JP6327963B2 - Character recognition device and character recognition method - Google Patents

Description

  The present invention relates to a character recognition apparatus and a character recognition method capable of recognizing handwritten characters entered on a form or the like with high accuracy, and in particular, Chinese characters written in simplified Chinese characters or traditional Chinese characters are enhanced. The present invention relates to a character recognition device and a character recognition method that can be recognized accurately and at high speed.

  A character recognition device that optically reads handwritten Chinese characters written on a form or the like and extracts character features extracted from each character, and the Chinese character structure is located on the left side of the Chinese character. ), 旁 located on the right side of the Chinese character structure, crown on the upper side of the Chinese character structure, leg located on the lower side of the Chinese character structure, and the entire Chinese character structure Arranged in multiple areas of the structure that is located, the hanging that is hung from the upper side to the lower side of the kanji, and the ridge that is located along the lower side from the left side of the kanji structure In order to form the combination of the kanji elements, the divided regions for dividing each region are discriminated, and the feature amount of the character image is extracted using the Euclidean distance value to perform character recognition.

  As shown in FIG. 18, the character recognition processing according to this prior art is performed in step S1801 in which the computer of the character recognition device extracts the features of the character image, and whether the character image is a Chinese character or not based on the feature amount extracted in step S1801. Step S1802 for determining whether the character is kanji, step S1804 for calculating the Euclidean distance value of the character image determined in step S1802, step S1803 for determining a divided region such as a radical of the character image of kanji, and region by step S1803 Step S1805 for calculating the cosine similarity for each n-divided region of the determined character image, and step S1806 for selecting the isomorphic character and the similar character candidate by comparing the similarity and the isomorphic character table calculated in step S1805. , Calculated in step S1804 By executing step S1807 for determining the recognition candidate character based on the Clic distance and the isomorphic character and similar character candidate selected in step S1806, and outputting the determination result of the character candidate determined in step S1807. It is configured to perform character recognition. The isomorphic character table is a table in which kanji characters are classified by common radicals, wrinkles, etc., and kanji characters are classified by character shapes that are similar to human eyes.

  In step S1807, candidate determination is performed by (1) a condition for selecting a candidate from the kanji / non-kanji result obtained in step S1802, (2) a condition for selecting a candidate having a small Euclidean distance value in step S1804, and (3) step S1805. (4) a condition for selecting a candidate with a high cosine similarity in a partial region, and (5) an isomorphic character classification table to which a character type with a high cosine similarity belongs. The recognition character candidate is determined using the selection condition, and the Urid distance value and the cosine similarity are calculated by the mathematical formula shown in FIG.

  In other words, the conventional character recognition process is configured to determine a recognized character candidate by combining various techniques such as a Urid distance value, cosine similarity, region division, and isomorphic character table comparison of a character image.

  In addition, as a document in which a character recognition technique according to the prior art is described, the following Patent Document 1 can be cited. In Patent Document 1, a feature amount extraction unit from a character image input by an input unit includes a feature amount of a character image. When the matching unit calculates the degree of difference from the dictionary using the pseudo Bayes discriminant function, the number of principal components (k of the pseudo Bayes discriminant function equation) calculated by the pseudo Bayes discriminant function is adjusted by the character image. By calculating the degree of difference from the dictionary, by reducing the calculation of the principal component correction term that contributes to the improvement of character recognition accuracy, small features unnecessary for character recognition are deleted, and feature deficient components The following Patent Document 2 is cited as a document in which a technique for supplementing the above is described, and a technique for performing character recognition by dividing the radicals described above. Was Based on the coordinate series of each stroke of the written character, it is determined whether or not the stroke extends over the divided area in the character frame. Based on the integrated divided area, the handwritten character is cut out in radical units, and the radical recognition unit performs a recognition process for each cut out radical by referring to the radical recognition dictionary. A technique is described in which a recognition result of handwritten characters is easily acquired based on the result of the neck recognition process and the character composition dictionary, thereby improving the recognition accuracy by easily performing radical extraction, and a technique using the Euclidean distance for character recognition The following Patent Document 3 is cited as a document in which is described, and in this Patent Document 3, normalization means for normalizing a character image portion to a rectangular image of a predetermined size, and normalization by the normalization means. HOG features of an image portion obtained by combining feature amount calculation means for calculating HOG (Histograms of Oriented Gradient) feature amounts of each of the divided images obtained by dividing the rectangular image, and HOG feature amounts of the respective divided images in the HOG feature amount space. A distance calculating means for calculating the Euclidean distance between the amount and the HOG feature value of each predetermined character image, a comparing means for comparing the distance calculated by the distance calculating means, and the character image based on a comparison result of the comparing means A technique for providing character recognition means for recognizing characters that can be included in a portion is described.

JP 2000-215272 A JP-A-11-184971 JP 2014-29608 A

  In the conventional character recognition technology, a large number of features are extracted from a character image to be recognized and a correct class (for example, “A”, “A”, etc.) is estimated. Therefore, processing time becomes redundant when performing high-accuracy recognition, and in recent years, there is a need for character recognition not only for Japanese kanji but also for simplified characters in China and traditional characters in Taiwan. As a result, the processing time becomes redundant.

  In addition, character recognition devices have customer needs for high-speed character recognition of large-scale forms even with relatively low accuracy, and customer needs for high-precision character recognition of large-scale forms even when processing time is long. However, the conventional character recognition device has a problem that it is difficult to adjust the processing speed and the recognition accuracy according to customer needs.

  Furthermore, the technique described in Patent Document 1 is intended to achieve high speed and high accuracy by deleting small features unnecessary for character recognition and complementing the feature deficient components. However, there is a problem that a certain number of misidentifications are allowed, and the techniques described in other patent documents have a problem that high-speed processing is difficult.

  A first object of the present invention is to provide a character recognition device and a character recognition method capable of recognizing kanji characters at high speed and high accuracy. The second object is to provide a character recognition device and a character recognition method capable of adjusting the speed and recognition accuracy according to customer needs.

The invention of claim 1, wherein in order to achieve the object of, by using the input unit for reading in optical scan characters including Chinese characters consisting of Chinese elements a plurality of areas as a character image, the input unit scan A character identification dictionary storing a plurality of standard character images, and classifying kanji with similar radicals and similar shapes by classifying them together with a common radical, 旁, etc. An external storage unit for storing an isomorphic character table, and an arithmetic unit for controlling character recognition of a character image with reference to the character identification dictionary and the isomorphic character table, wherein the arithmetic unit stores the input character image A character recognition device that performs character recognition processing with reference to a character identification dictionary,
The computing unit is
A divided region determination step of determining a divided region for dividing the character image in units of one character scanned by the image acquisition unit for each Chinese character element;
A cosine similarity calculation step of calculating the cosine similarity of the kanji elements in the same region position of the character image determined by the divided region determination step;
A Euclidean distance value calculating step of calculating a Euclidean distance value of a Chinese character element at the same region position of the character image determined by the divided region determining step;
Character images having the same kanji composition based on the divided areas determined in the divided area determining step and having a similarity between kanji elements in the same region position based on the cosine similarity calculated in the cosine similarity calculating step higher than a predetermined value And an isomorphic character table creating step for storing character images whose Euclidean distance values between Kanji elements calculated by the Euclidean distance value calculating step are smaller than a predetermined value and grouped as isomorphic characters with high similarity,
A character image whose cosine similarity calculated by the cosine similarity calculation step is larger than a predetermined cosine similarity, a character image whose Euclidean distance value calculated by the Euclidean distance value calculation step is smaller than a predetermined value, and creation of the isomorphic character table And a standard character recognition step of performing character recognition using a character image of a logical sum with a character image of isomorphic characters obtained by grouping character images having high similarity in the same region position by a process as a first character. And

  According to a second aspect of the present invention, in the character recognition device according to the first feature, the calculation unit includes a plurality of divided regions divided by the divided region determination step from a character image obtained by a logical sum in the standard character recognition step. A second feature is that a heterogeneous candidate character having a different configuration and a high-definition character recognition step of performing character recognition based on a candidate character excluding a character image included in a group to which the heterogeneous candidate character belongs are performed.

According to a third aspect of the present invention, in the character recognition device according to the first or second feature, a character image having a Euclidean distance value calculated by the Euclidean distance value calculating step smaller than a predetermined value as a candidate character is used as a candidate character. The third feature is to execute a high-speed character recognition process for recognition.

  According to a fourth aspect of the present invention, in the character recognition device having any one of the first to third features, the kanji based on the divided region determined by the divided region determining step in the isomorphic character table creating step by the arithmetic unit. A large grouping step for dividing character images having the same structure into large groups, and a cosine similarity of a kanji element at a specific region position in a large groove grouped by the large grouping step is higher than a predetermined value and a character image And a small grouping step of dividing character images whose Euclidean distance values are smaller than a predetermined value into small groups, and storing the large grouped and small grouped kanji images in an isomorphic character table, as a fourth feature, The invention according to Item 5 is the character recognition device according to the third or fourth feature, wherein the calculation unit is configured to perform the standard character recognition step. Switching between accurate character recognition process and the high-speed character recognition process, or the fifth aspect, further comprising a function for adjusting.

The invention of claim 6 wherein acquires the input unit for reading in optical scan characters including Chinese characters consisting of Chinese elements a plurality of areas as a character image, a character image scanned by using the input unit An image acquisition unit, a character identification dictionary storing a plurality of standard character images, and a homomorphic character table for classifying kanji characters that are similar by visual inspection while classifying kanji characters by a common radical or 旁A character recognition method of a character recognition device comprising: an external storage unit for storing; and an arithmetic unit that controls character recognition of a character image with reference to the character identification dictionary and the isomorphic character table,
In the calculation unit,
A divided region determination step of determining a divided region for dividing the character image in units of one character scanned by the image acquisition unit for each Chinese character element;
A cosine similarity calculation step of calculating the cosine similarity of the kanji elements in the same region position of the character image determined by the divided region determination step;
A Euclidean distance value calculating step of calculating a Euclidean distance value of a Chinese character element at the same region position of the character image determined by the divided region determining step;
Character images having the same kanji composition based on the divided areas determined in the divided area determining step and having a similarity between kanji elements in the same region position based on the cosine similarity calculated in the cosine similarity calculating step higher than a predetermined value And an isomorphic character table creating step for storing character images whose Euclidean distance values between Kanji elements calculated by the Euclidean distance value calculating step are smaller than a predetermined value and grouped as isomorphic characters with high similarity,
A character image whose cosine similarity calculated by the cosine similarity calculation step is larger than a predetermined cosine similarity, a character image whose Euclidean distance value calculated by the Euclidean distance value calculation step is smaller than a predetermined value, and creation of the isomorphic character table And a standard character recognition step of performing character recognition using, as a candidate character, a character image that is a logical sum of character images of isomorphic characters obtained by grouping character images having high similarity in the same region position by the process. And

  The invention according to claim 7 is the character recognition method of the sixth feature, wherein a plurality of divided regions divided by the divided region determining step from the character image obtained by the logical sum in the standard character recognition step in the arithmetic unit. And a high-definition character recognition step of performing character recognition based on candidate characters excluding character images included in a group to which the different candidate characters belong, and a different character candidate having different configurations.

The invention according to claim 8 is to cause the arithmetic unit to execute a high-speed character recognition step of performing character recognition using a character image whose Euclidean distance value calculated by the Euclidean distance value calculation step is smaller than a predetermined value as a candidate character. Eight features.

  The invention according to claim 9 is the character recognition method according to any of the sixth to eighth features, wherein the kanji based on the divided area determined by the divided area determining step in the isomorphic character table creating step is provided in the arithmetic unit. A large grouping step for dividing character images having the same structure into large groups, and a cosine similarity of a kanji element at a specific region position in a large groove grouped by the large grouping step is higher than a predetermined value and a character image Including a small grouping step of dividing character images whose Euclidean distance values are smaller than a predetermined value into small groups, and storing the large grouping and small grouped kanji images in an isomorphic character table, In the character recognition method according to any of the eighth and ninth features, the standard character recognition step and the high-precision character recognition are included in the calculation unit. Switching between step and high speed character recognition process, or a tenth aspect of the letting a function of adjusting.

  In the character recognition device and the character recognition method according to the present invention, the arithmetic unit has the same kanji composition in which kanji composed of kanji elements in a plurality of areas are divided for each area, and the similarity between kanji elements in the same area position based on cosine similarity For character identification, an isomorphic character table in which character images whose degrees are higher than a predetermined value and character images whose Euclidean distance values between Kanji elements calculated in the Euclidean distance value calculation step are smaller than a predetermined value are grouped as isomorphic characters with high similarity. By performing character recognition with reference to candidate characters stored in the dictionary 165 and stored in the isomorphic character table, it is possible to recognize kanji characters including simplified characters and traditional characters with high accuracy and high speed.

The figure which shows the basic composition of the character recognition apparatus by embodiment of this invention. The figure which shows the outline | summary of the character recognition process by this embodiment. The figure which shows the recognition output candidate determination process in the high speed mode by this embodiment. The figure which shows the recognition output candidate determination process in the standard mode by this embodiment. The figure which shows the isomorphic character table and character selection process in the standard mode by this embodiment. The figure which shows the recognition output candidate determination process in the high precision mode by this embodiment. The figure which shows the character selection process in the high precision mode by this embodiment. The figure which shows the isomorphic character / classical character selection processing operation by this embodiment. The figure which shows the division area determination processing operation | movement by this embodiment. The figure which shows the division area determination processing operation | movement by this embodiment. The figure which shows the creation processing operation | movement of the isomorphic character table by this embodiment. The flowchart which shows the division area determination processing operation by this embodiment. The figure explaining the division position candidate calculation process in the division area determination process by this embodiment. The figure which shows the large group classification table creation processing operation by this embodiment. The flowchart which shows the similar area | region determination processing operation | movement by this embodiment. The figure which shows the table at the time of the similarity calculation according to area | region by this embodiment. The figure which shows the small group classification table creation processing operation by this embodiment. The figure explaining the isomorphic character table by this embodiment. The figure explaining the numerical formula which calculates the similarity degree by this embodiment. The figure which shows the output candidate determination processing operation by a prior art. The figure explaining the calculation range by Euclidean distance. The figure explaining the determination method by Euclidean distance and cosine similarity. The figure which shows the relationship between the processing time which character recognition requires, and recognition accuracy. The figure which shows the numerical formula of the similarity and cosine similarity by a Euclidean distance value and a Mahalanobis distance value. The figure which shows the outline | summary of the character recognition process by a prior art. The figure which shows the operation screen which switches the recognition mode by this embodiment, or adjusts.

Hereinafter, an embodiment of a character recognition apparatus using a character recognition method according to the present invention will be described in detail.
[Constitution]
As shown in FIG. 1, the character recognition apparatus according to the present embodiment has an input unit 110 such as a scanner that optically reads characters entered in a form and the like, and an image that acquires a character image scanned using the input unit 110. An acquisition unit 130, a character identification dictionary that stores a plurality of standard character images, and an isomorphic character table for classifying kanji characters that are similar by visual inspection while classifying kanji characters by a common radical, 旁, etc. An external storage unit 160 that stores a character identification dictionary 165 that includes the calculation unit 150 that controls character recognition of a character image with reference to the character identification dictionary 165 and the isomorphic character table stored in the external storage unit 160, and the calculation A display unit 120 for displaying characters identified by the unit 150, unreadable characters, and the like, and a communication unit 140 for transmitting data of character recognition results to the outside. 150 is configured to perform a character recognition process character images stored in the external storage unit 160 with reference to the character recognition dictionary 165.

In the calculation unit 150 according to the present embodiment, the recognition accuracy is lower than the standard (default) mode in consideration of the balance between the recognition speed and the recognition accuracy and the standard mode according to an instruction from the operator using the display unit 120. Three modes can be set: a high-speed mode with high recognition processing and a high-accuracy mode with high recognition accuracy, although the recognition speed is lower than the standard mode, and fine adjustment of the recognition speed and recognition accuracy is possible. It is configured as possible.
As shown in FIG. 24A, the recognition mode setting window screen displayed on the display unit 120 for adjusting the three types of modes and adjusting the recognition speed and the recognition accuracy is switched between the three types of modes of high accuracy, standard, and high speed. Radio button 990 for designating, an arbitrary designation button 991 for arbitrarily designating accuracy by inputting a numerical value in the range of 0 to 100, and adjustment of accuracy in the range of 0 to 100 by moving left and right using a mouse or the like The slider bar 992 is displayed, and for example, when the radio button 990 high-speed button is clicked according to the business, it is displayed as shown in FIG. At this time, it is displayed as shown in FIG.

  In addition to alphanumeric characters and symbols, the characters to be read by the character recognition technology of the present embodiment include Japanese kanji, Chinese kanji (simplified characters), and Taiwanese kanji (traditional characters). The left side, the right side, the right side, the upper side, the lower side leg, and the lower side leg. (Kamae), and a plurality of regions in which the arrangement of drooping (position) that hangs down from the upper side to the lower side, and a ridge that is positioned along the lower side from the left side is predetermined. In this application, an area such as “bias” or “旁” is referred to as a divided area, a partial kanji image in the divided area is referred to as a kanji element, and an arrangement position such as “bias” or “旁” is defined. It is called the area position, and the composition of the Chinese characters with the same area position of the divided areas (for example, “Mine” “Bee” 2 divided by Kanji configured longitudinally as kanji configured to be divided into three parts in the vertical direction as "Se", "dialect", "Akira" vertically divided into two parts by Kanji construction, etc.) that the Chinese character configuration as.

[Overall operation]
As shown in FIG. 2, the outline of the character recognition processing by the calculation unit 150 of the character recognition apparatus 100 according to the present embodiment is to optically scan and input characters entered in a form or the like under the control of the calculation unit. Step S202 for extracting the feature amount of the character image input in Step S201, Step S203 for inputting the recognition mode (high speed / standard / high accuracy mode) selected by the operation of the operator, and Step S201 Step S205 for creating an isomorphic character table for grouping the character images to be read based on the character image inputted in step S205 and storing it in the character identification dictionary 165, and the feature amount of the character image inputted in step S201 and step S203. With the selected recognition mode as input, refer to the character identification dictionary 165 Character recognition is performed by calculating a similarity dc (x) based on a formula of Euclidean distance and Mahalanobis distance for the recognition character output candidate determined in step S204, and determining a recognition character output candidate. By executing step S206 and step S207 of outputting candidate characters in ascending order of similarity dc (x) calculated in step S206, the operation is performed to perform character identification.
The processing for determining the recognition character output candidates in step S204 described above differs in processing method depending on the high-speed mode, the standard mode, and the high-accuracy mode, and each processing content will be described next.

[Standard mode output candidate decision processing]
In step S204 of the standard mode output candidate determination process, as shown in FIG. 4A, in order to narrow down the number of candidates, it is determined whether the character image is a kanji character or a non-kanji character (number, alphabetic character, symbol, kana character, etc.). Following step S402, a step S403 for calculating a Euclidean distance value to narrow down a character type having a large Euclidean distance value from candidates, and a character type having a low cosine similarity (close to -1) following the step S402. Step S404 for calculating the cosine similarity for each predetermined NXN divided region of the character image in order to exclude (candidate character) from the candidate, and comparing the cosine similarity with a predetermined value after step S404 In step S405, a group to which a large character type belongs is selected from the isomorphic character table, and in step S403, Based on the candidate character excluding the candidate character excluding the recognition candidate having a large Euclidean distance value and the character type (candidate character) having a small cosine similarity determined for each divided region based on the specified value (NXN) in step S405, (1 (2) a condition for selecting a candidate from the result of the kanji / non-kanji determination unit, (2) a condition for selecting a candidate with a small Euclidean distance value, (3) a condition for selecting a candidate with a high cosine similarity, and (4) Step S406 for determining a recognized character candidate by performing a logical OR with a condition for selecting a group to which a character type having a high cosine similarity belongs from the isomorphic character classification table is performed, and the Euclidean distance value and the cosine distance value are set as factors. It operates so as to output the character image as an output candidate.

  Further, in the isomorphic character / typical character candidate selection process in step S405, a group to which a character type having a high cosine similarity belongs is selected from the isomorphic character classification table shown in FIG. 4B, and belongs to the selected groove. The character type operates to be added to the recognition candidate, and details will be described later. In FIG. 4, other circles in the isomorphic character table are external characters in the electronic application software, and are therefore expressed in parentheses in this specification.

[High-speed mode output candidate decision processing]
Step S204 of the high-speed mode output candidate determination process, as shown in FIG. 3, is a step S321 for determining whether the character image is a Kanji character or a non-Kanji character (number, alphabetic character, symbol, kana character, etc.) in order to narrow down the number of candidates. Subsequent to step S321, step S322 for calculating a Euclidean distance value to narrow down the character types having a large Euclidean distance value from candidates, and recognition candidate characters obtained by excluding character types having a large Euclidean distance value from the candidates in step S322 (1) Judgment of Kanji / Non-Kanji Character according to the determination result of Step S321 and selection of candidate characters as Kanji or non-Kanji (2) Selection of candidate characters whose Euclidean distance value is smaller than a predetermined threshold The cosine similarity is obtained by executing step S323 of determining a recognized character candidate based on the logical sum of the conditions to be Since the exception of the determination by, it operates to output the high-speed candidate characters those accuracy is lower than that determined by the standard mode as an output candidate.

[High-precision mode output candidate decision processing]
In step S204 of the high-accuracy mode output candidate determination process, as shown in FIG. 5A, in order to narrow down the number of candidates, it is determined whether the character image is a kanji character or a non-kanji character (number, alphabetic character, symbol, kana character, etc.). Step S504 for calculating the Euclidean distance value for narrowing down the character types having a large Euclidean distance value from the candidates following Step S502, Step S503 for determining the divided region of the character image following Step S502, Subsequent to step S503, in order to exclude a character type (candidate character) having a small cosine similarity (close to -1) from candidates, a cosine similarity is calculated for each region obtained by dividing the character image in step S503. And a group to which a character type having a relatively high cosine similarity belongs after step S506. Step S506 selected from the glyph classification table, candidate characters excluding recognition candidates having a large Euclidean distance value in Step S507, and candidate characters excluding character types (candidate characters) having a low cosine similarity for each region in Step S506 (1) Conditions for selecting candidates from the results of the kanji / non-kanji determination unit, (2) Conditions for selecting candidates with small Euclidean distance values shown in FIG. 5B, and (3) High cosine similarity From the logical sum of the conditions for selecting a candidate and (4) the condition for selecting a group to which a character type with a high cosine similarity belongs from the isomorphic character classification table, (5) does not match the basic structure obtained by the divided area determination processing Euclidean by performing a negative OR to remove the character type (heterogeneous candidate character) and the character type of the group to which the character type belongs from the candidate The distance value and the character image are divided based on the kanji composition in addition to the Euclidean distance value by operating to output the character image as an output candidate with the cosine distance value of each divided area such as the radical of the kanji as a factor. Since the determination based on the cosine similarity for each divided area is performed, the recognition speed is lower than that in the determination in the standard mode, but the operation is performed so that highly accurate candidate characters are output as output candidates.

[Homomorphic / classical character candidate selection processing]
The details of the isomorphic / typical character candidate selection processing in step S405 are as follows. As shown in FIG. 6, the correct character is “凋”, and the handwritten character is “Nice” and “旁” ) When the “circumference” is in contact, the divisional element discrimination process determines the type and position of the radical of the input pattern (character image), thereby dividing the vertical and horizontal divisions into two elements. Step S601 for determining the composition of the kanji composed, and cosine similarity and Euclidean value similarity are obtained for each divided region divided by step S601, and the first candidate and second place in order from the left in the figure in ascending order of similarity. In step S601, which is assumed to be a candidate, third place candidate, and the character “凋” having the same region position as the isomorphic character table (3) to which the first candidate character assumed in step S601 belongs is the same. Step S603 for adding the character type of the isomorphic character table (2) to the candidate and adding the character types of the isomorphic character table (1) and (5) to the candidate, and the third character has a high degree of similarity, ) And 旁 (construction) (because the composition of Chinese characters is different), the group to which a character type having a high cosine similarity / Euclidean value similarity belongs is executed by executing step S604 that is excluded from candidates. The character type belonging to the selected groove is selected from the isomorphic character classification table, and operates so as to be added to the recognition candidate.

[Division area determination processing 1]
As shown in FIG. 7, the details of the character image segmentation area determination process in step S503 in the output candidate determination process in the high-accuracy mode described above are as follows. (Hen) In the case of a vertically divided Chinese character structure in which “Nisui” and “Tsubaki” (Tsukuri) are in contact with each other, an A core pattern that extracts the core of the character component of the input pattern (character image); A character line segment pattern in the B vertical direction extracted from the character vertical direction component, a character line segment pattern in the C horizontal straight direction extracted from the character horizontal direction component, and a D line segment approximate pattern extracted from the circumscribed line of the character image, Generating step S701, obtaining a horizontal and vertical histogram from the generated pattern ABC, and dividing the six vertical and horizontal histograms and pattern D obtained in step S702. By executing the step S703 for determining the position to determine an area division position of the Kanji characters.

  In the process of obtaining the division position in step S703, in the case of the illustrated character graphic, since there are two peaks (mountains) in the vertical histograms A-1 and B-1, the valley position between the two peaks is divided. Candidates a and b, division position candidates c and d based on the E-line circumscribed approximation pattern, horizontal histograms A-2, B-2, C-2, and histogram C-1 have small differences between peaks and valleys, This is performed by excluding the candidate from the division position candidates due to being one and determining the division position by combining them.

[Division area determination processing 2]
The divided region determination process in step S503 is not limited to the above process, and the A core line pattern obtained by extracting the core line of the character component of the input pattern (character image) and the B vertical direction obtained by extracting the character vertical direction component. A character line segment pattern and a D-line segment approximated pattern obtained by extracting a circumscribing line of the character image are generated. From the generated horizontal histogram A-1 of pattern A and horizontal histogram B-1 of pattern B. Even if the division position candidates a and b are extracted and the division position candidates a and b are extracted from the circumscribed approximate pattern D, a pattern having a similar basic configuration of the divided area is selected from the isomorphic character table and added. good.

  In this technique, for example, as shown in FIG. 8, it is determined that the divided area is divided into right and left by the dividing positions a to d, and an isomorphic character table (2) having a kanji structure in which kanji elements are divided into two vertical columns. And (3) are extracted, the character types of the isomorphic character tables (2) and (3) are added as candidates, and excluded from the isomorphic character table (4) and later after which the kanji composition is greatly different, such as top and bottom, enclosure, etc. Is done by.

[Specific description of division position candidate determination]
In the case of the horizontal histogram shown in FIG. 11A, a specific method for determining the division position candidate from the horizontal histogram A-1 described with reference to FIG. 7 described above is a mountain (pixel) higher than the upper threshold value α of the horizontal histogram. The first step of setting a position having a large number of peaks as a mountain a, a mountain b, a mountain c, and a mountain d, and among the plurality of mountains set by the first step, the lowermost threshold β is between adjacent mountains. From the second step of setting the low position as the valley a, the third step of setting the position lower than the lower threshold β as the valley b between other peaks, and the valley position set by the second step and the third step, 1 The fourth step of selecting one and whether or not the valley selected in the fourth step is valid as the decomposition position, Y _{x is} the minimum coordinate of the valley shown in FIG. 11C, and X is the center of the coordinate Y _x. x -1 position coordinates apart around the _{Y x-1,} the coordinate _{Y x} When a distant position coordinates and the _{Y x + 1,} a fifth step according to whether or not satisfies the equation conditions _{"Y x <Y x-1} and _{Y x <Y} x _{+ 1",} the fifth step When the valley a is determined as the division candidate, the division position candidate is determined by repeating the sixth step of setting the valley a as the division candidate. Incidentally, it is judged by the determination of whether a mountain in the first step, and Equation conditions shown in FIG. 11 (b) _{"Y x-1 <Y x} and _{Y x} + 1 _{<Y x".}

[Uniform character table creation process]
As shown in FIG. 9, the outline process of step S205 for creating the isomorphic character table is as follows. Character images are divided into large groups (no division, division area divided in the vertical direction, division in the horizontal direction) for each pattern having a similar kanji structure. A divided region, a divided region divided into a plurality of + character shapes, and other divided shapes), and a character type having a small Euclidean distance value or a character type having a high cosine similarity in the group roughly classified in step S901. Step S902 for dividing the group into small groups, and step S903 for storing the information divided into large groups according to the region structure in step S901 and the information divided into small groups based on the Euclidean distance value and cosine similarity in step S902 in the isomorphic character table. To create an isomorphic table by running To work. For example, in the case of a kanji composition vertically divided into left and right by this step S902, for example, a kanji image having a high degree of similarity between the left side kanji element and the right side kanji element. It operates to divide into small groups from kanji images with high similarity.

  That is, the isomorphic character table creation processing according to the present embodiment divides a kanji image into large groups based on the division shape according to the kanji composition rule, and further, the kanji image having a high similarity between the kanji elements in the same divided region position in the same large group The isomorphic character table is created by dividing into small groups.

  As shown in FIG. 10, this homomorphic character table creation process registration process to a large group includes step S1001 for reading a kanji image to be recognized from the homomorphic character table creation database, and a cord from the kanji image read in step S1001. Step S1002 for creating a pattern, Step S1003 for creating a horizontal history from the core pattern created in Step S1002, and a valley position between adjacent peak positions (mountains) from the horizontal history created in Step S1003 Step S1004 for obtaining the position, Step S1005 for selecting one valley position from the valley positions obtained in Step S1004, and whether or not the valley position selected in Step S1005 is a division position candidate (the magnitude of the distance between the peaks and valleys) Etc.) Step S1006, step S1007 that excludes the division position candidate selected in step S1005 when it is determined that it is not a division position candidate in step S1006, and division position candidate that is determined in step S1006 that it is a division position candidate As step S1009, and whether or not all valley positions determined in step S1004 have been determined. If it is determined that the determination has not been completed, step S1009 returns to step S1006, and step S1009. When it is determined that all valley positions have been determined in step S1010, the division candidate positions are determined based on the division position candidates stored in step S1008, and the division area determined in step S1010 is the same character table. Large guru Step S1011 for determining whether or not it has already been registered in step S1011, and step S1011 for registering the new major classification in the isomorphic character table and the character type read in step S1001 in the major classification When it is determined in S1012, that it is registered in Step S1011 above, Step S1013 in which the character type read in Step S1001 is added to the corresponding major classification already registered in the isomorphic character table, and all character images are large. It is determined whether or not the table registration to the group has been completed, and when it is determined that the registration of all the character images to the large group has not been completed, step S1014 returns to step S1001, and in step S1014 all of the character images are registered. Text table for large groups When it is determined that the recording has been completed, the processing is ended.

[Large group creation processing]
The large grouping process roughly classified according to the area structure in step S901 of FIG. 9 is performed by the above-described divided area determination process in the case of the Chinese character image (image pattern) “凋” shown on the left side of FIG. (B) When the vertical division area shown on the right side is acquired, and the acquired vertical division area is registered in advance in the isomorphic character table but the character type is not registered (FIG. 12 [b] left side) ), The Chinese character image (image pattern) “凋” is registered in the # 2 group which is the vertical division area (right side of FIG. 12B), and the obtained vertical division area is registered in the isomorphic character table in advance. If not (FIG. 12 [c] left side), a large group # 2 of vertically divided areas is added and a Chinese character image (image pattern) “凋” is newly registered in the added large group # 2 (FIG. 12). [C] Right) To work.

[Similar area judgment processing]
The processing of grouping character types having a small Euclidean distance value or character types having a large cosine similarity into a small group in the large group in step S902 of FIG. 9 is performed by the calculation unit as shown in FIG. Step S1301 for selecting one groove from the large groups registered in (left side of FIG. 14), Step S1302 for reading one character image from the large group selected in Step S1301, and the character image read in Step S1302 Step S1303 for selecting one Kanji element from the above, and calculating the similarity between the Kanji element selected in Step S1303 and the Kanji element at the same region position in another Kanji image using the Euclidean distance value and the cosine similarity. In the similarity calculation result table (right side of FIG. 14) To run and step S1304 to be.

  For example, as shown in FIG. 14, the similarity is calculated in steps S1301 to S1304 by selecting a large group of the # 2 vertical division format, and selecting the left region (deviation [hen] of any kanji image belonging to this large group. ])), The degree of similarity between the kanji elements in the same area position (left biased position) of other kanji in the same large group is calculated, and small groups are divided according to the degree of similarity. As shown on the right side of FIG. 14, for each character type in the large group of # 2, the degree of similarity is determined according to the division position of the calculation target area (for example, center vertical division, left vertical division, right vertical division, etc.). Can be divided into small grooves. The similarity is set within a range of 0 to 1, and the larger the value is, the more similar it is.

  Next, the process of grouping character types having a high degree of similarity back into a small group returns to FIG. 13 and the calculation of the degree of similarity between Chinese character elements (divided regions) in all character images in the large group selected in step S1301 is completed. In step S1305, the process returns to step S1304 when it is determined that the process has not been completed. In step S1305, the similarity calculation of kanji elements (divided areas) for all character images in the same large group is performed. When it is determined that the processing has been completed, it is determined whether or not the similarity calculation of the kanji elements (divided regions) other than those selected in step S1303 has been completed. When it is determined that the processing has not been completed, the process returns to step S1303. In step S1306, in step S1306, it is determined that the similarity calculation for all the kanji elements (divided areas) has been completed. Step S1307 returns to Step S1302 when it is determined whether the similarity calculation of all the character images in the large group selected in Step S1301 has been completed. When it is determined in S1307 that the similarity calculation for all the Kanji elements (divided areas) in the same large group has been completed, it is determined whether the similarity calculation for all large groups has been completed, and it is determined that the calculation has not been completed. If it is determined in step S1308 that the process returns to step S1301 and the similarity calculation for all large groups has been completed in step S1308, classification is made into small group candidates based on the similarity calculation result calculated in step S1304. Step S1309 and the small group classified in Step S1309 Step S1310 for selecting one group and calculating the degree of similarity with another small group, Step S1311 for determining whether the degree of similarity calculated in Step S1310 is equal to or greater than a predetermined threshold, and the similarity calculated from Step S1311 When it is determined that the degree is not equal to or greater than the predetermined threshold, the small group candidate is not connected, and when the similarity is determined to be equal to or greater than the predetermined threshold in Step S1311, the small group whose threshold is equal to or greater than the predetermined threshold is selected. Step S1312, which is combined with one small group and stored in the small group candidate classification table (FIG. 15), determines whether the calculation of the similarity of all the small groups classified in step S1309 has been completed. When it is determined that the process has not been performed, the process returns to step S1310. By executing step S1313 that ends the processing, an isomorphic character table (FIG. 15) in which small groups having the similarity of the Kanji element at the specific region position equal to or higher than a predetermined value is created (for example, vertically divided into two parts). A small group with similar left side bias and a small group with similar vertical wrinkles on the right side of the vertical split).

[Homomorphic character table]
As shown in FIG. 16A, the isomorphic character table created in step S205 includes a large group representing a rough division format (no division, vertical division, cross-cutting, etc.) of kanji characters that are divided areas, and a large group. Items of a small group divided by detailed division positions (for example, center vertical division, left vertical division, right vertical division, etc.), division position allowable range of the small group, and character types included in the small group As shown in FIG. 16 (b), the division position allowable range includes an X axis in the horizontal direction and a Y axis in the vertical direction, and the character image has a size of nx, ny, as shown in FIG. If the allowable range of the X-axis direction and / or the Y-axis method is expressed in dot units and the division area cannot be determined, the character image is equally divided by NXN as shown in FIG. , Similarity in the same area Candidate character types can be selected by comparing a small group with a region with a small Euclidean distance value or a large cosine similarity, and added to a recognition candidate because the divided region format (structure) is similar Can do.

[Summary]
In this way, the character recognition device to which the character recognition method according to the present embodiment is applied, when plotting the Euclidean distance values of all classes and the character images to be calculated by the character recognition technology using only the Euclidean distance values as in the prior art, As shown in FIG. 19, the class is roughly divided into a class having a high possibility of a correct answer indicated by a circle, a class having a medium possibility of being indicated by a hatched circle, and a class having a low possibility of a correct answer indicated by a circle. Classes with a high probability of correct answer indicated by a circle above the threshold T may not be calculated and may be missed, and conversely, a class with a moderate probability of correct answer indicated by the diagonal line below the threshold T is also calculated. While there was a problem that there was a possibility of misreading as a target, in the present invention, by using an isomorphic character table in which the similarity of kanji elements at a specific region position is determined, Class can be included in the calculation target, and the class with the possibility of erroneous recognition indicated by the slash ◯ mark below the threshold T can be excluded from the calculation target. be able to.

  In addition, as shown in FIG. 20, the character recognition apparatus to which the character recognition method according to the present embodiment is applied plots class A patterns (character images) with circles and class B patterns (character images) with X marks. The input pattern (character image) is represented by Δ, and the Euclidean distance value between the input pattern (character image) represented by Δ and the average pattern of class A and class B (large ○ mark, large X mark) is the same. In this case, in the conventional technique, since it has to be compared with both the class A and class B patterns, the processing time becomes redundant. On the other hand, the cosine similarity is also used as in this embodiment. Since the cosine similarity angle α for class A is larger than the cosine similarity angle β for class B, character recognition can be performed at high speed by excluding class A from the comparison target.

  This character recognition processing time is a processing time when the recognition degree is set to high accuracy as shown in FIG. 21 in which the processing time for the number of output candidates to be compared at the time of recognition is represented by a bar graph / recognition accuracy by a broken (dashed) line graph. If the recognition speed is increased when the recognition time is increased, the recognition speed decreases. Therefore, by narrowing down the number of output candidates as in this embodiment, a large memory size is not required. , High-speed character recognition can be performed.

100 character recognition device, 110 input unit, 120 display unit, 130 image acquisition unit,
140 communication unit, 150 calculation unit, 160 external storage unit, 165 dictionary for character identification

Claims (10)

And input unit for reading in optical scan characters including Chinese characters consisting of Chinese elements a plurality of areas as a character image, an image acquisition unit for acquiring a character image scanned by using the input unit, a plurality of standard character image A character identification dictionary storing characters, an external storage unit for classifying kanji characters by a common radical, 旁, etc., and storing an isomorphic character table for classifying similar and isomorphic kanji characters visually, and the character identification A character recognition unit that performs character recognition processing with reference to the character identification dictionary with respect to the input character image. A device,
The computing unit is
A divided region determination step of determining a divided region for dividing the character image in units of one character scanned by the image acquisition unit for each Chinese character element;
A cosine similarity calculation step of calculating the cosine similarity of the kanji elements in the same region position of the character image determined by the divided region determination step;
A Euclidean distance value calculating step of calculating a Euclidean distance value of a Chinese character element at the same region position of the character image determined by the divided region determining step;
Character images having the same kanji composition based on the divided areas determined in the divided area determining step and having a similarity between kanji elements in the same region position based on the cosine similarity calculated in the cosine similarity calculating step higher than a predetermined value And an isomorphic character table creating step for storing character images whose Euclidean distance values between Kanji elements calculated by the Euclidean distance value calculating step are smaller than a predetermined value and grouped as isomorphic characters with high similarity,
A character image whose cosine similarity calculated by the cosine similarity calculation step is larger than a predetermined cosine similarity, a character image whose Euclidean distance value calculated by the Euclidean distance value calculation step is smaller than a predetermined value, and creation of the isomorphic character table A standard character recognition step of performing character recognition using a character image of a logical sum with a character image of isomorphic characters obtained by grouping character images with high similarity in the same region position by the process as a candidate character Recognition device.   Characters included in a group to which the different candidate characters belong and different candidate characters having different configurations of a plurality of divided regions divided by the divided region determination step from the character image obtained by the logical sum in the standard character recognition step. The character recognition apparatus according to claim 1, further comprising: a high-definition character recognition step of performing character recognition based on candidate characters excluding images. 3. The high-speed character recognition step of performing character recognition using a character image whose Euclidean distance value calculated by the Euclidean distance value calculating step is smaller than a predetermined value as a candidate character. Character recognition device. In the isomorphic character table creating step, the computing unit divides character images having the same kanji composition based on the divided regions determined in the divided region determining step into large groups, and groups by the large grouping step. A small grouping step of dividing the character image and the character image and the Euclidean distance value smaller than the predetermined value into small groups, wherein the cosine similarity of the kanji element at a specific region position in the divided large groove is higher than a predetermined value,
4. The character recognition apparatus according to claim 1, wherein the kanji images divided into large groups and small groups are stored in an isomorphic character table.   5. The character recognition device according to claim 3, wherein the arithmetic unit has a function of switching or adjusting the standard character recognition step, the high-precision character recognition step, and the high-speed character recognition step. And input unit for reading in optical scan characters including Chinese characters consisting of Chinese elements a plurality of areas as a character image, an image acquisition unit for acquiring a character image scanned by using the input unit, a plurality of standard character image A character identification dictionary storing characters, an external storage unit for classifying kanji characters by a common radical, 旁, etc., and storing an isomorphic character table for classifying similar and isomorphic kanji characters visually, and the character identification A character recognition method for recognizing a character, comprising an arithmetic unit that controls character recognition of a character image with reference to a dictionary and an isomorphic character table,
In the calculation unit,
A divided region determination step of determining a divided region for dividing the character image in units of one character scanned by the image acquisition unit for each Chinese character element;
A cosine similarity calculation step of calculating the cosine similarity of the kanji elements in the same region position of the character image determined by the divided region determination step;
A Euclidean distance value calculating step of calculating a Euclidean distance value of a Chinese character element at the same region position of the character image determined by the divided region determining step;
Character images having the same kanji composition based on the divided areas determined in the divided area determining step and having a similarity between kanji elements in the same region position based on the cosine similarity calculated in the cosine similarity calculating step higher than a predetermined value And an isomorphic character table creating step for storing character images whose Euclidean distance values between Kanji elements calculated by the Euclidean distance value calculating step are smaller than a predetermined value and grouped as isomorphic characters with high similarity,
A character image whose cosine similarity calculated by the cosine similarity calculation step is larger than a predetermined cosine similarity, a character image whose Euclidean distance value calculated by the Euclidean distance value calculation step is smaller than a predetermined value, and creation of the isomorphic character table And a standard character recognition step of performing character recognition using a character image of a logical sum with a character image of isomorphic characters obtained by grouping character images having high similarity in the same region position by a process as a candidate character. Recognition method.   Characters included in a group to which the heterogeneous candidate characters belong to different candidate characters having different configurations of a plurality of divided areas divided by the divided area determination step from a character image obtained by logical sum in the standard character recognition step The character recognition method according to claim 6, further comprising: performing a high-definition character recognition step of performing character recognition based on candidate characters excluding images.   8. The high-speed character recognition step of performing character recognition using a character image whose Euclidean distance value calculated by the Euclidean distance value calculating step is smaller than a predetermined value as a candidate character is executed by the computing unit. Character recognition method.   In the isomorphic character table creating step, the arithmetic unit has a large grouping step in which character images having the same kanji composition based on the divided regions determined in the divided region determining step are divided into large groups, and the large grouping step A small grouping step of dividing a character image and a character image having a Euclidean distance value smaller than a predetermined value as a small group, wherein the cosine similarity of the Kanji element at a specific region position in the divided large groove is higher than a predetermined value, 9. The character recognition method according to claim 6, wherein the kanji images divided into large groups and small groups are stored in an isomorphic character table.   The character recognition method according to claim 8 or 9, wherein the arithmetic unit is provided with a function of switching or adjusting the standard character recognition step, the high-precision character recognition step, and the high-speed character recognition step.