JPH11238099A - Character recognition device, method therefor and computer readable recording medium stored with character recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a character recognition device capable of correctly recognizing characters even when the character images of a recognition object are crushed. SOLUTION: A character image input part 302 inputs normalized binarized recognition object character images and an image area division part 303 divides the existence area of the character images into 16. A feature amount extraction part 304 extracts a feature amount for the respective divided areas and an image area classification part 305 classifies the respective areas into two groups. A recognition part 306 calculates the street distance of the feature amount of the recognition object character image and the standard feature amount of a recognition dictionary 301 separately for a first group and a second group, multiplies a prescribed coefficient by the respective street distances of the first group and the second groups, totals them and defines a closest standard character as a recognized result.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition apparatus and method for recognizing characters from an input character image to be recognized.

[0002]

2. Description of the Related Art In order to improve the accuracy of character recognition, there is a character recognition device described in Japanese Patent Application Laid-Open No. 61-125688. This apparatus has a recognition dictionary in which a standardized standard character image is divided into 16 regions each having a length and width, and a standard feature amount and a character code for each region are registered. This standard feature amount is obtained by classifying patterns of arrangement of black pixel portions representing character element portions into four directions of horizontal / upper left / vertical / upper right, and calculating the numbers of the four patterns of the character element portions included in the area. Things. Therefore, the recognition dictionary contains:
A four-dimensional standard feature amount of each area is described for each character code, and a “4 × 16” 64-dimensional standard feature amount is registered.

In this apparatus, a character image to be recognized is input as a binary image having a character element portion of "black" and a background portion of "white", and is normalized to an area of "16.times.16" pixels. The image is divided into 16 regions each having 4 pixels, and the same feature amount as in the recognition dictionary is extracted in each region. Next, as a similarity between the character image to be recognized and the character indicated by the character code registered in the recognition dictionary, a standard feature amount (64-dimensional)
Is calculated with respect to the extracted feature amount (similar 64 dimensions). Since the smaller the calculated city distance is, the higher the similarity is, the input character image is recognized as a character represented by a character code having a standard feature amount that minimizes the city distance.

[0004] By dividing the area in this manner, for example, when the number of four patterns of “horizontal / upper left / vertical / upper right” is used as the standard feature amount without performing division, “soil” which cannot be correctly recognized is used. And "shi" can be distinguished and recognized.

[0005]

However, in such a conventional character recognition apparatus, when a character image is input as a binary image, if the input character image is destroyed due to, for example, the resolution of a scanner, etc. No matter how much the area is subdivided, the number of feature elements is changed from 4 to 8, and the accuracy of the feature amount is increased, and the accuracy of the similarity calculation is increased. I can't recognize.

A specific example will be described. FIGS. 1 and 2 are binary character images of the character “den” both normalized. The character image “den” 101 shown in FIG. 1 is correctly read by the scanner, but the character image “den” 201 shown in FIG. 2 is in a state where the characters are crushed when read by the scanner. . When the two character images “den” 101 and 201 are recognized by a conventional character recognition device, the character image “den” 101 is correctly recognized as the character “den”.
Is erroneously recognized as the letter "t".

SUMMARY OF THE INVENTION It is an object of the present invention to provide a character recognition device with improved recognition accuracy capable of correctly recognizing a character even when a character image to be recognized is crushed.

[0008]

According to the present invention, there is provided a character recognition apparatus for recognizing a character image to be recognized as a character code.
(N ≧ 2) is divided into partial image areas, and a recognition dictionary in which standard features of partial character images included in each partial image area are registered in advance along with character codes in character units, and a recognition target character image is divided into N parts Image region dividing means for dividing into image regions;
A feature extracting unit for extracting a feature of a partial character image included in the partial image region divided by the image region dividing unit; and a partial image region classification for classifying the partial image regions divided by the image region dividing unit into a plurality of groups. Means, for each group, calculate the elementary similarity between the feature of the partial character image of the recognition target character image and the corresponding standard feature of the recognition dictionary, weighting the elementary similarity of each group a predetermined weight, Similarity calculation means for calculating the similarity between the recognition target character image and the characters in the recognition dictionary,
A recognition character selecting means for selecting a character having the highest similarity as a recognition character and selecting a character code thereof.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the character recognition device according to the present invention will be described below with reference to the drawings. (Embodiment 1) FIG. 3 is a configuration diagram of Embodiment 1 of a character recognition device according to the present invention. The character recognition device includes a recognition dictionary 301, a character image input unit 302, an image region division unit 303, a feature amount extraction unit 304, an image region classification unit 305, a recognition unit 306, and a recognition result output unit 307. It has.

[0010] The recognition dictionary 301 pre-registers a character code and a standard feature amount indicating the characteristics of a standard character image of each character. FIG. 4A is a diagram illustrating an example of registered contents of the recognition dictionary 301. The registration content 401 includes a character code 402 and a standard feature amount 403 for the number “1”.
And FIG. 4B shows a standard character image of the numeral “1” from which the standard feature amount 403 is extracted.

In the character image 404, a character element portion is indicated by "black" pixels and a background portion is indicated by "white" pixels in a normalized area of 16 pixels in each of the vertical and horizontal directions. This character image 40
4 to extract the standard feature amount 403 of the character image 404.
Is divided into N regions (N = 16 in this case). In the figure, as indicated by broken lines, the area is “16” divided equally into “4” pixels in each of the vertical and horizontal directions.

The standard feature quantity 403 is a feature element which is a pattern of arrangement of “black” pixels at a boundary point between “black” pixels and “white” pixels of the character image 404 as shown in FIG. The number of coincidences is counted for each divided area. The characteristic element “horizontal” 501 is indicated by patterns 502 and 503 in which “black” pixels are arranged in the horizontal direction. “*” Attached to the pixel indicates that the boundary point between the character element portion and the background portion is set as the target pixel. The characteristic element “upper left” 504 is indicated by patterns 505 and 506 in which black pixels are arranged in the upper left direction. Where "◎" is "black"
This indicates that it may be a pixel or a "white" pixel. The characteristic element “vertical” 507 is indicated by patterns 508 and 509 in which “black” pixels are arranged in the vertical direction. The characteristic element “upper right” 510 is indicated by patterns 511 and 512 in which “black” pixels are arranged in the upper right direction.

In the area 405 of FIG.
When the number of appearances of 504, 507, 510 is counted, "horizontal"
501, a pattern 502 in which the “black” pixel 406 is a target pixel and a pattern 503 in which the “black” pixel 407 is a target pixel appear once, and the standard feature amount of the feature element “horizontal” 501 is It becomes "2". When the “white” pixel 408 in the area 405 is set as a target pixel, “upper left”
Pattern 505 appears once. Thus, the standard feature amount of the feature element “upper left” 504 becomes “1”. The pattern 508 of how to arrange the characteristic elements "vertical" 507,
Since 509 does not appear in the area 405, the standard feature amount is “0”. Also, “white” pixels 409 in the area 405,
The pattern 51 of how to line up each with 410 as a pixel of interest
1 appears twice. Thereby, the characteristic element “upper right” 51
The standard feature value of 0 is “2”. The standard feature amount (corresponding to the area 405) specified by <horizontal 1> and <vertical 1> of the standard feature amount 403 in FIG. 4A is “2, 1, 0, 2”.

Similarly, in the area 411, the "black" pixel 41
The pattern 502 in the arrangement manner appears four times with 2,413,414,415 as the pixel of interest. Thus, the standard feature amount of the feature element “horizontal” 501 becomes “4”. The standard feature amount (corresponding to the area 411) specified by <horizontal 1> and <vertical 2> of the standard feature amount 403 in FIG.
0 ". Note that the upper side 416 of the region 411 is “white”
The standard feature amount is counted as a pixel.

In the remaining 14 areas, standard feature values are similarly counted. Thus, the standard feature amount 403 is
The number of appearances of four types of characteristic elements “horizontal” 501, “upper left” 504, “vertical” 507, and “upper right” 509 are counted for each of the divided “16” regions from the pattern of the arrangement of “black” pixels. Yes, the character code 402 is “4 × 1
6 ", which is a 64-dimensional quantity.

FIGS. 6 to 12 show standard character images corresponding to other examples of the registered contents of the recognition dictionary 301, respectively.
FIG. 6A shows registered contents 601 of the alphabet “i”, and FIG. 6B shows a character image 602 thereof. The registration content 601 includes a character code 603 and a standard feature amount 604 obtained by dividing the area of the character image 602 into “16”, like the numeral “1” shown in FIG.

Similarly, FIG. 7 shows the alphabet "t", FIG. 8 shows the kanji "cloud", FIG. 9 shows the kanji "rai", FIG. 10 shows the kanji "home", FIG. 11 shows the kanji "den", and FIG. The registered content of "dragon" and a standard character image are shown. The character image input unit 302 includes a scanner or the like, and inputs a binary image in which a character image part to be recognized is a pixel element in “black” pixels and a background part is “white” pixels, and “16 × 16” Normalize to pixels. The character image to be recognized is the character image "den" 101, shown in FIGS.
This is the same as “Den” 201.

The image area dividing section 303 converts the character image input by the character image input section 302 into N areas (here, N = N).
16), that is, the image is divided into 16 areas each of which is equal to four in the vertical and horizontal directions. FIG. 13 shows a character image obtained by dividing the character image 101 shown in FIG. 1 into 16 regions. The feature amount extraction unit 304 determines the pattern of the arrangement of “black” pixels “horizontal” 502 and 503, “upper left” 505 and 506, which are the characteristic elements shown in FIG.
“Vertical” 508 and 509 and “upper right” 511 and 512 are stored together with the target pixel (marked with “*” in the figure). For each area divided by the image area dividing unit 303 (for example, <horizontal 1><vertical1>,<horizontal1><
<Vertical 2>, <Horizontal 1><Vertical3>, ..., <Horizontal 4><Vertical 3
>, <Horizontal 4> and <vertical 4>), and if there is a pattern that matches the stored pattern of “black” pixels with the pixel marked with “*” at the boundary point as the pixel of interest, Are increased by “1” at a time.

By performing such processing, the characteristic amount extracting section 304 extracts a four-dimensional characteristic amount from each of the 16 regions, and extracts a 64-dimensional characteristic amount from the character image.
The extracted feature amount is stored. FIG. 14 shows a 64-dimensional feature quantity extracted from the character image "den" 101 shown in FIG. Similarly, the character image “den” 201 shown in FIG.
, A 64-dimensional feature quantity shown in FIG. 15 is extracted.

The image area classifying unit 305 classifies each of the N areas (N = 16) divided by the image area dividing unit 303 into a plurality of groups. Here, it is classified into two groups, a first group and a second group. FIG. 16 shows a state in which the respective regions of the character images 101 and 102 are classified into a first group 1601 in the peripheral portion and a second group 1602 in the central portion (hatched portion). Therefore, the first group 1601 includes the <horizontal 1>, <vertical 1> to <vertical 4>, <horizontal 2>, and <vertical 1> divided by the image area dividing unit 303.
>, <Horizontal 2><vertical4>,<horizontal3><vertical1>,<horizontal3><
Twelve areas of <vertical 4>, <horizontal 4>, <vertical 1> to <vertical 4> are classified. Similarly, the second group 1602 includes <horizontal 2><vertical 2
>, <Horizontal 2><vertical3>,<horizontal3><vertical2>,<horizontal3><
Four regions of vertical 3> are classified.

When the classification of each image region is completed, the image region classification unit 305 notifies the recognition unit 306 of the image regions classified into the first group and the second group. When the recognition unit 306 receives the notification of each image region classified into the first group and the second group from the image region classification unit 305, the feature amount extraction unit 30
4 is read out and the recognition dictionary 301 is read out.
For all the character codes registered in, the distance between the city area and the standard feature amount corresponding to each character code is calculated.

The recognizing unit 306 is provided with a first group 1601 and a second group 160 as shown in FIG.
When the notification that the second group is received, the distances d1 and d2 are calculated for the first group 1601 using the equation (1) and for the second group using the equation (2). Here, Foi is the i-th feature amount classified into the first group stored in the feature amount extraction unit 304, and Fsi is the recognition dictionary 3
01 is the corresponding i-th standard feature amount registered in 01. Here, Foj is the j-th feature amount classified into the second group stored in the feature amount extraction unit 304, and Fsj is the recognition dictionary 3
01 is the corresponding j-th standard feature amount registered in 01.

Next, the recognizing unit 306 multiplies the calculated city distance d1 of the character by α and multiplies the city distance d2 by β to add a total city distance d of each character. Recognition dictionary 3
The total city distance is calculated for all the characters stored in “01”, and the character having the smallest total city distance d is determined to have the highest similarity to the character image to be recognized, and is output to the recognition result output unit 307. Notify the character code of the character.

Now, assuming that α = 1.0 and β = 0.5,
When the total city distance d is calculated for the character image “den” 101 shown in FIG. 1, the result is as shown in FIG. As a result, the total city distance d of the character "den" is minimized to "14".
"5" (see FIG. 11) is notified to the recognition result output unit 307.

When the total city distance d is calculated for the character image "den" 201 shown in FIG. 2, the result shown in FIG. 17B is obtained. Thus, the recognition result output unit 307 is similarly notified of the character code “4545”. Even with the crushed character image “den” 201 shown in FIG. 2, the present apparatus can correctly recognize characters. Note that the character images “den” 101 and “den” 20 calculated by the above-described conventional device are used.
FIG. 17C shows the distance from the city center to the area 1. As a result, in the character image "den" 201,
It is clearly shown that it was not recognized correctly.

The recognition result output unit 307 has a display such as a CRT, and displays a standard character image corresponding to the character code notified from the recognition unit 306. Next, the operation of the present embodiment will be described with reference to the flowchart shown in FIG. The character image input unit 302 inputs the character image to be recognized as a normalized binary character image (S18).
02).

The image area dividing section 303 divides the character image existing area into N areas (S1804). The feature amount extraction unit 304 extracts a feature amount of the dimension of the number of feature elements from the character image for each of the divided regions (S1806).
The image area classifying unit 305 assigns each of the divided areas to a first group that is a peripheral part of the character image and a second group that is a central part of the character image.
It is classified into groups (S1808).

The recognizing unit 306 determines whether or not each feature extracted by the feature extracting unit 304 belongs to the first group of peripheral parts (S1810). When it is determined that the character belongs to the peripheral part, the city distance d1 with the standard feature amount of the character registered in the recognition dictionary 301 is calculated and multiplied by a coefficient 1.0 (S
1812), and the process moves to S1816. When it is determined that the character does not belong to the peripheral part, since it is the central part, the city distance d2 with the standard feature amount of the character registered in the recognition dictionary 301 is calculated and multiplied by a coefficient 0.5 (S1814).

Next, the recognizing unit 306 adds the value obtained by multiplying the city distance d1 of the peripheral part obtained in S1812 by 0.5 to the city distance d2 of the central part obtained in S1814 by 0.5 to obtain the total city distance. d is calculated, and the character code of the standard feature amount used and the total city distance d are stored (S1816).
Steps S1810 to S1816 are repeated until all characters registered in the recognition dictionary 301 have been collated (S181).
8). The character code with the smallest total city distance d is selected as a character of the character image to be recognized, and is notified to the recognition result output unit 307 (S1820).

The recognition result output unit 307 stores the character code, displays a standard character image corresponding to the character code, and ends the processing. In the present embodiment, the first group 16
01 and the second group 1602 are classified into the peripheral part and the central part of the character image. However, as another classification method, the character group is classified into the left part and the right part of the character image. Alternatively, classification may be made between the upper part and the lower part, and attention may be paid to the “nearby” of the character component to improve the accuracy of character recognition. (Embodiment 2) FIG. 19 is a configuration diagram of a character recognition apparatus according to Embodiment 2 of the present invention. This character recognition device
Recognition dictionary 301, character image input unit 302, image region division unit 303, feature amount extraction unit 304, image region classification unit 305, candidate character recognition unit 1901, recognition unit 190
2 and a recognition result output unit 307. The same components as those of the character recognition device according to the first embodiment are denoted by the same reference numerals, and the description thereof will be omitted. Only the components unique to the present embodiment will be described.

When the image area classifying section 305 classifies the area into the peripheral area of the first group and the central area of the second group, the candidate character recognizing section 1901 determines the feature amount of the peripheral area of the first group and the recognition dictionary. 30
The city distance d1 between the character registered in No. 1 and the corresponding standard feature amount is calculated using Expression (1). FIG. 20 (a)
Indicates a character having a short city distance d1 between the character images "den" 101 and "den" 201, that is, a character having a high degree of similarity at the periphery of the character image. Note that the candidate character recognition unit 1901 stores a character code instead of the actual characters “den”, “lightning”, “cloud”,.

The candidate character recognizing unit 1901 includes the recognition dictionary 30
When the calculation of the city distance d1 with respect to the character image to be recognized is completed for all the characters registered in No. 1, the character code of the small city distance d1, for example, the upper three characters and the city distance d1 are notified to the recognizing unit 1902. I do. Recognition unit 19
02 has substantially the same configuration as the recognition unit 306 of the above embodiment, but the calculation of the city distance d1 is not performed because the candidate character recognition unit 1901 has already performed the calculation. Also, with regard to the calculation of the city area distance d2, only the character of the character code notified from the candidate character recognition unit 1901 is given by the equation (2).
Calculate using The calculated city distance d2 is multiplied by 0.5 and the smallest total city distance d added to the city distance d1 is notified to the recognition result output unit 307 as a recognized character.

FIG. 20B shows the total city distance d between the character image “den” 101 and the character image “den” 201. As a result, both the character images “den” 101 and 201 are correctly recognized as the character “den”. In the present embodiment, the candidate character recognizing unit 1901 causes the city area distance d2 between the feature amount at the center of the character image to be recognized and the standard feature amount.
Can be narrowed down (from all the characters in the recognition dictionary 301 to, for example, three characters), the processing speed is increased.

Next, the operation of this embodiment will be described with reference to the flowchart of FIG. Note that S2102 to S2
Steps S1802 to S180 in the first embodiment
6 and the description is omitted. In step S2108, the image area classification unit 305 classifies each area into a peripheral part of the first group and a central part of the second group, and the candidate character recognition unit 1901
And the recognition unit 1902 of the classification.

The candidate character recognizing section 1901 calculates the city area distance d1 between the notified feature amount of the peripheral area and the corresponding standard feature amount registered in the recognition dictionary 301 by using equation (1). (S2110). The processing of S2110 is repeated for all the characters registered in the recognition dictionary 301 (S2112), and three characters with a small (close) city distance d1 are selected, and their character codes are recognized by the recognition unit 19.
02 is notified (S2114).

The recognizing unit 1902 calculates the characteristic amount of the central region classified into the second group of the character images to be recognized and the standard characteristic amount of the central region of the character code notified from the candidate character recognizing unit 1901. Is calculated using equation (2) and multiplied by a factor of 0.5 (S2116). A value obtained by multiplying the city distance d1 notified from the candidate character recognition unit 1901 by the coefficient 1.0 and a value obtained by multiplying the city distance d2 by the coefficient 0.5 are added to obtain a total city distance d (S21).
18). The processes of S2116 and S2118 are repeated until calculation of the city area distance d with the notified standard feature amount of three characters is completed (S2120). The recognizing unit 1902 selects a character having the smallest total city distance d as the recognition result, and notifies the character code to the recognition result output unit 307 (S212).
2), end the process. (Embodiment 3) FIG. 22 is a configuration diagram of a character recognition apparatus according to Embodiment 3 of the present invention. This character recognition device
Recognition dictionary 301, character image input unit 302, image region division unit 303, feature amount extraction unit 304, image region classification unit 305, large classification dictionary 2201, large classification unit 2202
, A recognition unit 2203, and a recognition result output unit 307. The same components as those in the first embodiment are denoted by the same reference numerals, and the description thereof will be omitted. Only the components unique to the present embodiment will be described.

The large classification dictionary 2201 groups characters in similar character units having similar standard character shapes, and registers a group standard feature amount in similar character units and a character code group of characters belonging to similar character units. I have. FIG. 23 is a diagram illustrating an example of the contents of the large classification dictionary 2201. The drawing exemplifies “D1”, “D2”, and “D3” in similar character units. In the similar character unit “D1”, a numeral “1” and alphabets “i” and “t” are grouped, and their character code groups (in the figure, actual characters “1”, “ i ”and“ t ”2301) and a group standard feature value 2302 are registered.

The similar character unit “D2” includes the kanji “cloud”,
“Den” and “Lightning” are grouped, and a character code group 2303 and a group standard feature value 2304 are registered. The similar character unit “D3” includes the kanji “house”,
“Den” and “dragon” are grouped, and a character code group 2305 and a group standard feature amount 2306 are registered.

The grouping in units of similar characters is performed by cluster analysis of standard features of the recognition dictionary 301 and classifying characters into clusters in units of predetermined similar characters. In addition,
The cluster analysis is described in "Multivariate Statistical Analysis", written by Yutaka Tanaka and Kazumasa Wakimoto, Gendai Mathematics, pp. 230-244. Here, overlapping registration of characters in units of similar characters is permitted. The group standard feature amount is given by a simple average of the standard feature amounts of characters that are the basis of grouping.

Of course, similar character units may be grouped in other methods, for example, groups of identical Chinese characters such as "biased" and "nearby" may be grouped in similar character units. Large classification section 22
02 denotes a feature amount of the character image to be recognized in the area classified into the first group by the image area classification unit 305 and the large classification dictionary 2201.
Is calculated using the equation (1). At this time, Fsi is a group standard feature amount. Large area dictionary 2 for city distance d1
The calculation is performed for all the similar character units registered in 201, a similar character unit with the closest city area distance d1 (high similarity) is selected, and the character code group is recognized by the recognition unit 2203.
Notify.

FIG. 24A is a diagram showing the calculation result of the city distance d1 of the character images 101 and 201 to be recognized by the large classification unit 2202. Similar character units are shown in descending order of similarity. In the character image “den” 101, “cloud”, “den”, and “thunder” 2303 of the similar character unit “D2” are selected, and in the character image “den” 201, the similar character unit “D”
“Home”, “Den”, and “Dragon” 2305 of “3” are selected, and their character codes are notified to the recognition unit 2203.

When the recognition unit 2203 receives the notification of the character code group from the large classification unit 2202, it recognizes the feature amount of the peripheral portion of the first group of the character image to be recognized and the notified character registered in the recognition dictionary 301. The city distance d1 with the corresponding standard feature value of the code is calculated using equation (1). Similarly, the city distance d2 between the feature amount at the center of the second group and the corresponding standard feature amount is calculated using the equation (2), and the city distance d2 is multiplied by a coefficient 0.5. The total city distance d is calculated by summing the city distance d1 and the city distance d2 multiplied by the coefficient 0.5, and is stored together with the character code. After performing this process for the notified character code group, the character code having the smallest total city distance d is notified to the recognition result output unit 307.

FIG. 24B is a diagram showing a calculation result of the total city distance d of the character images “D” 101 and 201 to be recognized by the recognition unit 2203. Character image "den" 101, 20
In both cases, the character with the highest similarity is recognized as “den”. In the present embodiment, among the image regions of the character image to be recognized, a similar character unit grouped from the feature amount of the image region that more clearly reflects the character shape feature is selected,
Recognition character candidates are narrowed down, and then, the similarity is calculated by multiplying the city area distance by a different coefficient according to the image area classified as in the first embodiment. As a result, the time required for comparison with the recognition dictionary 301 is greatly reduced.

Next, the operation of this embodiment will be described with reference to the flowchart of FIG. Steps up to S2506 are the same as S1806 of the first embodiment, and a description thereof will not be repeated. The image area classification unit 305 classifies each area into a first group in the peripheral part and a second group in the central part, and notifies the large classification unit 2202 and the recognition unit 2203 of the classified area (S250).
8).

The large classification unit 2202 calculates the city distance d1 between the characteristic amount of the peripheral portion of the recognition target character image and the corresponding group standard characteristic amount of similar character units in the large classification dictionary 2201 (S2510). S2510 is repeated until all similar character units registered in the large classification dictionary 2201 are collated (S2512). Next, the large classification unit 2202 selects a similar character unit with the closest city area distance d1, that is, a similar character unit with the highest similarity, and notifies the recognition unit 2203 of a character code group of characters included in the similar character unit (S2514). ).

Upon being notified of each character code from the large classifying unit 2202, the recognizing unit 2203 determines whether or not the feature amount of the character image to be recognized belongs to the peripheral part of the area (S251).
6) If it belongs to the peripheral part, the city distance d1 between the feature amount and the corresponding standard feature amount of the notified character code is calculated by using the expression (1) with reference to the recognition dictionary 301, and the city distance d1 is calculated. The coefficient is multiplied by 1.0 (S2518). When the feature amount does not belong to the peripheral portion, the city area distance d2 between the feature amount and the corresponding standard feature amount of the notified character code is calculated using Equation (2) with reference to the recognition dictionary 301, and the city area distance is calculated. d2 is multiplied by a coefficient of 0.5 (S2520). Next, the total city distance d obtained by summing the values obtained in S2518 and S2520 is obtained and stored together with the character code (S
2522). For the notified character code, the collation with the recognition dictionary 301 is repeated from S2516 to S2522 (S2524), and the character code with the closest total city distance d (high similarity) is selected (S2526), and the process is terminated. .

In this embodiment, the large classification dictionary 2201
Has registered the group standard feature of the character group belonging to the similar character unit for all of the image regions, but has registered only the group standard feature of the peripheral region classified into the first group together with the character code group. You may. (Embodiment 4) FIG. 26 is a configuration diagram of a character recognition apparatus according to Embodiment 4 of the present invention. This character recognition device
Recognition dictionary 301, character image input unit 302, image area dividing unit 303, feature amount extracting unit 304, crush degree calculating unit 2601, recognition unit 2602, recognition result output unit 3
07. The same components as those in the first embodiment are denoted by the same reference numerals, and the description thereof will be omitted. Only the components unique to the present embodiment will be described.

The crushing degree calculating section 2601 calculates the ratio of the character element portion to the area for each of the 16 areas divided by the image area dividing section 303. Specifically, the number of “black” pixels included in the area is counted, the number is divided by the total number of pixels “16” of the area, and multiplied by “100”. An area where this ratio is less than the threshold value is regarded as a normal part and is classified into the first group, and an area where the ratio is equal to or more than the threshold value is regarded as the collapsed part and is classified into the second group. Here, the threshold is “75%”.

FIG. 27 shows a normal part 2701 of the first group and a crushed part 2 of the second group classified by the crush degree calculating unit 2601.
702 to the character images “Den” 101 and 201 in FIGS.
It is shown about. According to this, it is understood that the character image “den” 201 has a larger number of collapsed parts than the character image “den” 101. Crush degree calculation unit 2601
Recognizes each area classified into the first group and the second group
Notify 2.

The recognizing unit 2602 includes the crush degree calculating unit 26
When the area of the first group and the second group is notified from 01, the first group is notified.
Features of recognition target character image of group area and recognition dictionary 301
Is calculated, and a city distance d1 between the corresponding standard feature amount of the second dictionary and the corresponding standard feature amount of the recognition dictionary 301 is calculated by multiplying the same by the coefficient α. Multiply by the coefficient β. The obtained values are summed to obtain a total city distance d. Total city distance d obtained
And a character code corresponding to the standard feature amount, and this process is performed on all the characters registered in the recognition dictionary 301.

In the character image “den” 101 to be recognized, the first
Since there are 13 groups, the city distance d1 is calculated by equation (3). Since the area of the second group is three areas, the city distance d2 is
It is calculated by equation (4). In the recognition target character image “den” 201, the first group area is 8
Since there are eight regions in the second group, i is 1 to i = 32 in Expression (3), and j is j = 1 to j = 32 in Expression (4). Note that Foi, Fsi, Foj, F
sj, i, and j are the same as those described in the first embodiment.

Here, when α = 1.0 and β = 0.5 and the total city distance d is obtained for the character images “Den” 101 and 201, the recognition target character images “Den” 101 and 201 are both characters “Den”. Is the minimum total city distance d, that is, the similarity is the highest, is selected as the recognition result, and is notified to the recognition result output unit 307. Next, the operation of the present embodiment will be described.

The operation of the present embodiment is different from the operation of the first embodiment only in steps S1808 to S1814 (see FIG. 18). S18
In step 08, the image area classifying unit 305 classifies the image into the first group of the peripheral part and the second group of the central part, which are set in advance. The ratio of “black” pixels corresponding to the character element portion in the area is calculated, and if the value is less than the threshold value, it is classified into the first group as a normal part. Classify into the second group.

In step S1810, the recognizing unit 306 determines whether or not the feature amount of the character image to be recognized belongs to the first group of the peripheral part. However, in the present embodiment, the recognizing unit 2602 uses the first part of the normal part. It is determined whether it belongs to a group. If it belongs to the first group of the normal part, the process moves to S1812 and the second part of the collapsed part
If it belongs to the group, the process moves to S1814. Recognition unit 2602
Calculates the city distances d1 and d2 in S1812 and S1814 with the peripheral portion being a normal portion and the central portion being a collapsed portion, and multiplying d1 and d2 by coefficients 1.0 and 0.5, respectively.

As described above, in the present embodiment, the regions divided according to the recognition target character image input from the character image input unit 302 are classified into the first group and the second group. Recognition accuracy is improved compared to classification into groups. (Fifth Embodiment) FIG. 29 is a configuration diagram of a character recognition device according to a fifth embodiment of the present invention. This character recognition device
Recognition dictionary 301, character image input unit 302, image region division unit 303, feature amount extraction unit 304, crush degree calculation unit 2601, candidate character recognition unit 2901, recognition unit 29
02 and a recognition result output unit 307. In addition,
The same components as those in the first and fourth embodiments are denoted by the same reference numerals, and the description thereof will be omitted. Also, the candidate character recognition unit 2901 and the recognition unit 2902 of the present embodiment
Has almost the same configuration as the candidate character recognizing unit 1901 and the recognizing unit 1902 of the second embodiment, but has a slight difference from the difference between the image area classifying unit 305 and the crush degree calculating unit 2601.

The crushing degree calculating section 2601 determines whether or not the ratio of the character element portion in each region of the character image to be recognized is less than a threshold value, for example, 75%. It notifies the candidate character recognizing unit 2901 as a normal part and notifies the recognizing unit 2902 of the above area as a second group crushed part. The candidate character recognizing unit 2901 includes a crush degree calculating unit 26.
01 is read from the feature value extraction unit 304 and the feature value of the normal portion area classified into the first group notified from
Calculate the city distance d1 with the corresponding standard feature amount registered in 01.

Now, the recognition target character image "den" 101, 20
As shown in FIG. 27, when each area of the first group is classified into the first group and the second group, the correspondence between the feature amount of the normal part of the first group and the recognition dictionary 301 is calculated using Expression (3). The city distance d1 with the standard feature value to be calculated is calculated. When the city distance d1 is obtained for all the characters in the recognition dictionary 301, the result is as shown in FIG. The recognition unit 2902 is notified of a set of the character codes of the top three characters having a short city distance d1, that is, the top three characters having a high degree of similarity, and the city distance d1.

The recognizing unit 2902 calculates the city distance d2 between the feature quantity classified into the second group and the standard feature quantity corresponding to the notified character code registered in the recognition dictionary 301 by using equation (4). To calculate. The recognizing unit 2902 sums a value obtained by multiplying the city distance d1 by the coefficient α and a value obtained by multiplying the calculated city distance d2 by the coefficient β to obtain a total city distance d. When α = 1.0 and β = 0.5, the total city distance d of the character images 101 and 201 to be recognized is as shown in FIG.
(b). As a result, the character code of the character “den” having the highest similarity is notified to the recognition result output unit 307.

The operation of the present embodiment is slightly different from the operation of the second embodiment, and therefore, different portions will be briefly described with reference to FIG. In step S2108, the crushing degree calculation unit 2601 classifies the areas into a first group of normal parts and a second group of crushed parts. In step S2110, the candidate character recognition unit 2
A step 901 obtains an urban area distance d1 using the feature amount of the first group. Thereby, in S2114, the recognition dictionary 3
The calculation target of the city distance d2 with the standard feature value of 01 is narrowed down to three characters. (Embodiment 6) FIG. 31 is a configuration diagram of a character recognition apparatus according to Embodiment 6 of the present invention. This character recognition device
A recognition dictionary 301, a character image input unit 302, an image region dividing unit 303, a feature amount extracting unit 304, a crush degree calculating unit 2601, a large classification dictionary 2201, and a large classification unit 31
01, a recognition unit 3102, and a recognition result output unit 307.

The same components as those in the first, third, and fourth embodiments are denoted by the same reference numerals, and description thereof is omitted. The large classifying unit 3101 includes a crush degree calculating unit 2
The urban area distance d1 between the feature amount of the area classified into the first group notified from 601 and the corresponding group standard feature amount registered in the large classification dictionary 2201 is calculated. The city distance d1 is obtained for all similar character units registered in the large classification dictionary 2201, the most similar similar character unit is selected, and the character code of the character included in the similar character unit is notified to the recognition unit 3102.

Now, as shown in FIG. 27, when the areas are classified into a normal part 2701 of the first group and a crushed part 2702 of the second group, the large classification unit 3101 The city distance d1 is calculated using (3) and the like. At this time, Fsi is a group standard feature amount of the large classification dictionary 2201. FIG. 32 (a) shows the city area distance d1 for the recognition target character images "den" 101 and 201, respectively. The most similar character unit is “D2” for “den” 101 and “D3” for “den” 201.

When the character code group is notified from the large classification unit 3101, the recognition unit 3102, when the feature amount of the character image to be recognized is classified into the first group, recognizes the notified character in the recognition dictionary 301. The city distance d1 with the corresponding standard feature value of the code is calculated using the equation (3) or the like. When the object is classified into the second group, the city distance d2 between the feature amount and the corresponding standard feature amount of the recognition dictionary 301 is determined.
Is calculated using Equation (4) and the like. Obtained city distance d
1 and d2 are multiplied by coefficients α and β, respectively, to obtain a total city distance d. This process is performed only for the characters in the notified character code group. The character code of the character having the minimum total city distance d is notified to the recognition result output unit 307.

FIG. 32B shows the total city distance of the character images 101 and 201 to be recognized when the coefficients α and β are 1.0 and 0.5, respectively. The recognition results are both characters "den". Thus, the large classification dictionary 2201
By calculating the city distance d1 with the group standard feature amount of, it is possible to narrow down the processing target for calculating the total city distance d using the recognition dictionary 301, thereby improving the recognition accuracy and shortening the time required for the recognition processing. Can be achieved.

Next, the operation of the present embodiment will be described in Embodiment 3.
This will be briefly described with reference to the flowchart (FIG. 25).
Instead of S2508, the degree-of-crush calculation unit 2601 determines each area divided by the image area division unit 303 as the first part of the normal part.
It is classified into a group and a second group of crushed portions. In step S2510, the large classification unit 3101 calculates an urban distance d1 between the feature amount of the first group of the recognition target character image and the group standard feature amount of the large classification dictionary 2201.

In S2516, the recognition unit 3102
It is determined whether or not the feature quantity belongs to the normal part of the first group, and thereafter, the same processing as in the third embodiment is performed. In the first to sixth embodiments, the standard normalized character image registered in the recognition dictionary 301 is 16 × 16 pixels. However, it is needless to say that other sizes may be used. , The character images to be recognized input by the character image input unit 302 may have the same size.

Although the standard character image area is divided into 16 equal parts of 4 pixels each in the vertical and horizontal directions, it is needless to say that the area may be divided by pixels different in the vertical and horizontal directions and may be divided into “p × q”. . In this case, the character image to be recognized may be similarly “p × q” divided by the image area dividing unit 303. Further, although the image region dividing unit 303 divides the normalized character image to be recognized into each region, the character region input unit 302 does not normalize the character image. The character image to be recognized is divided into “p × q” regions at the center of gravity or at equal pitches, and the feature amount extraction unit counts the number of feature elements for each region, and then calculates the area such as the area and length of each region. The number of feature elements may be normalized by size to obtain a feature amount. In this case, the recognition dictionary registers the standard feature amounts of the standard character images extracted by the same method.

Further, in the above embodiment, the similarity is weighted by classifying the area into two groups, a first group and a second group, and multiplying the city distance indicating the similarity by two kinds of coefficients. In another embodiment, the distance may be classified into three or more groups, and different coefficients may be multiplied by the distances in the city area to further improve the accuracy of character recognition. Further, in the above embodiment, the city distance is used for obtaining the similarity. However, the present invention is not limited to this, and the Euclidean distance or the Mahalanobis distance may be used.

In the first to sixth embodiments, the structure is shown in FIGS. 3, 19, 22, 26, 29, and 31, respectively. However, as another embodiment of the present invention, A program in which the functions of the constituent elements shown in the drawings are exhibited is recorded on a computer-readable recording medium, and the program is read by being mounted on a character recognition device having no such function, and is similar to the present invention. The effect of may be exerted.

[0069]

As described above, the present invention relates to a character recognition apparatus for recognizing a character image to be recognized as a character code, wherein a standard image of a character is stored in N (N ≧ 2) partial image areas. A recognition dictionary that divides and pre-registers standard features of partial character images included in each partial image region together with character codes in character units, an image region dividing unit that divides a recognition target character image into N partial image regions, A feature extracting unit for extracting a feature of a partial character image included in the partial image region divided by the image region dividing unit; and a partial image region for classifying the partial image regions divided by the image region dividing unit into a plurality of groups. Classifying means, for each group, calculate the elementary similarity between the partial character image feature of the recognition target character image and the corresponding standard feature of the recognition dictionary, and weight the elementary similarity of each group by a predetermined weight. , Sentence to be recognized A similarity calculation means for calculating a similarity between the image and the recognition dictionary character, is set to be equipped with a recognized character selection means for selecting the character codes of the similarity highest character as the recognized character.

With such a configuration, the overall similarity is calculated by adding the weight corresponding to the contribution of each group to character recognition to the elemental similarity of each group, and the character image to be recognized is crushed. Even if there is, a character recognition device with improved character recognition accuracy can be obtained. The partial image area classifying unit includes a peripheral / central classifying unit that classifies the partial image area into a first group of peripheral parts and a second group of central parts of the recognition target character image, The means for separately calculating elementary similarity, which is a city area distance, a Euclidean distance or a Mahalanobis distance, between the feature of the partial character image and the corresponding standard feature of the recognition dictionary in the first and second groups. A similarity calculation unit configured to assign predetermined coefficients to the elementary similarities between the first group and the second group such that the contribution of the first group to the similarity is greater than the contribution of the second group to the similarity; And a similarity totalization unit for multiplying and summing.

With such a configuration, the contribution of the elemental similarity at the center of a character image in which the character image to be recognized is generally collapsed is reduced, and the similarity is calculated to improve the accuracy of character recognition. it can. Further, the partial image area classifying means has a peripheral / central classifying section that classifies the partial image areas into a first group of peripheral parts and a second group of central parts of the character image to be recognized, including: The similarity calculating means includes:
Calculating the elementary similarity of the first group, and recognizing a character group having a higher elementary similarity, which is an urban area distance Euclidean distance or a Maharabis distance, between the feature of the partial character image and the corresponding standard feature of the recognition dictionary, in the recognition dictionary A second group element similarity calculation unit that calculates a second group element similarity degree for the candidate character selected by the candidate character selection unit; A similarity totalization unit that multiplies the elementary similarity between the first group and the second group by a predetermined coefficient so that the contribution to the similarity is greater than the contribution to the similarity of the second group; I have to do that.

With such a configuration, in general, candidate characters are selected from the features of the peripheral portion that well represent the shape characteristics in the character image to be recognized, and the overall similarity is determined only for the selected candidate characters. Since the calculation is performed, the recognition speed can be improved and the processing speed can be increased. The partial image area classifying unit has a peripheral / central classifying unit that classifies the partial image area into a first group of peripheral parts and a second group of central parts of the character image to be recognized, and has a similar shape. A standard image group of characters is divided into N (N ≧ 2) partial image regions, and character group standard features of partial character images included in each partial image region are registered in advance together with character code groups of similar character units. Calculating the similarity between the character group classification dictionary and the character group partial feature of the character image to be recognized classified into the first group and the corresponding character group standard feature registered in the character group classification dictionary. Character group selection means for selecting a character code of a character group having a high degree, the similarity calculation means,
Calculating the degree of similarity with the feature of the partial character image to be recognized only with the standard feature registered in the recognition dictionary corresponding to the character code selected by the character group selecting means; Are elementary similarities that separately calculate the elementary similarity, which is the city area distance, the Euclidean distance or the Mahalanobis distance, between the feature of the partial character image and the corresponding standard feature of the recognition dictionary in the first and second groups. A degree calculation unit that multiplies the elementary similarity between the first group and the second group by a predetermined coefficient such that the contribution of the first group to the similarity is greater than the contribution of the second group to the similarity. And a total sum of similarities.

With such a configuration, a character group of a similar character unit having a similar shape is selected by using features of a portion that well represents the feature of the character image to be recognized, and the accuracy of the selected character group is improved. Since the similarity is calculated, the recognition speed can be improved and the processing speed can be further increased. The partial image area classifying unit may calculate a percentage of the partial character image in the partial image area for each partial image area, and a partial character image ratio calculating unit, wherein the ratio of the partial character image is less than a threshold value. As described above, the image processing apparatus includes a normal / crushed portion determination unit that classifies the partial image region into a first group of normal portions and a second group of collapsed portions.

With such a configuration, each of the partial image regions of the individual character image to be recognized is determined to be a normal part and a blurred part, and is classified into a first group and a second group. Since a predetermined weight is assigned to the elementary similarity between the partial image regions of the two groups, the reliability of the similarity is increased, and the accuracy of character recognition is further improved. In addition, the similarity calculating means may include an elementary similarity which is a city area distance, a Euclidean distance, or a Mahalanobis distance between a feature of the partial character image and a corresponding standard feature of the recognition dictionary in the first and second groups. And a contribution to the similarity of the first group is a second
There is provided a similarity totaling unit that multiplies the elementary similarity between the first group and the second group by a predetermined coefficient so as to be greater than the contribution to the similarity of the group, and sums them.

With such a configuration, the contribution of the crushed portion of the partial image area to the similarity is reduced, so that the recognition accuracy can be improved. Further, the similarity calculating means calculates the elementary similarity of the first group, and calculates the elementary similarity which is a city area distance, a Euclidean distance or a Mahalanobis distance between the feature of the partial character image and the corresponding standard feature of the recognition dictionary. A candidate character selection unit for selecting a character group having a higher degree as a candidate character from the recognition dictionary, and a second group for calculating a second group elementary similarity for the candidate character selected by the candidate character selection unit An elementary similarity calculation unit, and a first similarity calculation unit configured to determine that the first group has a greater contribution to the similarity than the second group has a similarity.
It has a similarity summing section for multiplying the elementary similarity between the group and the second group by a predetermined coefficient, respectively, and summing them.

With such a configuration, candidate characters are selected from the similarities of only the first group of the partial image areas determined to be normal, and the similarity is calculated for the selected candidate characters. Speed and recognition accuracy can be improved. Further, a standard image group of characters having similar shapes is divided into N (N ≧ 2) partial image regions, and a character group standard feature of a partial character image included in each partial image region is represented by a character code in units of similar characters. A similarity between the character group classification dictionary registered in advance with the group, the characteristic of the partial character image of the recognition target character image classified into the first group, and the corresponding character group standard characteristic registered in the character group classification dictionary Character group selecting means for calculating a degree and selecting a character code of a character group having a high degree of similarity, wherein the similarity degree calculating means stores the character dictionary in the recognition dictionary corresponding to the character code selected by the character group selecting means. Only between the registered standard feature and the registered standard feature, the similarity with the feature of the partial character image to be recognized is calculated, and the similarity calculating means calculates the feature of the partial character image between the first group and the second group. City with corresponding standard features of the recognition dictionary A similarity calculating unit that separately calculates a prime similarity that is a distance, a Euclidean distance, or a Mahalanobis distance; and a similarity calculating unit that makes a contribution to the similarity of the first group larger than a contribution to the similarity of the second group. 1st group and 2nd
And a similarity summing section for multiplying the elementary similarity with the group by a predetermined coefficient and summing them.

With such a configuration, it is possible to narrow down a character group including the characters of the character image to be recognized from the characteristics of the partial image area determined to be a normal part, so that the similarity calculation can be speeded up. Further, the coefficient multiplied by the similarity total part is 1.0 for the elementary similarity of the first group and 0.5 for the elementary similarity of the second group. Is to select the character having the lowest similarity as the character having the highest similarity. .

With such a configuration, the contribution to the similarity of the central portion of the recognition target character image is reduced to half of that of the peripheral portion, thereby dramatically improving the character recognition rate. The threshold value is set to 75%. With such a configuration, the first group and the second group of the normal part and the crushed part are appropriately classified, and the accuracy of character recognition is improved.

The present invention is also a character recognition method for recognizing a character image to be recognized as a character code, comprising: an image area dividing step of dividing the character image to be recognized into N (N ≧ 2) partial image areas; A feature extraction step of extracting features of a partial character image included in the partial image area; a partial image area classification step of classifying the partial image areas into a plurality of groups; and a partial character image of the recognition target character image for each group. And the standard image of the character is divided into N partial image regions, and the standard characteristic of the partial character image included in each partial image region corresponds to a corresponding recognition dictionary registered in advance together with a character code in character units. A similarity calculating step of calculating a similarity with the standard feature, weighting the basic similarity of each group with a predetermined weight, and calculating a similarity between the character image to be recognized and the character of the recognition dictionary; Degree is to having a recognized character selection step of selecting the character code having the highest character as the recognized character.

With this configuration, the overall similarity is calculated by adding the weight corresponding to the contribution of each group to character recognition to the elementary similarity of each group, and the character image to be recognized is crushed. Even if there is, a character recognition method with improved character recognition accuracy can be obtained. Further, according to the present invention, a standard image of a character is divided into N (N ≧ 2) partial image regions, and standard features of the partial character images included in each partial image region are registered in advance along with character codes in units of characters. And an image area dividing step of dividing the recognition target character image into N partial image areas, and extracting characteristics of the partial character images included in the partial image areas divided in the image area dividing step. A feature extraction step, a partial image region classification step of classifying the partial image regions divided in the image region division step into a plurality of groups, and a feature of the partial character image of the recognition target character image and the recognition dictionary for each group. A similarity calculating step of calculating a similarity between the corresponding standard feature and a predetermined weight for the basic similarity of each group to calculate a similarity between the character image to be recognized and the character of the recognition dictionary; , And with the degree of similarity is highest character selects the character codes as recognition character recognized character selection step and a computer-readable recording medium having a program recorded with.

With such a configuration, by mounting the recording medium on a conventional character recognition apparatus and reading the program, even if the recognition target character image is crushed, the apparatus can be used as a character recognition apparatus with improved recognition accuracy. be able to.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an example of a character image to be recognized input to an image input unit.

FIG. 2 is a diagram illustrating another example of a character image to be recognized input to an image input unit.

FIG. 3 is a configuration diagram of a first embodiment of a character recognition device according to the present invention.

FIG. 4A is a diagram illustrating an example of registered contents of the numeral “1” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "1" based on the registration content of (a).

FIG. 5 is a diagram illustrating feature elements for extracting a standard feature amount.

FIG. 6A is a diagram illustrating an example of registered contents of the alphabet “i” in the recognition dictionary according to the embodiment. (b)
FIG. 9 is a diagram showing a standard character image “i” that is the basis of the registered content of (a).

FIG. 7A is a diagram illustrating an example of registered contents of the alphabet “t” in the recognition dictionary according to the embodiment. (b)
FIG. 9 is a diagram showing a standard character image “t” that is the basis of the registered content of (a).

FIG. 8A is a diagram showing an example of registered contents of a kanji “cloud” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "cloud" based on the registration content of (a).

FIG. 9A is a diagram illustrating an example of registered contents of a kanji character “Rain” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "lightning" based on the registration content of (a).

FIG. 10A is a diagram showing an example of registered contents of a kanji character “home” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "home" based on the registration content of (a).

FIG. 11A is a diagram showing an example of registered contents of a kanji character “den” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "den" which becomes the basis of the registration content of (a).

FIG. 12A is a diagram showing an example of registered contents of a kanji character “dragon” in the recognition dictionary according to the embodiment. (b)
It is a figure which shows the standard character image "dragon" based on the registration content of (a).

FIG. 13 is a diagram illustrating a state where the recognition target character image “D” 101 illustrated in FIG. 1 is divided into regions by the image region dividing unit according to the embodiment.

FIG. 14 is a diagram showing characteristic amounts of a recognition target character image “den” 101 extracted by the characteristic amount extraction unit of the embodiment.

FIG. 15 is a diagram illustrating the feature amounts of the recognition target character image “den” 201 illustrated in FIG. 2 extracted by the feature amount extraction unit according to the embodiment.

FIG. 16 is a diagram showing a region divided by the image region dividing unit in the image region classifying unit according to the embodiment described above, as a first group in a peripheral part and a second group in a central part.
It is a figure showing the state classified into groups.

FIG. 17A is a diagram illustrating characters having a short total city distance in the recognition target character image “D” 101 calculated by the recognition unit according to the embodiment. (b) is a recognition target character image “den” 2 calculated by the recognition unit of the above embodiment.
It is a figure which shows the character of 01 with a short total city area distance. (c)
FIG. 3 is a diagram showing, as a reference example, a character having a short city distance in a character image 101 or 201 to be recognized obtained by a conventional character recognition device.

FIG. 18 is a flowchart illustrating an operation of the embodiment.

FIG. 19 is a configuration diagram of a character recognition device according to a second embodiment of the present invention.

FIG. 20 (a) shows that the candidate character recognition unit of the embodiment selects a candidate character having a short city distance using the feature amount of the first group of the recognition target character images “den” 101 and 201. FIG. (b) shows the recognition target character image “D” 1 using the result of candidate character recognition in the recognition unit of the above embodiment.
It is a figure which shows that 01 and 201 are selected.

FIG. 21 is a flowchart illustrating an operation of the embodiment.

FIG. 22 is a configuration diagram of a character recognition device according to a third embodiment of the present invention.

FIG. 23 is a diagram showing an example of registered contents of the large classification dictionary according to the embodiment.

FIG. 24A is a diagram illustrating that a similar character unit is selected using the first group of feature amounts of the recognition target character images “den” 101 and 201 in the large classification unit according to the embodiment. It is. (b) is a diagram showing that the characters of the recognition target character images “den” 101 and 201 are selected by the recognition unit of the above embodiment.

FIG. 25 is a flowchart illustrating the operation of the above embodiment.

FIG. 26 is a configuration diagram of a character recognition device according to a fourth embodiment of the present invention.

FIG. 27 is a diagram illustrating that the crushing degree calculating unit according to the embodiment classifies the recognition target character images “den” 101, 201 into a first group of normal parts and a second group of crushing parts.

FIG. 28 is a diagram illustrating that characters of a recognition target character image “den” 101 and 201 are selected by the recognition unit according to the embodiment.

FIG. 29 is a configuration diagram of a character recognition device according to a fifth embodiment of the present invention.

FIG. 30A is a diagram illustrating a first example of a normal part of the character images 101 and 201 to be recognized by the candidate character recognition unit according to the embodiment;
It is a figure showing that a candidate character is chosen using a feature quantity of a group. (b) shows that the recognition unit of the above-described embodiment uses the result of the candidate character recognition unit to recognize the recognition target character image “den” 101,
It is a figure which shows that the character of 201 is selected.

FIG. 31 is a configuration diagram of a character recognition device according to a sixth embodiment of the present invention.

FIG. 32 (a) shows that a similar character unit is selected by using the feature amount of the first group of the normal part of the recognition target character image “den” 101, 201 in the large classification unit of the above embodiment. FIG. (b) is a diagram showing that the characters of the recognition target character images “den” 101 and 201 are selected by the recognition unit of the above embodiment.

[Explanation of symbols]

301 Recognition dictionary 302 Character image input unit 303 Image region division unit 304 Feature extraction unit 305 Image region classification unit 306, 1902, 2203, 2602, 2902, 3
102 recognition unit 307 recognition result output unit 1901, 2901 candidate character recognition unit 2201 large classification dictionary 2202, 3101 large classification unit 2601 loss degree calculation unit

Claims (15)

[Claims] 1. A character recognition device for recognizing a character image to be recognized as a character code, wherein a standard image of a character is divided into N (N ≧ 2) partial image regions, and each character image is included in each partial image region. A recognition dictionary in which standard features of partial character images are registered in advance along with character codes in units of characters, an image region dividing unit that divides a recognition target character image into N partial image regions, A feature extraction unit for extracting a feature of a partial character image included in the partial image region; a partial image region classification unit for classifying the partial image regions divided by the image region division unit into a plurality of groups; Calculate the elementary similarity between the characteristic of the partial character image of the target character image and the corresponding standard characteristic of the recognition dictionary, weight the elementary similarity of each group by a predetermined weight, and calculate the elementary similarity between the recognition target character image and the recognition dictionary. Letters and Character recognition apparatus characterized by comprising: a similarity calculation means for calculating the similarity, a recognition character selection means for selecting the character codes of the similarity highest character as the recognized character. 2. The partial image region classifying means includes a peripheral / central classifying unit that classifies the partial image region into a first group of peripheral portions and a second group of central portions of a character image to be recognized, The similarity calculating means separately calculates the elementary similarity, which is a city distance, a Euclidean distance, or a Mahalanobis distance, between the feature of the partial character image and the corresponding standard feature of the recognition dictionary in the first group and the second group. An elementary similarity calculating unit for calculating the elementary similarity between the first group and the second group so that the contribution of the first group to the degree of similarity is greater than the contribution of the second group to the degree of similarity 2. A character recognition apparatus according to claim 1, further comprising: a similarity summing unit for multiplying the sum by multiplying the sum by a coefficient. 3. The partial image area classifying means includes: a first partial image area including:
A peripheral / central classification unit for classifying the partial group into a group and a central second group; wherein the similarity calculating means calculates the elementary similarity of the first group, and the feature of the partial character image and the recognition dictionary. A candidate character selection unit for selecting a character group having a higher elemental similarity, which is a city area distance Euclidean distance or a Maharabis distance, with the corresponding standard feature as a candidate character from the recognition dictionary, to a candidate character selected by the candidate character selection unit. On the other hand, a second group element similarity calculation unit for calculating the elementary similarity of the second group, and a second group element similarity calculation unit that makes the contribution to the similarity of the first group larger than the contribution to the similarity of the second group. 2. The character recognition device according to claim 1, further comprising: a similarity totaling unit that multiplies the elementary similarity between the first group and the second group by a predetermined coefficient to sum the result. 4. The partial image area classifying means includes a peripheral / central classifying section for classifying the partial image area into a first group of peripheral parts and a second group of central parts of the character image to be recognized. Is divided into N (N ≧ 2) partial image areas, and the character group standard features of the partial character images included in each partial image area are pre-set together with the character code group of similar character units. Calculating the similarity between the registered character group classification dictionary, the feature of the partial character image of the recognition target character image classified into the first group, and the corresponding character group standard feature registered in the character group classification dictionary And a character group selecting means for selecting a character code of a character group having a high degree of similarity, wherein the similarity calculating means is registered in the recognition dictionary corresponding to the character code selected by the character group selecting means. Recognized parts only between standard features Calculating the degree of similarity with the features of the character image; and the similarity calculating means includes: an urban area distance between the first group and the second group, the features of the partial character image and the corresponding standard features of the recognition dictionary; A primary similarity calculating unit that separately calculates a primary similarity that is a Euclidean distance or a Mahalanobis distance; and a first similarity calculation unit that makes a contribution to the similarity of the first group larger than a contribution to the similarity of the second group. 2. The character recognition device according to claim 1, further comprising: a similarity totaling unit that multiplies the elementary similarity between the group and the second group by a predetermined coefficient, respectively, and sums them. 5. A partial character image ratio calculating unit for calculating a ratio of a partial character image in the partial image region for each partial image region, a threshold value of the partial character image ratio. 2. The character recognition apparatus according to claim 1, further comprising a normal / crushed portion determination unit that classifies the partial image area into a first group of normal portions and a second group of crushed portions based on a value less than or equal to the value. . 6. The similarity calculating means is a city distance, a Euclidean distance, or a Mahalanobis distance between a feature of the partial character image and a corresponding standard feature of the recognition dictionary in the first group and the second group. A first similarity calculating unit for separately calculating a first similarity; a first similarity calculating unit configured to calculate a first similarity of the first group and a second second similarity such that a contribution of the first group to the similarity is greater than a second similarity to the second similarity. 6. The character recognition device according to claim 5, further comprising: a similarity totalization unit that multiplies each of the elementary similarities by a predetermined coefficient and sums them. 7. The similarity calculating means calculates the elementary similarity of the first group, and calculates an urban distance between a feature of the partial character image and a corresponding standard feature of the recognition dictionary.
A candidate character selection unit that selects a character group having a higher elemental similarity that is a Euclidean distance or a Mahalanobis distance as a candidate character from the recognition dictionary; and a second group of candidate characters selected by the candidate character selection unit. A second group prime similarity calculation unit for calculating a prime similarity; and a first group and a second group, such that a contribution of the first group to the similarity is greater than a contribution of the second group to the similarity. 6. The character recognition device according to claim 5, further comprising: a similarity totaling unit that multiplies each of the elementary similarities by a predetermined coefficient and sums them. 8. A standard image group of characters having similar shapes is divided into N (N ≧ 2) partial image regions, and a character group standard feature of a partial character image included in each partial image region is converted into a similar character unit. A character group classification dictionary registered in advance together with the character code group of the character group; characteristics of partial character images of the recognition target character images classified into the first group; and corresponding character group standard characteristics registered in the character group classification dictionary And character group selection means for calculating a character code of a character group with a high degree of similarity, wherein the similarity calculation means corresponds to the character code selected by the character group selection means. Only in the case of the standard feature registered in the recognition dictionary, the similarity with the feature of the partial character image to be recognized is calculated, and the similarity calculating means calculates the similarity between the first group and the second group. Image features and corresponding standard features of the recognition dictionary A similarity calculating unit for separately calculating the elementary similarity that is a city distance, a Euclidean distance, or a Mahalanobis distance to the first group, and the contribution to the similarity of the first group is greater than the contribution to the similarity of the second group 6. The character recognition device according to claim 5, further comprising a similarity totaling unit for multiplying the primary similarity between the first group and the second group by a predetermined coefficient and summing them. 9. The coefficient multiplied by the similarity total part is:
It is 1.0 for the elementary similarity of the first group, and 0.5 for the elementary similarity of the second group. The recognition character selecting means assigns the character having the smallest similarity to the similarity. 10. The character recognition device according to claim 2, wherein the character is selected as the best character. 10. The character recognition device according to claim 5, wherein said threshold value is 75%. 11. A character recognition method for recognizing a character image to be recognized as a character code, comprising: an image region dividing step of dividing the character image to be recognized into N (N ≧ 2) partial image regions; A feature extraction step of extracting a feature of a partial character image included in a partial image region classification step of classifying the partial image regions into a plurality of groups; a feature of a partial character image of the recognition target character image for each group; The standard image of the character is divided into N partial image areas,
The basic features of the partial character images included in each partial image area are calculated with the corresponding standard features of the recognition dictionary in which character codes are registered in advance along with the character codes in units of characters. Weighting, calculating a similarity between the recognition target character image and the character in the recognition dictionary, and a recognition character selecting step of selecting a character code of the character having the highest similarity as a recognition character. Character recognition method characterized by having. 12. The partial image area classification step includes a peripheral / central classification sub-step of classifying the partial image area into a first group of peripheral parts and a second group of central parts of the recognition target character image, In the similarity calculating step, the elementary similarity, which is a city area distance, a Euclidean distance or a Mahalanobis distance, between the feature of the partial character image and the corresponding standard feature of the recognition dictionary in the first group and the second group is separately determined. Calculating the elementary similarity between the first group and the second group such that the contribution to the similarity of the first group is greater than the contribution to the similarity of the second group. 12. The character recognition method according to claim 11, further comprising: a similarity synthesis sub-step of multiplying the sum by a predetermined coefficient and summing them. 13. The partial image area classifying step includes: a partial character image ratio calculating substep of calculating a ratio of the partial character image in the partial image region for each partial image region; 12. The character recognition according to claim 11, further comprising a normal / crush determination sub-step of classifying the partial image region into a first group of normal portions and a second group of crush portions according to a value less than or greater than a threshold value. Method. 14. A standard image of a character is divided into N (N ≧ 2) partial image areas, and standard features of the partial character images included in each partial image area are registered in advance in character units together with character codes in advance. An image area dividing step of dividing a recognition target character image into N partial image areas, and extracting a feature of a partial character image included in the partial image area divided in the image area dividing step. An extracting step; a partial image area classifying step of classifying the partial image areas divided in the image area dividing step into a plurality of groups; and a correspondence between a feature of the partial character image of the recognition target character image and the recognition dictionary for each group. A similarity calculation step for calculating the similarity between the character image to be recognized and the characters in the recognition dictionary by calculating the elementary similarity with the standard feature to be recognized and weighting the elementary similarity of each group with a predetermined weight. If the similarity is highest character recognition character selection step and a computer-readable recording medium having a program recorded with to elect the character codes as recognition character. 15. The partial image area classifying step, wherein a partial character image ratio calculating substep of calculating a ratio of the partial character image in the partial image region for each partial image region; 15. The computer-readable medium according to claim 14, further comprising a normal / crush determination sub-step of classifying the partial image region into a first group of normal portions and a second group of crush portions according to a value less than or equal to a threshold value. Possible recording medium.