to multiply the output of. A divider 35 calculates the reciprocal of the output of the multiplier 34.

FIG. 4 is a block diagram specifically showing the configuration of the first embodiment of the subclassification section 22. In FIG. 4, reference numeral 40 denotes an input unit for inputting the feature vector of the character pattern output from the subclassification unit input signal selection unit 24.
Reference numeral 41 denotes a character category dictionary, which stores a plurality of character category reference feature vectors indicating a representative value of each character category of a character pattern. Reference numeral 42 is a distance calculation unit that calculates the distances between all the character category reference feature vectors of the character category dictionary 41 and the feature vectors. A divider 43 calculates the reciprocal of the output of the distance calculator 42. Reference numeral 44 is an adder, which calculates the total sum of the outputs of the respective dividers 43. A multiplier 45 multiplies the output of the adder 44 and the output of the distance calculation unit 42. 46 is a divider, which calculates the reciprocal of the output of the multiplier 45. The operation of the character recognition device configured as described above will be described below.

From the document image input by the image input unit 10,
As shown in FIG. 5, the character segmentation unit 11 extracts a character pattern for each character region using a plurality of parameters, and the character pattern is extracted to each feature extraction unit 12, and the size information of the extracted character pattern (character Pattern width a, height b),
The position information h is output to the primary candidate character selection unit 17 and the post-recognition processing unit 15 in the candidate character selection unit 14. Here, specific examples of the parameter in the character cutout unit 11 include a standard size of a character area, a pitch, a space length of an adjacent character area, and the like.For example, in a Japanese document, the standard size of the character area is a full-width Kanji character. Set to font size.

In each of the plurality of feature extraction units 12 prepared, a different feature vector X _k (k = 1 to N _F : N _F is the number of types of feature vectors) from the character pattern extracted from the character extraction unit 11. To extract. Here, each feature vector X _k is composed of n _k feature data,

[0048]

## _EQU7 ## This is expressed as X _k = (x _k1 , x _k2 , ..., X _knk ).

The following are examples of feature vectors of a plurality of different character patterns.
Feature vectors other than the following may be used. (A) The image information of the character pattern already described in the conventional example is divided into, for example, 8 × 8, and the 64-dimensional grayscale feature mesh feature having the frequency of black pixels in each area as an element (b) The image information of the character pattern For example, divide into 4x4,
An element is the appearance frequency of the direction elements (horizontal, vertical, diagonally right, diagonally left) of the character contour points in each small area.
64-dimensional contour direction density feature (c) Image information of a character pattern is divided into, for example, 4 × 4,
The area of a region that is not the character part from the one circumscribing frame to the first circumscribing frame until the first character part is encountered is counted. This operation is performed on four circumscribing frames, and 16-dimensional primary peripheral features are obtained. Similarly, the first encounter of the character part continues counting, and the area until the second character part is encountered is counted. 32 dimensional peripheral features combined with 16 dimensional secondary peripheral features obtained by Each feature vector X _k of the character pattern extracted by each feature extraction unit 12 is input to the corresponding single feature recognition unit 13. The plurality of prepared single feature recognition units 13 independently perform the following identification processing to obtain the similarity of each character category with respect to each feature vector X _k . In addition,
The feature vector X _k used in the following description of the operation of the single feature recognizing unit 13 is a feature vector obtained by a certain feature extraction method, and hereinafter, X _k and n _k are described as X and n for simplicity.

The specific operation of the single feature recognition unit 13 will be described below.

In the single feature recognition unit 13, first, the feature vector X is input to the fuzzy major classification unit 21. In the fuzzy major classification unit 21, the feature vector X is input to the input unit 30 and m _r
X is output to the individual distance calculation units 31. Each distance calculator
31 is a group reference feature vector V _i (1 ≦ i ≦ m _r ; m _r is the number of group reference feature vectors representing the character category groups stored in the group dictionary 20; (Number) is read out, and the distance d _i between X and V _i shown in equation (8)
Are calculated and output to the corresponding divider 32 and multiplier 34, respectively.

[0052]

D _i = ‖X−V _i ‖ ^{2 / (f-1)} (1 ≦ i ≦ m _r ) where f is a real number that satisfies f> 1. Each divider 32
Calculates the reciprocal of the distance d _i and outputs the output to the adder 33. The adder 33 calculates the sum of the outputs of all the dividers 32 and outputs the output to the m _r multipliers 34. Each of the multipliers 34 multiplies the outputs of the corresponding distance calculation unit 31 and the adder 33, and inputs the output to the corresponding divider 35. Each divider 35 calculates the reciprocal of the output of the corresponding multiplier 34. Finally, in the fuzzy major classification unit 21, the feature vector X of the character pattern to which the output of each divider 35 is input.
Group membership degree μ _i (1 ≦ i ≦
m _r ) is output to the group selection unit 23. In other words, the group membership degree μ _i (1 ≦ i ≦ m _r ) of each character category group can be expressed as in Expression (9).

[0053]

[Equation 9]

The group reference feature vector representing each of the character category groups stored in the group dictionary 20 is stored in advance by a conventional clustering method, for example, “Character pattern information processing” by Shin Nagao edited by the Institute of Electronics, Information and Communication Engineers. (Corona) K-means algorithm and Isodata algorithm, Y. Linde, A. Buzo, and RM gray
(Gray)
lgorithm for Vector Quantizer design), "I-E-E-Transaction-on-Communication-Volume 28, pages 84-95 19
January 80 (IEEE Trans. Commun., COM-28, 1, pp.84-95, J
It is designed using the LBG algorithm shown in an.

In the following, the group dictionary will be calculated using the K-means algorithm.
A brief description of how to design the 20. (1) From the set of character patterns for designing a group dictionary of recognition objects, m _r (where m _r is the predetermined number of character category groups) character patterns are appropriately selected, and these feature vectors are set. m _r group reference feature vectors V _i (1 ≦ i ≦ m
_r ). (2) For all feature vectors X of the character patterns for designing the group dictionary, the distances d _i shown in the equation (10)

[0056]

[Number 10] seek a d _i = ‖X-V _i ‖ a to minimize V _i. At this time, X is assumed to belong to the character category group S _i (1 ≦ i ≦ m _r ). (3) Feature vector X of the character pattern belonging to each S _i
Of V _i 'is obtained. (4) If V _i ′ = V _i holds for all i, the group reference feature vector V _i at this time is stored in the group dictionary 20. If not, V _i 'is set as a new group reference character pattern signal V _i and the process returns to (2).

By designing the group reference feature vector in this manner, all the character patterns can be divided into some subsets (character category groups) of the character patterns having similar feature vectors. Note that Is
odata algorithm and LBG algorithm are also K
The method is basically the same as the averaging algorithm.

The group selection unit 23 selects a plurality of character category groups in the descending order of group membership obtained by the fuzzy large classification unit 21 and finely inputs group selection information indicating which character category group has been selected. The signal selecting unit 24 and the corresponding group membership degree are output to the single feature similarity calculating unit 25. As a method of selecting the character category group, a category group having a group membership degree equal to or higher than a certain threshold may be selected.

Based on the group selection information obtained from the group selection unit 23, the fine classification input signal selection unit 24 selects the fine classification unit 22 to which the feature vector X of the input character pattern is input,
The feature vector X is output to these subclassification units 22.

In each subclassification section 22 corresponding to the character category group selected by the group selection section 23 (that is, the subclassification section 22 to which the feature vector is input from the subclassification input signal selection section 24),
First, the character vector feature vector X is input to the input unit 40, and X is output to the m _c distance calculation units 42. Each distance calculation unit 42 is a character category reference feature vector W _i (1 ≦ i ≦ m _c ; m _c is a character category reference feature vector) indicating a representative value of each character category stored in the character category dictionary 41. X) shown in (Equation 11)
And the distance d _i between W _i and W _i are calculated and output to the corresponding divider 43 and multiplier 45, respectively.

[0061]

D _i = ‖X−W _i ‖ ^{2 / (f-1)} (1 ≦ i ≦ m _c ), where f is a real number satisfying f> 1. Each divider 43
Calculates the reciprocal of the distance d _i and outputs the output to the adder 44. The adder 44 calculates the total sum of the outputs of all the dividers 43 and outputs the output to the m _c multipliers 45. Each of the multipliers 45 multiplies the output of the corresponding distance calculation unit 42 and the output of the adder 44, and inputs the output to the corresponding divider 46. Each divider 46 calculates the reciprocal of the output of the corresponding multiplier 45. Finally, in the sub-classification unit 22, the output of each divider 46 is the single feature similarity as the intra-group similarity ν _i (1 ≦ i ≦ m _c ) of each character category with respect to the feature vector X of the character pattern. It is output to the degree calculation unit 25. That is, each subclassification unit 22
In-group similarity ν _i (1 ≦ i ≦
m _c ) can be expressed as in Expression (12).

[0062]

[Equation 12]

The character category reference feature vector, which indicates the representative value of each character category stored in the character category dictionary 41, is used for a character pattern having each character category in the character category group. Is designed in advance so that the output of the distance calculation unit 42 corresponding to the? Generates the minimum output as compared with the output of the other distance calculation units 42.

The method for designing the character category reference feature vector is, for example, a learning vector quantization method (LV).
This is done by a learning algorithm called Q). As for the learning vector quantization method, for example, T. Kohonen (Kohon
en) "Learning Vector Quantization for Pattern Recog"
nition) ", Helsinki University of Tec
hnology), Report TKK-F-A601, November 1986 (Report TKK-F-
It is shown in A601 (1986.11).

The learning vector quantization method will be briefly described below.

First, a character category reference feature vector W _i having m _c character categories is prepared. As the initial value of this W _i , a feature vector of a character pattern arbitrarily selected for each character category from a character pattern for designing a character category dictionary including a set of character patterns included in each character category group, and a group A reference feature vector obtained by a conventional clustering method such as the K-means algorithm explained when designing the dictionary 20 is used. Then, the selected feature vector X of the character pattern from the character pattern for the design of the character category dictionary with any one category C _X is, for this X sequentially, the following steps are repeated. (1) Select the character category reference feature vector W _C closest to X. However, the character category of this W _C is C _C. (2) If C _X = C _C , W _C approaches the X direction. On the other hand, if C _X ≠ C _C , W _C is moved away from X.
Further, character category reference feature vectors other than W _C are not updated.

The above-mentioned updating of the character category reference feature vector when X is presented is repeatedly performed for all the designed character patterns of the character category dictionaries prepared in advance.

In this way, by designing the character category reference feature vector of the character category dictionary 41,
The character vector reference feature vector having each character category is always located at the closest distance to the feature vector of the character pattern having each character category in the character category group. Therefore, the character category of the input character pattern can be recognized in each character category group by selecting the distance calculation section 42 that produces the minimum output among all the distance calculation sections 42.

In the single feature similarity calculation unit 25, first, the multiplier 26 determines the group membership degree of the character category group selected by the group selection unit 23 and each subclassification unit corresponding to the character category group.
22. That is, all the in-group similarities obtained from 22 (that is, the fine classification unit 22 to which the feature vector is input from the fine classification input signal selection unit 24) are multiplied, and the outputs are output to the category similarity calculation unit 27. That is, the number of multipliers 26 provided is (the number of group-reference feature vectors × the sum of the number of character category-reference feature vectors in each subclassification unit), and one category group p selected by the group selection unit 23. Group membership degree μ _p (1 ≦ p ≦ m _r ;
m _r is the number of character category groups) and the in-group similarity ν _pq (1 ≦ q ≦ m _c ; m _C is a character pattern) of a certain character category q obtained from the fine classification unit 22 corresponding to the character category group p The number of character categories of data) is input. That is, the multiplier 26
The output value ξ _pq of

[0070]

[ _Expression 13] ξ _pq = μ _p · ν _pq

The category similarity calculator 27 classifies the output values of all the multipliers 26 into each character category,
A plurality of ones having a large output value are selected. Then, the sum of these selected output values is calculated for each character category, and this is calculated as the similarity r _i of each character category to the feature vector of the character pattern input to the single feature recognition unit 13.
(1 ≦ i ≦ N _C ; N _C is the number of character categories) is output to the candidate character selection unit 14. As a method of selecting a plurality of output values of the multiplier 26 for each category, it is possible to select an output value of the multiplier 26 that is equal to or higher than a certain threshold value.

As described above, the single feature recognition unit 13
The feature vector of the character pattern obtained from a certain feature extraction unit 12 is subjected to a large classification, then a sub-classification, and finally a hierarchical identification that integrates the outputs thereof, whereby a certain feature data is obtained. The similarity of each character category of the character pattern is calculated.

In the candidate character selection section 14, first, the respective integrated similarity conversion sections 16 have their corresponding single feature recognition sections 13 respectively.
From, the similarity of each character category to each feature vector of the input character pattern is input, and this similarity is converted into an integrated similarity for each character category. Specifically, (Equation 14)
As shown in the formula, the similarity r _ki of each character category obtained from the corresponding single feature recognition unit 13 (1 ≦ k ≦ N _F : N _F is the number of types of feature vectors, 1 ≦ i ≦ N _C ; N _C is the number of character categories) and is divided by the maximum value r _kmax among these similarities to perform normalization, and the integrated similarity t _ki (where 0 ≦ t _ki ≦ 1, 1 ≦ i ≦ N _C ; N _C is the number of character categories) and is output to the corresponding primary candidate character selection unit 17.

[0074]

[ _Mathematical formula-see original _document ] t _ki = r _ki / r _{kmax In} each primary candidate character selection unit 17, the size information (width a and height b of the character pattern) of the character pattern and the position information obtained by the character cutting unit 11 h and standard size information (width A _i , height B _{i of} standard character pattern) of all character categories stored in the character information storage unit 50, standard position information H _i (1
≦ i ≦ N _C ; N _C is the number of character categories) and a plurality of candidate character categories (primary candidate character categories) are selected and output to the secondary candidate character selection unit 18. . For example, as a method of selecting a specific candidate character category,
The character width evaluation function F (a), the character height evaluation function F (b), and the character position evaluation function F (h) as shown in the formulas (15), (16), and (17) are

[0075]

F (a) = | a−A _i |

[0076]

## EQU16 ## F (b) = | b−B _i |

[0077]

## EQU17 ## F (h) = | h−H _i | may be set, and one having a value of each function F smaller than a certain threshold may be selected as a candidate category. In addition,
As the evaluation function, the ratio between the character width and the character height, that is, the character shape evaluation function as shown in the formula (18) using the character shape information a / b and the standard character shape information A _i / B _i May be used for.

[0078]

[Equation 18] F (a / b) = | (a / b) − (A _i / B _i ) | In addition, as a method of selecting the primary candidate character category in the primary candidate character selection unit 17, Characters that not only select a candidate character category using only the size information s and position information h of the pattern but also have a combined similarity of each character category obtained from the corresponding integrated similarity conversion unit 16 that is equal to or greater than a certain threshold value First, after selecting a plurality of categories, the size information s of the character pattern, the position information h, and the standard size information (A _i , B _i ) of all the character categories stored in the character information storage unit 50 as described above. , The standard position information H _i may be compared to select the primary candidate character category.

In the secondary candidate character selection section 18, first, the adder
181 adds the integrated similarities t _ki of all the primary candidate character categories obtained from the plurality of primary candidate character selecting units 17 for each same category as shown in the formula (19), The final similarity u _i (where 1 ≦ i ≦ N _C ; N _C is the number of categories) is calculated.

[0080]

[Formula 19]

The final candidate character selection unit 182 compares the final similarities u _i and selects a plurality of those having a large final similarity value, and the post-recognition processing unit selects them as secondary candidate character categories. Output to 15.

As described above, the candidate character selection unit 14
Is for selecting a plurality of candidate character categories by integrating the identification results, that is, the degrees of similarity, obtained by recognizing the respective feature vectors of the character patterns input in each single feature recognizing unit 13. Here, the integrated similarity conversion unit 16 sets the similarity obtained from the single feature recognition unit 12 to the integrated similarity that enables the secondary candidate character selection unit 18 to successfully integrate the identification results regarding each feature vector. It is to convert. Specifically, in order for the secondary candidate character selection unit 18 to finally select a candidate character category with high confidence, the most important value is the value of each character category obtained from each single feature recognition unit 13. The maximum value of the similarity and values in the vicinity thereof, that is, the first and second values obtained by each single feature recognition unit 13,
..., the value of the similarity of the candidate character category. However, the maximum similarity obtained from each single feature recognition unit 13 is affected by the total number of output units of each subclassification unit in the single feature recognition unit 13 (specifically, the output units of the subclassification unit). The larger the total number is, the smaller the maximum similarity tends to be.) Therefore, if the similarity obtained from each single feature recognition unit 13 is directly output to the primary candidate character selection unit 17, the average is always averaged. The priority is given to the identification result from the single feature recognizing unit 13 that has the largest maximum similarity. Therefore, as it is, it is difficult to recognize the advantage of using a plurality of types of feature vectors, that is, to correctly recognize a character pattern that is erroneously recognized by a certain feature vector based on the identification result by different feature vectors. However, in the integrated similarity conversion unit 16, by normalizing the similarity of each character category obtained from each single feature recognition unit 13 by its maximum similarity, the maximum similarity of each single feature recognition unit 13 Is always 1. Therefore, in the final selection of the candidate character category in the candidate character selection unit 14, it is possible to equally contribute the identification result of each single feature recognition unit 13, it is always a certain single feature recognition unit 13
The identification result from is no longer given priority, and a highly reliable candidate character category can be obtained.

The feature vector obtained by the feature extraction unit 12 is "I", "1", "-" as it is because the feature vector is extracted from the normalized character pattern.
It is not possible to accurately identify a character category that is characterized by the shape of a vertically or horizontally long character pattern such as. However, by effectively using the character size information and the position information in the primary candidate character selection unit 17, in the present embodiment, these character categories can be accurately identified.

Finally, the plurality of secondary candidate character categories obtained by the candidate character selection unit 14 are input to the post-recognition processing unit 15, and the size information and position information of the character pattern obtained by the character cutout unit 11 are input. Is used to determine the final recognized character category. Specifically, the post-processing target character set determination unit 52 performs the following processing.

First, all post-processing target character category sets stored in the post-processing character pair storage unit 51 are read out and compared with the first candidate character category of the secondary candidate character categories. At this time, if the first candidate character category is the post-processing target character category set, the first candidate character category stored in the character information storage unit 50 and each post-processing target character category set that is the same as the first candidate character category stored in the character information storage unit 50. The character size information and position information of the character category to be processed are read out, and the character pattern size information (character pattern width a and height b) and position information h obtained by the character cutout unit 11 are respectively compared and input. The character category that is more similar to the size information and the position information of the character pattern is determined as the recognized character category and is output. As a specific method of determining the recognized character category, there is a method of determining the character category having a small value as the recognized character category by using the evaluation function already shown in the equations (15) to (18). . The post-processing target character category set indicates similar upper and lower case letters such as "X, x", "tsu, tsu". For example, assume that the first candidate character of the secondary candidate character category is "X". Standard character width, standard character height stored in the character information storage unit 50,
In the character category "X", the standard position information is (0.6, 0.
9, 0.5), and "x" is (0.5, 0.6, 0.3). Then, at this time, the character width, the character height, and the position information of the character pattern obtained from the character cutout unit 11 are (0.52,
0.7, 0.32), for example, as an evaluation function,
Formula (15) + Formula (16) + Formula (17) (F (a, b, h) = F (a) + F (b) +
If F (h)) is used, the value of F is 0.4 for "X".
6 and 0.14 for "x", and the final recognized character category is "x".

As described above, the post-recognition processing section 15 effectively uses the character size information and the position information for the identification of the similar case which is difficult to be accurately identified only by the feature vector. Accurately identify character categories.

The above is the description of the operation of the character recognition apparatus of this embodiment.

Here, experimentally obtained results of character recognition in this embodiment will be shown. The character pattern data to be recognized is 3390 multi-font (Mincho, Gothic, etc.) character types (symbols, alphanumeric characters, hiragana, katakana, Greek characters,
JIS first level Kanji) characters were used. The feature vector for feature extraction is a 64-dimensional contour direction density feature and a 32-dimensional background density feature. The contour direction density is a feature amount indicating the direction and complexity of the character stroke in the local area, and the background density feature is a feature amount indicating the complexity of the background of the character. The detailed feature extraction method is Waki, Fujiwara, Takenouchi, Yokoe, and Shimizu, "Document Clean Copy System (3) -Character Recognition Algorithm and Its Hardware-", 1986 IEICE General Conference, 1512, 6-154.

Single feature recognition unit corresponding to each feature vector
The group dictionary 20 in 13, the character category dictionary 41 of each subclassification unit 22, the character information storage unit 50 in the post-recognition processing unit 15,
Designed with 27 fonts of 8 fonts. Here, the fuzzy major classification unit 21 in each single feature recognition unit 13 sets the number of character category groups (the number of feature vectors for group reference) to 7. Further, the number of character category reference feature vectors of each subclassification unit 22 is a total of 6483 of 1070,949,957,562,1029,920,996 in the character category dictionary 41 corresponding to the contour direction density feature, and also corresponds to the background density feature. Character category dictionary 41
Then, the total number of 1122,902,633,1287,751,1090,984 is 6769.

At the time of recognition, f in the formula 8 of the fuzzy major classification unit 21 in each single feature recognition unit was 1.2, and f in the formula 11 of the fine classification unit 22 was 1.3.
The number of character category groups selected by the group selection unit 23 is 7
That is, each feature vector is input to all the subclassification units 22. Further, in the category similarity calculation unit 27, the number of output values of the multiplier 26 selected for each character category is set to 4.

For comparison, a single feature recognition unit 13 was also designed in which the two feature vectors were combined into a one-dimensional 96-dimensional feature vector. The same character pattern for learning used in the design is used, the number of character category groups is 7, and the number of character category reference feature vectors in the character category dictionary 41 is 905,590,1255,382,933,1.
The total number of 272,1083 was 6420. For recognition, the same value was used for each parameter.

As the character pattern for evaluation, 27120 characters of 8 fonts different from the data used for designing each dictionary were used.

As a result of the experiment, a recognition rate of 98.33% was obtained in this example. In addition, the recognition rate of the character recognition device that was used for comparison and recognized by combining two feature vectors into one was 97.98%. That is, as in the present embodiment, it is better to identify a plurality of feature vectors individually and perform recognition by integrating the identification results, and
It can be seen that the recognition rate is improved as compared with the case where the recognition processing is collectively performed.

As described above, according to the present embodiment, a plurality of types of feature vectors are identified by the separate single feature recognition units 13 to obtain the similarity of each character category to each feature vector, The character selection unit 14 integrates the identification results obtained from the single feature recognition units 13 using all of these similarities, and finally recognizes the character pattern. Therefore,
It is easy to "correctly recognize a character pattern that is erroneously recognized by a certain feature vector from the identification result by different feature vectors", which is difficult when using a plurality of types of feature vectors together as in the conventional example. Therefore, it is possible to realize highly accurate recognition.

Further, in the conventional method, the time required for recognition increases due to the use of a plurality of types of feature vectors. The time required can be shortened as compared with the conventional example.

In the single feature recognition section 13 of this embodiment, the subclassification section 22 may have the following configuration instead of the configuration shown in FIG.

FIG. 6 is a block diagram specifically showing the configuration of the second embodiment of the subclassification unit 22. An input unit 60 is for inputting the feature vector of the character pattern output from the subclassification unit input signal selection unit 24. Reference numeral 61 is a multi-input / single-output signal processing unit, and the lower-layer input unit connected to it.
60, or the output of the multi-input / single-output signal processing unit 61 and the weighting factor, which is the degree of connection thereof, are multiplied and summed, and the result is thresholded and output. Here, these multiple input / output signal processing units have a layered structure and are connected to the network so that signals are propagated only to the upper layer without mutual coupling in each layer and the characteristics of the input character pattern are The in-group similarity, which is the degree of similarity to each character category in the character category group for the vector, is obtained.

FIG. 7 is a block diagram specifically showing the configuration of the multi-input / single-output signal processing section 61. In FIG.
Reference numeral 62 is an input unit for inputting an input signal. 63
Is a weighting coefficient storage unit that stores a weighting coefficient for weighting a plurality of input signals from the input unit 62. Reference numeral 64 denotes a multiplier that multiplies the weight coefficient of the weight coefficient storage unit 63 by the input signal from the input unit 62. An adder 65 sums up the output values of all the multipliers 64. 66
Is a threshold processing unit, which limits the output value of the adder 65 to a value within a certain range.

Second sub-classifying section 22 constructed as described above
The operation of the embodiment will be described below.

Similarly to the first embodiment, each subclassification section 22 corresponding to the character category group selected by the group selection section 23 (that is, the subclassification in which the feature vector is input from the subclassification input signal selection section 24). In the section 22), the feature vector X of the character pattern is first input to the input section 60. There are n input units 60, which are equal in number to the characteristic data of character patterns, and each characteristic data x _i is input to the corresponding input unit 60. Each of the multi-input one-output signal processing units 61 of the sub-classification unit 22 is shown in FIG.
The input unit 60 in the lower layer connected to it, or the output of the multi-input one-output signal processing unit 61 and the weighting coefficient storage unit, as shown in FIG.
The weighting factor w _ij stored in 63, which is the degree of the connection.
Are multiplied by the multiplier 64, the sum of the outputs of the respective multipliers 64 is calculated by the adder 65, the result is converted by the threshold processing unit 66, and the output value is output to the upper layer. That is, the output value I _i of the i-th multi-input / single-output signal processing unit 61 of a certain layer shown in FIG. 7 is I _j which is the input value to the input unit 62, and the input connected thereto and the degree of the connection. If a certain weighting coefficient is w _ij (the connection weight of the i-th multi-input one-output signal processing section and the j-th input), it can be expressed as in (Equation 20).

[0101]

[Equation 20]

The input / output characteristics of the threshold value processing unit 66 are shown in FIG. For example, the input / output characteristics of the threshold value processing unit 66 that limits the output to the range of (0, 1) can be expressed as in the equation (21).

[0103]

F (a) = 1 / (1 + exp (-a + θ)) where a is an input of the threshold processing unit 66. The input / output characteristic of the threshold value processing unit 66 may be a threshold value function other than the above.

The multi-input / single-output signal processing unit 61 in the uppermost layer is
It is set to the same number as the number of character categories of the character pattern included in each character category group, and each multi-input / one-output signal processing unit 61 in the uppermost layer corresponds to each of these character categories. That is, each multi-input one-output signal processing unit 61 in the top layer
Is the degree of similarity to each character category in the character category group with respect to the feature vector X of the character pattern, ν _i (1 ≦ i ≦ m _c ; m _c is the number of character category reference feature vectors) ) Is output to the single feature similarity calculator 25.

The weighting factor of each multi-input one-output signal processing unit 61 is such that, for a character pattern having each character category in the character category group, the multi-input one output of the uppermost layer corresponding to each character category. The signal processing unit 61 is designed in advance so as to generate the maximum output. As a result, each subclassification unit 22 can recognize the character category of the character pattern. The design method of these weighting factors is performed by a learning algorithm called an error back propagation method.
For the error backpropagation method, for example, DE Ramelhart (Rumel
hart), GE Hinton and RJ Williams (William
s) "Learning Representation by Back-Propagati
ng Errors), "Nature, vol.323, pp.533-536, Oc
t. 9, 1986, and its outline is as already explained in the conventional example.

FIG. 9 is a block diagram specifically showing the configuration of the second embodiment of the candidate character selection section 14 in the first embodiment of the present invention. In FIG. 9, 70 is a similarity normalization unit, the same number as the single feature recognition unit 13 is prepared, and the similarity of each character category obtained from the corresponding single feature recognition unit 13 The maximum value of the similarities is used for normalization. 71 is a primary candidate character selection section,
Similar to the similarity normalization unit 70, the same number as the single feature recognition unit 13 is prepared, and the normalized similarity obtained from the corresponding similarity normalization unit 70 and the character pattern obtained from the character cutout unit 11 respectively. Size information, position information, and character information storage unit 50
Using the standard size information and the standard position information of each character category stored in, a plurality of candidate character categories (which are called primary candidate character categories) are selected from all the character categories. is there. Reference numeral 72 denotes an integrated similarity calculation unit, which, like the similarity normalization unit 70, is provided in the same number as the single feature recognition unit 13, and performs non-linear conversion of the normalized similarity of each primary candidate character category to obtain them. The integrated similarity of each primary candidate character category is obtained by dividing the non-linearly converted normalized similarity by its sum value. Reference numeral 17 denotes a secondary candidate character selection unit, which has the same configuration as that of the first embodiment and uses the integrated similarities of the primary candidate character categories obtained from the integrated similarity calculation units 72 for all the first candidate character categories. A plurality of candidate categories (which are called secondary candidate categories) are selected from the next candidate character categories and output to the post-recognition processing unit 15.

Candidate character selection unit 14 configured as described above
The operation of the second embodiment will be described below.

First, the similarity normalization units 70 (Equation 2)
As shown in equation (2), the corresponding single feature recognizers
Similarity r _ki (1 ≦ k ≦
N _F : N _F is the number of types of feature vectors, 1 ≦ i ≦ N _C ; N _C
Is the number of character categories) and the maximum value r among these similarities
Normalization is performed by dividing by _kmax , and the normalized similarity s _ki (where 0 ≦ s _ki ≦ 1, 1 ≦ i ≦ N _C ; N _C is the number of character categories) corresponds to the first Output to the next candidate character selection unit 71.

[0109]

S _ki = r _ki / r _kmax Each primary candidate character selection section 71 basically uses the normalized similarity instead of the integrated similarity of each character category. The same operation as in the 14th embodiment is performed to select a plurality of candidate character categories (primary candidate character categories) and output them to the integrated similarity calculation unit 72.

In each integrated similarity calculation unit 72, the normalized similarity s _{ki of} each primary candidate character category obtained from each primary candidate character selection unit 71 (where 0 ≦ s _ki ≦ 1,1 ≦ i ≦
M _C ; M _C is the number of primary candidate character categories) is non-linearly transformed, and the normalized similarity of each non-linearly transformed primary candidate character category is normalized (divided) by their sum value.
Then, this is output to the secondary candidate character selection unit 18 as the integrated similarity t _ki (where 1 ≦ i ≦ M _C ). That is, the integrated similarity t _ki is expressed by the following equation (23) when the nonlinear function is f (X).
It can be expressed like an expression.

[0111]

[Equation 23]

The nonlinear function f (x) is a monotonically increasing function satisfying 0 ≦ f (x) ≦ 1, f (0) = 0, f (1) = 1 in 0 ≦ x ≦ 1. For example, as an example (Equation 24)
There are functions like expressions. However, 0 ≦ a ≦ 1.

[0113]

F (x) = x ^a The secondary candidate character selection unit 18 performs the same operation as that of the first embodiment to obtain all the primary candidates obtained from each integrated similarity calculation unit 71. A plurality of secondary candidate character categories are selected from all the primary candidate character categories using the integrated similarity of the character categories and output to the post-recognition processing unit 15.

Now, in each single feature recognizing unit 13, when closely observing the similarity of each character category when erroneously recognizing, (1) the magnitude of the maximum similarity is not so much related to erroneous recognition ( 2) It can be seen that there are many false recognitions when the difference between the maximum similarity and the second largest similarity is small. Therefore, in this embodiment, after the similarity obtained from each single feature recognition unit 13 is normalized by the maximum similarity, a process of emphasizing the difference between the maximum similarity and the second largest similarity, that is, the maximum similarity. When the difference between the degree of similarity and the second largest similarity is small, the difference is made smaller and the magnitude of the similarity itself is made smaller, so that a single feature that may cause misrecognition. Reliability (similarity) of the identification result obtained from the recognition unit 13
Is being lowered. Therefore, when the final candidate character category is selected by the secondary candidate character selection unit 18,
The identification result of the other single feature recognizing unit 13 becomes more important, and as a result, a highly reliable candidate character category can be obtained.

Here, the recognition results obtained experimentally in the character recognition device using the second embodiment of the candidate character selection unit 14 will be shown. The experiment was performed in the same manner as in the first example. However, in the integrated similarity calculation unit 72, the non-linear function used is shown in the equation (24), and a = 0.5.

As a result of the experiment, the recognition rate of 98.38% was obtained in this example, and the recognition rate was improved by 0.05% as compared with the first example.

As described above, according to this embodiment, by performing the conversion as described above on the similarity obtained from each single feature recognizing unit 13, the character pattern which is erroneously recognized by a certain feature vector becomes different from the different feature patterns. Correct recognition based on the vector identification result can be performed better than in the first embodiment, and more accurate recognition can be performed. (Embodiment 2) FIG. 10 is a block diagram of a character recognition apparatus according to the second embodiment of the present invention. In FIG. 10, 10 to 18 are the same as the character recognition device in the first embodiment shown in FIG. Reference numeral 80 denotes a character cutout control unit, which determines the character cutout unit according to the number of primary candidate character categories obtained from each primary candidate character selection unit 17.
It controls 11 operations. That is, in the second embodiment, the character clipping control unit 80 is added to the first embodiment of the character recognition device described above.

The operation of the character recognition device configured as described above will be described below. In this example,
Except for the character cut-out control unit 80, basically, the other constituent elements perform the same operation as in the first embodiment to perform the character recognition processing, and therefore, those operations will be briefly described here.

As in the first embodiment, first, the character cutting unit 11 uses a plurality of parameters to extract a character pattern from the document image input by the image input unit 10 as shown in FIG.
Each character area is extracted and the character pattern is extracted by each feature extraction unit 12
In addition, the size information and position information of the extracted character pattern are used as the primary candidate character selection unit in the candidate character selection unit 14.
17 and output to the post-recognition processing unit 15. Each of the feature extraction units 12 extracts different feature vectors X _k (k = 1 to N _F : N _F is the number of types of feature vectors) from the character patterns extracted from the character cutout unit 11, and they correspond to each other. Output to the single feature recognition unit 13. Each of the plurality of single feature recognition units 13 independently performs the same identification processing as in the first embodiment to obtain the similarity of each character category with respect to each feature vector X _k , and obtains this as a candidate character. Output to the selection unit 14. In the candidate character selection unit 14, first, each integrated similarity conversion unit 16, the similarity of each character category obtained from the single feature recognition unit 13 corresponding to each feature vector, the maximum value of those similarities. It is converted to integrated similarity by normalizing with. In each primary candidate character selection unit 17, the size information and position information of the character pattern obtained by the character cutout unit 11 and the standard size information and standard position information of all character categories stored in the character information storage unit 50. And a plurality of primary candidate character categories are selected and output to the secondary candidate character selection unit 18. As shown in the first embodiment, the method of selecting the primary candidate character category in the primary candidate character selection unit 17 is to select the candidate character category using only the size information and the position information of the character pattern. In addition, first select a plurality of character categories whose integrated similarity of each character category obtained from the corresponding integrated similarity conversion unit 16 is a certain threshold or more, and then, as described above, the size of the character pattern. A method of comparing the information and position information with the standard size information and standard position information of all the character categories stored in the character information storage unit 50 and selecting the primary candidate character category may be used.

Here, the character cut-out control unit 80 sums up the numbers of the primary candidate character categories obtained from the respective primary candidate character selection units 17, and if the total is 0, that is, it is selected. If there is no primary candidate character category
A plurality of parameters used when the character cutting unit 11 extracts a character pattern for each character area (as already described in the first embodiment, the standard size of a character area, the pitch, the space length of adjacent character areas, etc.). ) Is changed,
Once again, let the character cutout section 11 extract the character area,
The extracted character pattern is re-input to each single feature recognition unit 13.

On the other hand, the secondary candidate character selection unit 18 adds all the integrated similarities of the primary candidate character categories obtained from the respective primary candidate character selection units 17 for each same category, thereby After obtaining the final similarity to the primary candidate character category, a plurality of those having a large final similarity value are selected and output to the post-recognition processing unit 15 as secondary candidate character categories.

Finally, the post-recognition processing unit 15 uses the size information and position information of the character pattern obtained from the character cutting unit 11 to select from among the plurality of secondary candidate character categories obtained from the candidate character selection unit. Determine the final character category.

Generally, in a Japanese document written in full-width characters, when a half-width alphanumeric character and a character that is easy to separate such as "i" are consecutive, the character pattern is often cut out by mistake. In such a case, in the conventional example, it was difficult to correctly recognize the character pattern. However, according to the present embodiment, the first candidate character selection unit 17 compares the size information and position information of the character pattern extracted by the character cutout unit 11 with the standard size information and position information of each character category. By doing so, a plurality of primary candidate character categories are selected. Therefore, if the input character pattern differs from the standard size and position information of each character category to some extent or more, the candidate character category becomes 0. At this time, the character cutout control unit 80 determines that a character pattern cutout error has occurred in the character cutout unit 11, and minutely changes a plurality of parameters used when cutting out the character pattern in the character cutout unit 11 to change the character. Re-cut out the pattern. Then, since the recognition process is performed again, it is possible to accurately cut out a character string even with respect to a character string in which the above-described cutout error easily occurs, and it is possible to realize highly accurate recognition. (Embodiment 3) FIG. 11 is a block diagram of a character recognition apparatus according to a third embodiment of the present invention. In FIG. 11, 10 to 18 are the same as the character recognition device in the first embodiment shown in FIG. Reference numeral 90 denotes a teacher signal generation unit that generates a teacher signal using the recognized character category obtained from the post-recognition processing unit 15 and the correct character category of the character pattern extracted by the character cutout unit 11. Reference numeral 91 denotes an erroneous recognition frequency storage unit, which is the number of times the combination of the teacher signal (correct character category) and the recognized character category matches the post-processing target character category set stored in the post-processing target character set storage unit 51. The misrecognition frequency is stored for each post-processing target character category set. Reference numeral 92 denotes an erroneous recognition frequency updating unit that compares the combination of the correct answer character category and the recognized character category with each post-processing target character category set stored in the post-processing target character set storage unit 91 to detect the erroneous recognition frequency. The misrecognition frequency of each post-processing target character category set stored in the storage unit is updated, and
When the misrecognition frequency becomes higher than an arbitrary allowable number, the corresponding post-processing target character category set is output. 93 is a character information correction unit, which is standard size information of the post-processing target character category set obtained from the misrecognition frequency update unit 92 stored in the character information storage unit 50, standard position information is input character pattern size information, It is corrected by using the position information.

The operation of the character recognition device configured as described above will be described below. In this example,
Except for the teacher signal output unit 90, the erroneous recognition frequency storage unit 91, the erroneous recognition frequency updating unit 92, and the character information correction unit 93, basically, the other components perform the same operation as in the first embodiment, Since recognition processing is performed, these operations will be briefly described here.

As in the first embodiment, first, the character cutting unit 11 uses a plurality of parameters to extract a character pattern from the document image input by the image input unit 10 as shown in FIG.
Each character area is extracted and the character pattern is extracted by each feature extraction unit 12
In addition, the size information and position information of the extracted character pattern are used as the primary candidate character selection unit in the candidate character selection unit 14.
17 and output to the post-recognition processing unit 15. Each of the feature extraction units 12 extracts different feature vectors X _k (k = 1 to N _F : N _F is the number of types of feature vectors) from the character patterns extracted from the character cutout unit 11, and they correspond to each other. Output to the single feature recognition unit 13. Each of the plurality of single feature recognition units 13 independently performs the same identification processing as in the first embodiment to obtain the similarity of each character category with respect to each feature vector X _k , and obtains this as a candidate character. Output to the selection unit 14. In the candidate character selection unit 14, first, each integrated similarity conversion unit 16, the similarity of each character category obtained from the single feature recognition unit 13 corresponding to each feature vector, the maximum value of those similarities. It is converted to integrated similarity by normalizing with. In each primary candidate character selection unit 17, the size information and position information of the character pattern obtained by the character cutout unit 11 and the standard size information and standard position information of all character categories stored in the character information storage unit 50. And a plurality of primary candidate character categories are selected and output to the secondary candidate character selection unit 18. Secondary candidate character selection unit 18
Calculates the final similarity to all the primary candidate character categories by adding the integrated similarity of the primary candidate character categories obtained from each primary candidate character selection unit 17 for each same category. After that, a plurality of ones having a large final similarity value are selected and output to the post-recognition processing unit 15 as secondary candidate character categories. The post-recognition processing unit 15 uses the size information and position information of the character pattern obtained from the character cutting unit 11 to determine the final recognized character from the plurality of secondary candidate character categories obtained from the candidate character selection unit 14. Determine the category.

Here, the teacher signal generation unit 90 has the recognized character category and the character cutout unit 11 obtained from the post-recognition processing unit 15.
The correct answer character category of the character pattern extracted in step S1 is compared, and if they are different from each other, this correct answer character category is output to the misrecognition frequency update unit 92 as a teacher signal. The misrecognition frequency updating unit 92 reads all post-processing target character category sets stored in the post-processing target character set storage unit 51, and sequentially outputs a combination of a teacher signal (correct answer character category) and a recognized character category. Compare. If a combination of the teacher signal (correct answer character category) and the recognized character category matches the post-processing target character category set, the post-processing target character corresponding to the combination stored in the misrecognition frequency storage unit 91. Increase the misrecognition frequency of the category set by one. At this time, if the erroneous recognition frequency of the post-processing target character category set is higher than an arbitrary allowable number of times, the post-processing target character category set is output to the character information correction unit 93. Then, in the character information correction unit 93, when the post-processing target character category set is input, the standard size information and the standard position information of the post-processed character category pair stored in the character information storage unit 50 are actually used as characters. Correction is performed based on the size information and position information of the character pattern extracted by the cutout unit 11. Specifically, the correction method will now be described. In the post-processing character category pair (C ₁ , C ₂ ) (for example, “X”, “x”), the character category C stored in the character information storage unit 50 is described.
₁ of the standard character pattern width A _1, the standard character height B _1, the standard position information and H _1, also standard character pattern width A ₂ of the character category C _2, B ₂ a standard character height, standard The position information is H ₂ . Here, when the frequency of erroneously recognizing C ₁ as C ₂ is exceeded, the character information (A ₁ , B ₁ , H ₁ ) and (A ₂ , B ₂ , H ₂ ) of the corresponding character categories are obtained. Character information of each character category of the character pattern already input (a ₁ ,
Based on b ₁ , h ₁ ) and (a ₂ , b ₂ , h ₂ ), for example,
It is corrected as in the equation (25).

[0127]

[Expression 25] A ₁ = A ₁ + α (A ₁ −a ₁ ) B ₁ = B ₁ + β (B ₁ −b ₁ ) H ₁ = H ₁ + γ (H ₁ −h ₁ ) A ₂ = A ₂ + α (A ₂ −a ₂ ) B ₂ = B ₂ + β (B ₂ −b ₂ ) H ₂ = H ₂ + γ (H ₂ −h ₂ ) where α, β, and γ are arbitrary real numbers. , (A ₁ , b ₁ , h ₁ ) and (a ₂ , which are correction criteria for character information in each of the character categories C ₁ and C ₂ ).
b ₂ and h ₂ ) are average values of the character information of the character patterns already input.

As described above, according to this embodiment, each of the characters stored in the character information storage unit 50 used when the post-recognition processing unit 15 discriminates between similar case letters such as “X” and “x”. It is possible to correct the standard character size information and position information of the character category based on the size information and position information of the actually input character pattern. That is, since the size information and the position information of the similar case are learned by performing the recognition process of the document image in real time, the post-recognition processing unit 15
It is possible to adaptively improve the identification of the similar case in, and finally realize a highly accurate recognition performance.

Further, the character selection / recognition post-processing unit of the present invention according to claim 1 corresponds to the candidate character selection unit 14 and the post-recognition processing unit 15 in the above embodiment.

[0130]

As is apparent from the above description,
According to the present invention, feature vectors of a plurality of types of character patterns are identified by separate single feature recognition units to determine the similarity of each character category to each feature vector of a character pattern, and then the candidate character selection unit is selected. Integrates the identification results obtained from each single feature recognizing unit using all of these similarities, and finally recognizes the input character pattern. Therefore,
It is difficult to collectively use the feature vectors of a plurality of types of character patterns as in the conventional example, “correctly recognize a character pattern that is erroneously recognized by a certain feature vector from the identification result by a different feature vector”. This makes it possible to realize highly accurate character recognition.

Further, in the conventional method, the time required for recognition is increased by using the feature vectors of a plurality of types of character patterns, but in the present embodiment, each single feature recognition section carries out the identification processing in parallel. The time required for recognition can be shortened as compared with the conventional example.

[Brief description of drawings]

FIG. 1 is a block diagram of a character recognition device according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a single feature recognizing unit according to the first embodiment of the present invention.

FIG. 3 is a block diagram of a fuzzy major classification unit according to the first embodiment of the present invention.

FIG. 4 is a first subclassification section in the first embodiment of the present invention.
FIG.

FIG. 5 is an explanatory diagram showing an example of a character pattern extracted by a character cutout unit and its character size information and position information according to the first embodiment of the present invention.

FIG. 6 is a second subclassification section in the first embodiment of the present invention.
FIG.

FIG. 7 is a second subclassification section according to the first embodiment of the present invention.
Block diagram of a multi-input single-output signal processing unit in the embodiment of FIG.

FIG. 8 is a second subclassification section according to the first embodiment of the present invention.
FIG. 7 is an input / output characteristic diagram of a threshold processing unit in the multi-input / single-output signal processing unit in the example of FIG.

FIG. 9 is a block diagram showing a second embodiment of a candidate character selection unit in the first embodiment of the present invention.

FIG. 10 is a block diagram of a character recognition device according to a second embodiment of the present invention.

FIG. 11 is a block diagram of a character recognition device according to a third embodiment of the present invention.

FIG. 12 is a block diagram showing an embodiment of a conventional character recognition device.

[Explanation of symbols]

 10 Image input unit 11 Character cutout unit 12 Feature extraction unit 13 Single feature recognition unit 14 Candidate character selection unit 15 Post-recognition processing unit 16 Integrated similarity conversion unit 17 Primary candidate character selection unit 18 Secondary candidate character selection unit 181 Adder 182 Final candidate character selection unit 50 Character information storage unit 51 Post-processing target character set storage unit 52 Post-processing target character set determination unit 20 Group dictionary 21 Fuzzy large classification unit 22 Fine classification unit 23 Group selection unit 24 Fine classification unit Input signal selector 25 Single feature similarity calculator 26 Multiplier 27 Category similarity calculator 30 Input 31 Distance calculator 32 Divider 33 Adder 34 Multiplier 35 Divider 40 Input 41 Character category dictionary 42 Distance calculation unit 43 Divider 44 Adder 45 Multiplier 46 Divider 60 Input unit 61 Multi-input one-output signal processing unit 62 Input unit 63 Weighting coefficient storage unit 64 Multiplier 65 Adder 66 Threshold processing unit 70 Similar Degree normalization unit 71 primary candidate character selection unit 72 integrated similarity calculation unit 80 Shaped cutout control unit 90 teacher signal generator 91 misrecognized frequency storage unit 92 misrecognized frequency updating unit 93 character information correction unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Taiji Yuki, 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Koji Yamamoto, 1006 Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Co., Ltd. (72) Inventor Toshio Niwa 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd.

Claims (10)

[Claims] 1. An image input section for inputting a document image, and a character cutout section for cutting out a character from the document image input by the image input section for each character area to obtain a character pattern.
A plurality of feature extraction units that obtain different feature vectors from the character patterns; a plurality of single feature recognition units that obtain a similarity, which is the degree to which the character pattern belongs to each character category, from each of the feature vectors; And a character selection recognition post-processing unit that integrates the similarity of each character category obtained from the single feature recognition unit for each same character category and determines the recognized character category using the result. One feature recognition unit, a group dictionary in which a plurality of group reference feature vectors representing a character category group consisting of a set of character patterns having similar feature vectors are stored,
A fuzzy large classification unit that calculates a group membership degree that is a degree to which a character pattern belongs to each character category group using the group reference characteristic vector and the characteristic vector of the character pattern,
A plurality of sub-classification units that obtain the intra-group similarity that is the degree to which the character pattern belongs to each character category included in the character category group using the feature vector of the character pattern;
A group selection unit that selects a plurality of character category groups from the group membership degree, and a subclassification unit input that selects a subclassification unit that inputs the feature vector of the character pattern based on the group selection information obtained from the group selection unit A single feature that obtains the similarity of each character category to a character pattern using the signal selection unit, the group membership degree of the character category group selected by the group selection unit, and the intra-group similarity obtained by the sub-classification unit A similarity calculation unit, wherein the single feature similarity calculation unit inputs the group attribution of the character category group selected by the group selection unit and the feature vector of the character pattern from the subclassification unit input signal selection unit. A plurality of multipliers for multiplying all the in-group similarities obtained from the subclassifying unit, and a plurality of multipliers having a large output value for each character category are selected, The power to find the sum A character recognition device comprising a territory similarity calculation unit. 2. An image input unit for inputting a document image, and a character pattern and size information of the character pattern obtained by cutting out a character for each character region from the document image input by the image input unit using a plurality of parameter values. A character segmentation unit that obtains position information, a plurality of feature extraction units that obtain different feature vectors from the character patterns, and a plurality of unit vectors that obtain the degree of similarity of the character patterns belonging to each character category from each of the feature vectors. One feature recognition unit, a candidate character selection unit that obtains a plurality of secondary candidate character categories for the character pattern using the similarity of each character category obtained from the plurality of single feature recognition units, and the character pattern Size information and position information of the plurality of second
A post-recognition processing unit that determines a recognized character category from the next candidate character categories, wherein the single feature recognition unit is for group reference that represents a character category group composed of a set of character patterns having similar feature vectors. A group dictionary in which a plurality of feature vectors are stored,
A fuzzy large classification unit that calculates a group membership degree that is a degree to which a character pattern belongs to each character category group using the group reference characteristic vector and the characteristic vector of the character pattern,
A plurality of sub-classification units that obtain the intra-group similarity that is the degree to which the character pattern belongs to each character category included in the character category group using the feature vector of the character pattern;
A group selection unit that selects a plurality of character category groups from the group membership degree, and a subclassification unit input that selects a subclassification unit that inputs the feature vector of the character pattern based on the group selection information obtained from the group selection unit A single feature that obtains the similarity of each character category to a character pattern using the signal selection unit, the group membership degree of the character category group selected by the group selection unit, and the intra-group similarity obtained by the sub-classification unit A similarity calculation unit, wherein the single feature similarity calculation unit inputs the group attribution of the character category group selected by the group selection unit and the feature vector of the character pattern from the subclassification unit input signal selection unit. A plurality of multipliers for multiplying all the in-group similarities obtained from the subclassifying unit, and a plurality of multipliers having a large output value for each character category are selected, The power to find the sum And a candidate similarity selecting unit, wherein the candidate character selecting unit normalizes the similarity of each character category obtained from the single feature recognizing unit corresponding to each of the feature vectors by the maximum value of the similarity. A plurality of integrated similarity conversion units for converting to integrated similarity by converting, the integrated similarity and size information of the character pattern,
A plurality of primary candidate character selection units that select a plurality of primary candidate character categories from all the character categories by using position information, and an integrated similarity of each of the selected primary candidate character categories are displayed. A secondary candidate in which the final similarities of all the primary candidate character categories are obtained by adding up for each same category, and a plurality of those having a large final similarity value are selected as secondary candidate character categories. A character recognition device, comprising a character selection unit. 3. The post-recognition processing unit includes a character information storage unit that stores standard size information and standard position information of a character pattern for each character category, and a post-processing unit that stores a post-processing target character category set. A target character set storage unit, a first candidate character category of the secondary candidate character category obtained from the candidate character selection unit, and all post-processing target character category sets stored in the post-processing target character set storage unit. And if the first candidate character category is the post-processing target character category set, the standard size information and standard position of each character category corresponding to the post-processing target character category set from the character information storage unit. A character is recognized by reading out the information and comparing it with the size information and the position information of the input character pattern to recognize any one of the character categories in the character category set. A post-processing target character set determination unit that determines a category, and the primary candidate character selection unit includes the standard size information and standard position information of the character category stored in the character information storage unit, and the input character pattern. The character recognition device according to claim 2, wherein a plurality of primary candidate character categories are selected by comparing the size information and the position information of the first candidate character category. 4. The primary candidate character selection unit selects a category having an integrated similarity degree of each category obtained from the integrated similarity degree conversion unit that is larger than an arbitrary threshold value, and then selects the category in the character information storage unit. By comparing the standard size information and standard position information of the stored character category with the size information and position information of the input character pattern
The character recognition device according to claim 2, wherein the next candidate character category is selected. 5. The number of primary candidate character categories obtained from each primary candidate character selection unit is summed up, and if the total is 0, the character segmentation is performed by changing each parameter value of the character segmentation unit. The character recognition device according to claim 2, 3 or 4, further comprising: a character cut-out control unit for retriing. 6. A teacher signal generation unit for outputting a correct character category when the recognized character category obtained from the post-recognition processing unit and the correct character category of the input character pattern are different, the correct character category and the correct character category. A misrecognition frequency storage unit in which a misrecognition frequency, which is the number of times the combination with the recognized character category matches the post-processing target character category set, is stored for each post-processing target character category set, the correct character category, and the recognition The combination with the character category and each post-processing target character category set stored in the post-processing target character set storage unit are compared, and if they match, the corresponding post-processing target character category set in the misrecognition frequency storage unit is selected. The misrecognition frequency is updated, and when the misrecognition frequency exceeds an arbitrary allowable number, the corresponding post-processing target character category set is output. Using the new part and the standard size information and standard position information of the post-processing target character category set obtained from the misrecognition frequency update part stored in the character information storage part, using the size information and the position information of the character pattern. 4. A character information correcting unit for correcting is provided.
The character recognition device according to item 4 or 5. 7. The candidate character selection unit includes a plurality of similarity normalization units for normalizing the similarity of each character category obtained from the single feature recognition unit with the maximum value of the similarity, A plurality of primary candidate character categories are selected from all the character categories using the normalized similarity of each character category obtained from the similarity normalization unit and the size information and position information of the character pattern. A non-linear conversion of the primary candidate character selection unit and the normalized similarity of the primary candidate character category,
An integrated similarity calculation unit that obtains an integrated similarity of each character category by dividing the non-linearly converted normalized similarity by its sum value, and an integrated similarity of each selected primary candidate character category. By adding degrees for each of the same categories, the final similarity of all the first candidate character categories is obtained, and a plurality of those having a large final similarity value are selected as secondary candidate character categories. The next candidate character selection unit is provided, and the second candidate character selection unit is provided.
The character recognition device according to any one of 1 to 6. 8. The fuzzy major classification unit includes a plurality of distance calculation units that calculate distances between the feature vector of the input character pattern and all the group reference feature vectors stored in the group dictionary, and the distance calculation unit. A plurality of dividers for calculating the reciprocal of the output of the above, an adder for adding the outputs of the dividers, and a plurality of multipliers for multiplying the output of the adder and the output of the distance calculator. 8. The character recognition device according to claim 1, further comprising a plurality of dividers for calculating the reciprocal of the output of the multiplier. 9. The subclassification unit stores a character category dictionary in which a plurality of character category reference feature vectors indicating a representative value of each character category of a character pattern are stored, and the feature vector and the character category dictionary. A plurality of distance calculation units that calculate the distances to all the character category reference feature vectors, a plurality of dividers that calculate the reciprocal of the output of the distance calculation unit, and the outputs of the dividers. A plurality of multipliers for multiplying the output of the adder and the output of the distance calculator, and a plurality of dividers for calculating the reciprocal of the output of the multiplier. The character pattern recognition device according to claim 1, wherein the character pattern recognition device is a character pattern recognition device. 10. The subclassification unit comprises a plurality of multi-input / output signal processing units each having a layered structure, are connected to each other in a network so that signals are propagated only to an upper layer without mutual coupling in each layer, and The multi-input-output signal processing unit includes a weighting factor storage unit that holds a plurality of weighting factors, an input unit that inputs a plurality of input signals, and a weighting factor stored in the weighting factor storage unit from the input unit. A multiplying means for weighting the input signal; an adding means for adding the plurality of input signals weighted by the multiplying means; and a threshold processing section for limiting the output of the adding means to a value within a certain range. The character recognition device according to claim 1, wherein the character recognition device is a character recognition device.