JPH04199386A - Segmenting method for character - Google Patents

Abstract

PURPOSE:To surely segment a character even in the case that the character is blurred by segmenting the character on the basis of a density gradient. CONSTITUTION:A picture input means 10 picture-inputs a pattern to be recognized, and picture input information is pre-processed by a pre-processing means 11, and is inputted to a character segmenting means 12, and after the character is segmented, it is inputted to a picture element component extracting means 20, and a picture element component consisting of the density gradient direction component and the density gradient intensity component of every picture element of inputted picture data is extracted. Besides, a feature quantity extracting means 30 divides a picture inputted by the picture input means 10 into (mXn)- pieces of meshes by an area dividing means 31, and extracts the total sum of the density gradient intensity components of every density gradient direction component extracted by the picture element component extracting means 20 for every divided area in a vector quantity extracting means 32, and makes it the feature quantity of the pattern to be recognized. By segmenting the character by extracting density gradient information in a horizontal and a vertical directions in this way, the character can be correctly segmented even if the character is blurred.

Description

[Detailed description of the invention] Purpose of the invention.

(Industrial Application Field) The present invention relates to a method for cutting out characters used in character pattern recognition.

(Prior Art) Conventionally, as a method for cutting out characters, there is a method described in Japanese Patent Application Laid-Open No. 2-42598. To explain this method in detail using FIG. 23, each pixel of the read image data is binarized at a binarization level determined based on the density histogram of the image, and each pixel of the binarized image is A value is obtained by adding pixels in a direction perpendicular to the character cutting direction, and based on the added value, a projection His]-Durham is created in a horizontal direction. and,
A predetermined darkness value is set on the projection histogram, and the area that is larger than that threshold or where characters exist,
It is determined that the area smaller than the darkness value is an area where no characters exist, and the characters are extracted accordingly.

Therefore, each character can be extracted from a character group consisting of a plurality of characters.

(Invention or problem to be solved) However, in such conventional methods, the 24th
As shown in the figure, if the printed characters were faded, there was a problem in that the characters could not be cut out correctly. Particularly, such problems occur when the amount of printing ink remaining in the check writer becomes low. Specifically, the character r4. in FIG. , r2. , r
If there is a blur in the vertical direction of the character as shown in O,, rO,, creating a projection histogram of a binary image in the horizontal direction and simply binarizing it using the darkness value will not solve the blurred part. is less than the threshold, so it is determined that there is no character area. Therefore, when a character is cut out, one character ends up being cut out into a plurality of character areas, resulting in a problem that the cutout of the character fails or is partially cut out.

On the other hand, when looking at the density gradient, there is no density gradient in the part with the letter 2J, but there is a density gradient in the blurred part, so the two can be reliably distinguished by the density gradient.

The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to cut out characters based on the density gradient, thereby reliably cutting out characters even when the characters are blurred. The object of the present invention is to provide a character cutting method that can cut out characters.

Structure of the Invention: (Means for Solving the Problems) The present invention relates to a method for cutting out characters in character pattern recognition. a density gradient information extracting means for calculating the density gradient direction and elementary gradient strength for each pixel of the image data; a first density gradient direction distribution extraction means for extracting a density gradient direction distribution in the character cutting direction by adding perpendicularly to the character cutting direction; and a gradient direction along the opposite direction to the character cutting direction. a second concentration gradient direction distribution extracting means for extracting a concentration gradient direction distribution in a direction opposite to the oven exit direction of the characters by adding perpendicularly to the cutting direction of the characters; and the first concentration gradient direction distribution. A first character edge detection means detects one end of the character based on the distribution state extracted by the extraction means, and a first character end detection means detects one end of the character based on the distribution state extracted by the second concentration gradient direction distribution extraction means. a second character edge detection means for detecting the other edge; the character edge detected by the first character edge detection means and the other character edge detected by the second character edge detection means; This is achieved by making the space between the characters a character area.

(Function) The present invention provides a method for cutting out characters for reliably recognizing numbers written on a check or the like with a check writer. The image of the amount entry field of a check, etc. (hereinafter referred to as iJs stamp, etc.) is taken for a long time, and each character constituting the amount information is cut out character by character using the input image data.
It recognizes each cut out character and ultimately reads the amount of money on a check, etc.

In the present invention, the density gradient direction and density gradient strength for each pixel of a character are obtained by a density gradient information extraction means, and the gradient direction along the character cutting direction is perpendicular to the text-card cutting direction. The density gradient direction distribution in the character cutting direction is extracted by addition, and the gradient direction along the direction opposite to the character cutting direction is added perpendicularly to the character cutting direction to extract the character cutting direction. The density gradient direction distribution in the opposite direction is extracted, and both ends of the character are detected and cut out based on the extracted distribution state. According to the present invention, since density gradient information in the horizontal and vertical directions is extracted, accurate cutting can be achieved even if the characters are faded or scratched.

(Example) The present invention will be explained in detail below.

This embodiment is for recognizing the amount written on a check, etc., and the types of characters to be finally recognized are the numbers "0 to 9" and the day symbol [¥J, '*j There are 12 types of final digit codes such as . FIG. 2(A) shows an example of a bill, in which an amount entry field 1 is provided in the center for entering the amount with a check writer. Also, Figure 2 (B
) shows an example of a check, in which an amount entry field 4 is provided in the center for entering the amount with a check writer. The present invention provides an amount entry column (
The present invention relates to a character extraction method for recognizing the 1st day symbol and the last digit symbol such as "*" written in 1 and 4) with a check writer.

First, an example of a character recognition device will be described with reference to FIG. ), the input image data is inputted to the character extraction method 12 of the present invention to perform character extraction, and then inputted to the pixel component extraction means 20, and the density gradient direction component and density gradient of each pixel of the inputted image data are inputted to the pixel component extraction means 20. Extract pixel components consisting of intensity components. When the extraction of pixel components is completed, processing by the feature extracting means 30 is started. Feature extraction means 3
0 is first input to the image input means 1 in the area division means 31.
The input image is divided into mxn meshes, and the hector quantity extraction means 35 extracts the sum of the concentration gradient strength components for each concentration gradient direction component extracted by the pixel component extraction means 20 for each divided area. is used as the feature amount of the pattern to be recognized. The storage means 50 consists of a major classification dictionary 51 and a medium classification dictionary 52, and the comparison and matching means 40 compares the feature amount extracted by the feature amount extraction means 30 and the feature amount of the standard pattern stored in the storage means 50. The system performs a comparison and outputs a recognition result based on the comparison result.

Next, a recognition dictionary used for character recognition will be explained.

Fig. 3 is a schematic flowchart showing the operation of creating a recognition dictionary.In Fig. 3, first, an outline of the extraction of the standard time trace amount for each standard pattern will be explained.First, an image input for the standard pattern is performed. (Step S1), smoothing is performed as preprocessing (Step S2), and then a standard time minute amount is extracted for each standard pattern (Step S3). Here, input the standard pattern and its image.

The smoothing process and standard time trace extraction process will be explained in detail.

The types of standard patterns are as shown in FIG. 6(A), and 220 patterns have been created in advance for 12 types of characters, taking into consideration the form of the characters.

Each of these 220 patterns is a standard pattern and includes most of the commonly used character forms. Some of these standard patterns are shown in detail in FIGS. 6(B) to 6(D). Image input is performed by reading a standard pattern of a predetermined size created in advance in this manner, for example, in a rectangular range of 24 x 24 pixels, with the density gradation of each pixel being read in 256 8-bit gradations. An example of this input image data is shown in FIG. 4(A), and this image data is subjected to smoothing processing as pre-processing, and this smoothing processing will be explained next.

FIG. 7 is a flowchart showing an example of the operation of the smoothing process, and FIGS. 8 to 10 are explanatory diagrams thereof. In the smoothing process, the input image data is smoothed by scanning the input image data with a 3×3 pixel smoothing mask shown in FIG. 1O. That is,
First, as shown in FIG. 8, based on the density value of each pixel of the input image data, a histogram of the number of pixels of the density value is created (step 520), and the density value that becomes the peak is used as the background density value (BP). (Step 521)
. This is because a normal character pattern has more background parts than character parts, and the density value of the background part is constant, so the peak density value (BP) becomes the density value of the background.

Then, as shown in FIG. 9, one pixel of the background density value is arranged around the 24×24 pixels to obtain image data of 26×2B pixels (step 522).

This means that the 3x3 pixel smoothing mask is 24x24
This is to enable the operation to be performed on all input pixel data within a pixel. Next, the 26×26 pixel image data is scanned with the smoothing mask shown in FIG. 10 (step 523), the input image data is smoothed, and the result is used as subsequent image data. An example of image data at this stage is shown in FIG. 4(B). By performing the smoothing process in this way, the pattern of characters engraved by the check writer is filled with the density value of the characters, and the character engraved pattern can be erased.

Therefore, there are advantages to be gained from stable character features.

Next, a standard time trace amount is extracted from the image data smoothed as described above (step S3 in FIG. 3). The extraction operation of this standard time trace amount will be explained with reference to FIGS. 12 to 15.

Figure 12 is al! This is a flowchart showing the extraction operation of a standard time minute amount of a quasi-pattern, in which first, the Rohinson operator is used on the smoothed image data to obtain the density gradient strength and density gradient direction for each pixel (steps 530 to 532).・The operator is shown in Fig. 13, which consists of 8 types of masks M0 to M7 in which each pixel of 3 x 3 pixels is weighted, and each mask M o '' M 7 has a direction that the mask has. For example, mask M.

When the image data is scanned, the gradient strength (EPo) in the ↑ direction is determined for each pixel. In this way, by scanning the image data using the masks Mo to M7 in the 87i direction, the concentration gradient strength (EPo
~EP7) can be found (Ste Tsubu 530)
. Note that the concentration gradient intensities EP and -EP7 correspond to the masks MO to M7, respectively. The maximum value (EP, ) is determined from among the concentration gradient intensities (EPo-EP7) determined for each pixel in this manner (step 531). Then, the maximum value (EP+) is set as the density gradient strength of that pixel, and the direction corresponding to the maximum value EP is set as the density gradient direction (step 532). FIG. 4(C) shows the image data obtained in this manner.

Next, binarization processing is performed on the density gradient strength of each pixel obtained in this way (steps S33, 540 to
543). A method for determining a threshold value for binarization processing will be described with reference to FIG. 14(A). First, find the maximum value of concentration gradient strength (EPffi, ) among all pixels (
Step 533), check if this maximum value (EP□8) is within a predetermined range (Step 540), and if there is any other value, perform error processing as an image input error (Step 541). Image input errors can be detected at the early stage of image processing. And the maximum value (EP
-8) is within the predetermined range, this 1/8 EPmax
/8 is set as a threshold value (EPTH; hereinafter referred to as "effective level threshold value") (step 542). If the density gradient strength (Ej+) of each pixel is thicker than the effective level threshold (EPTH), the density gradient strength of that pixel is set to 1゛°, and if it is smaller than the effective level threshold (El”TH), the density gradient of that pixel is The intensity is set to "0" (step 543).As a result, the density gradient intensity of each pixel is binarized to 0 or "1".An example of this binarized image data is shown in the fourth example.
FIG. 4(E) is an example in which the concentration gradient direction is also displayed.

Note that here, it is possible to simplify by binarizing and determining the density gradient strength as 0 or 1, or to use the density gradient strength itself for each pixel as it is without binarizing it. As shown in Figure 14 (B) and (0), the concentration gradient strength can also be quantized and used. In the case of Figure 14 (B), the maximum value (EP,, a,) is divided into 8 steps. The concentration gradient strength is quantized in 8 steps from 0 to 7. In the case of Fig. 14 (C), the histogram of the concentration gradient strength is cumulatively added, and the maximum value is divided into 8 equal parts to calculate the concentration gradient. Intensity 0-7
It is quantized in 8 stages. If the density gradient strength for each pixel is determined using values quantized in multiple stages in this manner, accuracy will be further improved.

Next, as shown in FIG. 15(A), this binarized image data is equally divided (ixj) into a plurality of areas (step 544). Here, 16 areas of 4 x 4 (SP, [i] [j] [k] i-0 to 3
．． The case of classification into j-0 to j-3) is shown. however,
1 is the horizontal section number, J is the vertical section number, X
is the standard pattern number. And Figure 15 (B)
The sum of concentration gradient strengths for each concentration gradient direction for each area (sp, [i] [jl [k]; i・0~
3. j-0~3゜k-0~7) (step 5
45). However, k is the concentration gradient direction. The value SP obtained in this way, [i] [jl [k] is a standard time trace amount of the standard pattern, i X j X k (
−128) dimension vector quantity.

Through the above steps, the standard time trace amount of the standard pattern can be determined. Then, the feature amount is obtained for each standard time trace amount a predetermined number of times, for example, using the same 10 standard patterns, and the average value is stored as the regular standard time trace amount of that standard pattern (step S3).

In FIG. 3, the standard time trace amount (SP, [i] [jl [
k] ) is determined (step 54).

Next, the standard time trace amount SFX [i] [j
Create a major classification dictionary based on l[k] (step S
5). The contents of the major classification dictionary are as shown in Figure 16, and this major classification dictionary treats each standard pattern as a small group (hereinafter referred to as a cluster) of standard time trace amounts or those that have a certain similarity relationship. Consists of multiple clusters. Specifically, clusters are created using a generally known cluster analysis method, such as the Wart method, based on standard time trace amounts. In other words, the city-block distance indicates the degree of similarity between standard time traces of standard patterns! If (Dift
) for all standard pattern combinations, the following (1)
Obtained by the formula.

Dift (X, ×2) ・・・・・・・・・(1) (However, xi, x2 are standard pattern numbers) Then, as a result of calculating the city block distance for all combinations, the ones with the smallest distance are is one cluster,
The average value of the standard time trace amount of the standard pattern belonging to the cluster is taken as the feature quantity of the cluster. Next, the city-block distances of the features are calculated again for all combinations with clusters and other standard patterns, and the one with the smallest distance is defined as one cluster. From then on, the same operation is repeated until the city block distance becomes less than a predetermined value (for example, 100) or the number of clusters becomes a predetermined number (for example, 5).
Repeat until 0 pieces are left. In this way, it is a large classification dictionary that is divided into clusters based on the minimal similarity of standard patterns using Ward's method.

Furthermore, features are determined for each cluster. This is called a cluster feature. Cluster special (quantity (cspn [
i] [jl [k]; where n is the cluster number) is the average of the standard time trace amounts of the standard patterns belonging to each cluster. As will be described later, during actual recognition, major classification is performed based on this cluster feature amount, and the correspondence relationship between this cluster feature amount and clusters is a major classification dictionary. An example of the major classification dictionary at this stage is shown in FIG. On the other hand, the intermediate classification dictionary shows the correspondence between each standard pattern and its standard time trace amount. To explain one example in FIG. 16, for example, in cluster [14], the standard pattern belonging to this cluster is r3 of b3. , there is "5" in b5, and
The cluster feature amount C3PI4[mother[j
l [k] is the standard time trace amount of b3 SPy ti] [jl
Standard time trace amount of [k] and b5 5Rbs [i] [jl
It is the average value of [k].

Next, the operation of additionally registering a variation pattern in the major classification basic/dictionary shown in FIG. 16 will be explained from step 510 onward in FIG. The variation pattern is determined depending on the reading accuracy of the recognition medium (character pattern) of the recognition dictionary or the recognition device actually used. Specifically, the reading accuracy of the medium to be recognized by the recognition device is +4 degrees in rotation amount and ±± in deviation amount.
If it is 1 pixel (0,25+nm), 2 its fluctuation pattern is 14 degrees, -3 for example in the case of rotation.
10 types of variation patterns of 11 pixels and +1 pixel are created for the standard pattern (220 pieces) for degrees, -2 degrees, -1 degrees, +1 degrees, +2 degrees, +3 degrees, +4 degrees, and deviations. ~. An example thereof is shown in FIG. 17 for the character "0°". That is, with respect to the standard pattern shown in FIG. 17(^), the variation pattern is shown as shown in FIG. 17(8). Further, in this embodiment, the rotation and shift are made separately to create a variation pattern, but the shift is -1 pixel degree for -4 degrees of rotation.

It is also possible to add a variation pattern that combines rotation and displacement, such as:

A pattern signal of a standard pattern corresponding to the fluctuation pattern input as an image is input using the operation unit, and the fluctuation pattern is input as an image (step 510). The input image data is subjected to smoothing processing in the same way as when the standard pattern was input (step 511), and then the variable pattern feature amount (vp, [i] [jl[k]) is calculated (step 512), Based on this feature amount, a major classification is performed using a major classification dictionary. The major classification is based on the cluster standard feature value (C5P, [
i] [jl [k]) and the variation pattern feature (V
P, [i] [jl [k] ) similarity (D+5
tn) using the following formula (2).Step 513)
, is determined to belong to the cluster with the highest degree of similarity (D+stn is the minimum) (step 514).

1stn - Hiroko. ,=j, (vp, [i] [jl [k]
-C5P,, [i] l] [k]) :...
...(2) Where, n is the cluster number.

Then, it is checked whether a standard pattern corresponding to the variation pattern has already been registered in the failed cluster (step 515), and if it has not been registered, the pattern is additionally registered in the cluster (step s15). .517). For registration, only the pattern number corresponding to the m quasi-pattern is registered, and if it has already been registered, the above operation is performed for the next variation pattern. The above-described additional registration operation of variation patterns is performed for all of the variation patterns corresponding to the standard patterns (220 pieces), and ends when all of them are completed. Note that when additionally registering a variable pattern, only the pattern number of the standard pattern is additionally registered in the cluster, and the cluster feature used for major classification is not updated. This eliminates the need for major classification within all standard patterns again when adding a new standard pattern other than the existing 220, and also prevents the cluster features from slowing down. This is to expand the difference in similarity.

Next, when additionally registering the fluctuation pattern, the weight for each area of each pattern used in the detailed classification described later is calculated (step S18, 518AS19), and kneading will be explained. This is to check in advance which area is least affected by the change even if a change occurs during image input. This calculation operation is
This will be explained according to Figure 8.

n variable pattern features (VP-, p [
i] [jl [k]) For (P character variation pattern numbers 1 to n), the average feature amount μX for each 4×4 mesh area (ij) [1] [J] [k] and the feature for each area Quantity dispersion OX2 E;
) and (4), and the feature value variance for each area O
x' [i] [The reciprocal of jl is the weight WT [i] [j
Remember it as l. When the feature values have been calculated for all variation patterns, the calculated variation pattern characteristics 11
5! The average feature 11 and feature variance 0.2 for each mesh area (i, j) of the amount νPx f+J [J] fkl are calculated according to (3) and (4) below (step 55
0). Then, the reciprocal of the feature variance 08' is the weight WT,
[i] [jl step 551), jli mig? registration (step 552). This weight wT,
This is an area where [rJ[jl is large, or a stable area that is less affected by fluctuations that occur during image input.

σ X'[+][Jkon P”l Comparing with the dictionary, for example, cluster No.]4 is “e3. e5. b8
．． It can be seen that the four standard pattern numbers o8°゛ have been additionally registered. Note that the major classification dictionary shown in FIG. 19 is used in actual pattern recognition. here,
The major classification dictionary consists of clusters, which are groups of similar characters,
It shows the correspondence relationship between the cluster features and the cluster features, and major classification based on the major classification dictionary is to compare the features of the character to be recognized with the cluster characteristics and determine which cluster the character to be recognized belongs to. This is a process to determine whether it is a character.

Next, a character pattern recognition operation using the above-mentioned recognition dictionary will be explained based on the block diagram of FIG.

First, the image sensor of the image input means 10, for example, as shown in FIG.
Input the image of the amount entry column 4 of the check shown in B). This amount entry column 4 is a predetermined position, and the #light density in this column is acquired by an image sensor in 256 gradations.

FIG. 5(A) shows the original drawing of the amount entry field 4. The image data thus obtained is smoothed by the preprocessing means 11. As a result, Figure 5 (11)
Based on the smoothed image data, the character cutting means 2 cuts out the characters.

The °th figure shows the extraction of characters according to the present invention in the horizontal and vertical directions, and the upper left direction (M
1), the left direction (M2), and the lower left direction (M3) are scanned to obtain the concentration gradient strength of each. Then, a pixel having the maximum intensity or a predetermined value (EDIo) or more among the three intensities obtained for each pixel is detected, and the pixels are counted in the vertical direction to create a histogram in the horizontal direction (step 5101). '4.5 Figure (E) is a leftward histogram. Then, the leftward pistogram is sequentially scanned from the left side, and all rising points that are below a predetermined value (HT□) but above a predetermined value are detected and extracted as the left end candidate position of the character (step 5102).

Next, as in the case of the left direction, a right direction pistogram is created using three masks for the lower right direction (M5), right direction (M6), and upper right direction (M7) (step 5103).
. 74.5 Figure (F) is this rightward histogram. Then, in the same way as when extracting the left end, the right end candidate position of the character is extracted by sequentially scanning from the right side (step 5104).

Then, the area between the left end candidate and the right end candidate of the character extracted in this way is determined as each character region candidate (step 5
105). Figure 5 (G
), ()I), (1). Then, the width (W) of the character area is used to determine whether the obtained character area candidate is appropriate.
Then, the pitch (P) of the mutual spacing between the character areas is determined (step 5106), and if it is inappropriate, it is corrected and the character area at the optimal position is cut out (step 5107). Regarding the modification, see Figure 5 (G) and ()I).

To explain using (r), first, each of the determined widths (
Check whether W) and pitch (P) are within predetermined tolerance ranges. In the case of FIG. 5(G), the width and pitch are all within the permissible range, so it is cut out correctly and is cut out as is with the lever.

In the case of Figure 5 (H), the widths of area candidates C and d are too short, so the area candidates are corrected. If the width is too short or too wide, it is checked whether the pitches before and after that section are below the minimum allowable range of bits, and if so, they are combined with the M region to form one section. Repeat this action until the width is within the acceptable range. In FIG. 5() I), it is determined that areas C and d are combined into one. Also, Figure 5 (T)
Now, when there are duplicate candidates such as areas c and c', calculate the average value of the width (W) of the other areas, and determine the area that is closer to the average value as the correct area. . In the case of FIG. 5(D), it is determined that area C is the correct area. Characters are cut out in the horizontal direction based on each horizontal region obtained in this way (step 5110).

As in the case of horizontal cutting, create the distribution of the upward component (Step 5ill), extract the upper end position of the one-character area (Step 5112), and create the downward component distribution (Step 5113), of the one-character area. The lower end position is extracted (step S1.14). FIG. 4 (11) shows the distribution of the upward component, and FIG. 4 fI) shows the distribution of the downward component. In this embodiment, since there is one row in the horizontal direction, it is only necessary to perform this on the entire amount entry column 4. It is possible to cut each horizontally cut piece vertically even if the saws are not aligned horizontally. In step 5115, the space between the upper and lower ends of the characters thus formed is determined to be a character area.
), find the height of the character (step S I 1.6
). Then, it is determined whether h1≦character height≦h2 (step 5120), and if it is within the range, the character is cut out in the vertical direction (step 51.22), and the center point of each cut out area is (Step S1.23), and then set a 24×24 pixel area around each center point as a character area (Step 5124). Note that the above step 51
If the height of the character is not within the range of 20, h2, it is determined that the character has failed to be cut out (step 512).
).

Next, the feature amount extracting means 30 extracts a feature amount (hereinafter referred to as a human cover turn feature ball) for each character cut out as described above. Human cover turn characteristics IU (IP,
, [il [jl [k]; However, m is the cutting number, l is the horizontal division number, j is the vertical division number,
The density gradient direction number) is the same as the method used to extract the special Ii quantity of the standard pattern, and the Rohinson operator's mask is used on the image data of the cut out character to extract the density gradient strength and density gradient for each pixel. The direction is determined and binarized, and the character cutout area is divided into 16 4x4 areas (IP, [il
[jl)] and calculate the total number of pixels in each density gradient direction for each area. 128 thus obtained
It is a dimensional vector quantity or a human cover pattern feature quantity. and,
When the extraction of the dog covert turn feature amount is completed for all of the extracted characters, the large classification dictionary 51 is used based on this person covert turn feature fl15fv, and the large classification part 41 and the intermediate neck part 42 Pattern recognition proceeds through three stages: medium classification and detailed classification in the detailed classification section.

That is, first, major classification is performed to distinguish clusters, and the major classification is performed using each cluster standard feature amount C5P, [il [jl
The degree of similarity (Di,, n) with [k] is calculated using the following equation (5), the most similar character (Dl, or the minimum value) is determined, and the character is determined to belong to that cluster.

1stn As mentioned above, the standard patterns in the clusters of the major classification dictionary are additionally registered in the clusters based on the feature values when variations due to rotational shift occur during image input, so they may fail in the major classification. There's nothing to do. Once the cluster to which the extracted character belongs is determined by the major classification, the medium classification section 42 then performs medium classification. The middle classification is the human cover turn feature IP [il [jl [k]] and the standard time trace amount SP of all standard patterns belonging to the cluster, [il
The similarity (DI□X) with [jl [k] is calculated from (5) above.
It is calculated using a formula, and two oil fields that are similar by a predetermined value or more and are most similar (starting from the one with the smallest D: stx) are selected as candidates.

At this time, if there is no similar character (Dl, but below the specified value), it is determined that the character is not in the standard pattern, and the last digit of the number rO~9'' or the day symbol ``¥'', r*, etc. Suppose there is no sign character. Also, if there is only one similar pattern, it is determined to be the standard pattern. Normally, two candidates are extracted, but in this case, it is determined whether two standard patterns are the same character or not. For example, if the standard pattern "0" of the same character type "'ao" and "0" of -0°" are extracted, the most similar standard pattern is used and the character is determined to be "0". This means that the characters are finally "0-9"
", "\J, '*J, and the standard pattern can be either aO" or fO"'.Also, different character standard patterns or 2 When one is extracted, the process moves to detailed classification by the detailed classification section 44.

Detailed classification compares the standard patterns of the two extracted candidates, extracts stable dissimilar areas from the feature values of each category of both patterns, and performs pattern matching only on those areas. be. This will be specifically explained in accordance with the flowchart in FIG. 20 and using the explanatory diagram in FIG. 21.

As shown in FIGS. 21A and 21B, it is assumed that standard patterns x1 and x2 are extracted as two candidates in the intermediate classification. Then, the degree of difference D3rr[il[j] between the extracted standard patterns is calculated as below (
6) Calculate by formula.

1rf −Σ″(SP, + [il [j] [kl −SP,
2 [il [j] [kl) 'x (WT, effect i
]Ijl +WTx2[']i]) −・−・−(6)
The meaning of equation (6) is that the degree of difference in feature amounts between two candidate character standard patterns in a certain area is multiplied by the stability (weight) of the feature amounts in that area. Even when pattern fluctuations occur, the dissimilarity between two candidate characters is expressed as a measure for extracting a high and stable area. Then, (6)
WT, t[Tsukasa[''1] in the equation is obtained at the time of additional registration of the variation pattern in the dictionary creation process. (6)
The 21st example is a specific example of calculating the degree of difference in each area using the formula.
Figure fc), and from among the degrees of difference for each area obtained in this way, three areas with large degrees of difference are extracted.

In the example of FIG. 21(C), D+rr[1][1].

D, fr[++] [1], D+rr[3] [0 is extracted and the diagonal area in the same figure (D) is 4"4. These three areas have very different millet semi-patterns. This is the part that is most affected by fluctuations that occur during image input, so it is possible to more accurately determine which standard pattern the human cover turn is similar to in this area. extracted 3
Similarity AD1stx[! ]
[j] is determined by the following equation (7).

AD+5tx-, I:JIPm[I) [j] [k
l-5P, [il [''] [k] l...
・(7) Then, the sum of the dissimilarities of the three areas is defined as the area similarity u
+5txri] 1. j]. In the example of FIG. 21, each area similarity can be expressed by the following equations (8) and (9).

AD, s, -Jstxt El] [1 + Distxl [
(l [l:l÷Distxl [3] [+]
・・・・・・(8) ADistx2 − D+ttx2[1,] [i]+D+gtx
, [0] cl] + D+1tx2[”] rO]...
...(9) Two candidate standard patterns (X+ and
2) Area similarity A Distxl-ADls
Compare the similarity of tx2 and choose the one that is more similar (AD+st
(the smaller one) is determined to be a large Kabataan character. This human-kabataan character recognition operation is performed for all of the above-mentioned cut out characters, and then the recognition operation is ended.

At this stage, it is possible to recognize the amount information entered in the amount entry column of a check or the like.

In addition, in the above embodiment, the Wart method of cluster analysis is used when creating the major classification time points, but the method is not limited to this, and any method that classifies based on a certain similarity relationship between patterns is fine. . In addition, in the above embodiment, the characteristic IS5 of the pattern within the cluster is used as the cluster feature quantity.
! Either the average of the quantities is used, or the average is used, or one feature of the pattern within the cluster may be used as a representative cluster feature; in short, any feature that represents the pattern within the cluster may be used. good. Further, in the above embodiment, the density gradient strength and the density gradient direction are used as feature quantities, but the present invention is not limited to these, and any feature quantity that can express the feature of the pattern may be used.

Furthermore, in the above embodiment, the fluctuation pattern is created in advance,
By inputting the image as an image, the variable pattern characteristic amount is extracted, or the method is not limited to this method. Alternatively, image data of the fluctuation pattern may be created by performing the following steps, and the fluctuation pattern feature amount may be extracted.

In addition, Fig. 22 shows an example of detecting a postal label. In this case, the concentration gradient strength is also detected using the same method, and the parts where the distribution has the largest change in the left, right, upward, and downward directions are detected. By detecting it, it is possible to reliably compare the area of the label.

Effects of the Invention: As described above, according to the present invention, for character extraction in character pattern recognition, the concentration gradient direction and concentration gradient strength for each @element of a character are determined by the concentration gradient information extraction means, and 2.
After converting into values, the gradient direction along the character cutting direction is added perpendicularly to the character cutting direction to extract the concentration gradient direction distribution in the character cutting direction. The gradient direction along the opposite direction is added perpendicularly to the character cutting direction to extract the concentration gradient direction distribution in the opposite direction to the character cutting direction, and both ends of the character are calculated based on the extracted distribution state. Detected. Since density gradient information in the horizontal and vertical directions is extracted in this way, accurate extraction can be achieved even if the characters are faint.

In addition, as a recognition dictionary, clusters are formed based on the slight amount of changes that take into account fluctuation patterns, and after listing multiple candidates according to major to medium classification, detailed classification is performed based on the fluctuation of character features. There is. Therefore, high-speed and reliable recognition can be achieved. In addition, it is possible to prevent a large increase in dictionary capacity by creating a dictionary that takes into account fluctuation patterns, or by only increasing the character coats in the major classification dictionary and increasing the weight patterns in the detailed classification.

[Brief explanation of the drawing]

Figure 1 is a block configuration diagram showing an example of a character recognition device, Figure 2 (8) is a diagram showing an example of a bill, Figure 2 (B) is a diagram showing an example of a check, and Figure 3 is a recognition dictionary. 4(A)-(I) and FIG. 5 are diagrams showing specific examples of fidgeting images, and FIG. 6(A)-(D
) is a diagram showing an example of a standard pattern, FIG. 7 is a flowchart showing an example of smoothing processing, FIGS. 8 to 10 are diagrams for explaining smoothing, and FIG. 11 is a diagram showing the character cutting method of the present invention. The flowchart shown in Figure 12 is a sign! P
Flowchart showing the extraction operation of the feature quantity, Figures 13 to 15 are diagrams for explaining the extraction operation, Figure 16 is a diagram showing an example of a major classification dictionary, and Figure 17 is for explaining the standard pattern. Figure 18 is a diagram showing an example of a middle classification dictionary,
Fig. 19 is a diagram showing an example of a medium classification dictionary, Fig. 20 is a flow chart showing an operation example of detailed classification, Fig. 21 is a diagram for explaining detailed classification, and Fig. 22 is a diagram showing an example of detecting a postal label. 23 and 24 are diagrams for explaining the prior art. 1.4...Amount entry human rights, 2,5...Information column, 3.6
... Amount information column, 10... Image input means, 11...
- Preprocessing means, 12... Character extraction means, 2o... Pixel component extraction means, 3o... Feature amount extraction means, 40...
- Comparison/verification means, 4]... Major classification section, 42... Intermediate classification section, 44... Detailed classification section, 5o... Storage means. Applicant's agent Yu Yasukata Engraving of Figure 3 I (no change in content) ■ Q (A) Engraving of side β (no change in content) No. d (8) Engraving of drawing C (no change in content) Pregnancy ω (C) Engraving of the drawing (no change in content) Section 4 ([)) Figure 4 (E) Figure 4 (F) M3 M4 M5 Figure 13 (A) (B)
Figure 15 Figure 18 = 8s? ; Go I ECL x σ 0-~ ψt's > superior x -t,, x e E C) mQ- to ψt's 0%4+ ■0-~? 0Φ
t■φ> Φ Φ2; 72
f-9 people also strange 1 customer yen 00-~ginsui ψtono〉脣0~ginsu ψ001 sun1ψtoΦa)
ψto of gsneψ0! ! ■n size ~ size ~ ψ to one gin~tsu C gin~ 7ao 1ri) - ~ of! 0 cost 0 t ε ε E6
E ε j J ＃
w ←−~tsu? 1N mouth complaint 0 ψ
Toe-e −+ −+ −−−= −
←; Tsu 2 work 2 Sagi tomoe 2 ° C shrine ~ ) - no plaster]
−~1 sun0t CD cost approx. 000~~Φ)t1n ψ0+
N octopus? Ink's Kaigin0Q ~ ψNΦ Tsunomi - 1 sun) II
a@e・11115ku-o o o o o a a
CLri+ffσri~IOψsuYCOψΦginff1c+
〜〜〜n-Gingi〜-Shu, Mizube Kotoma 4 is in the main ball Honma Sumuizu Kotoma you are the main character
7! 11 2 H7+31 roll
-3 fake times -4 derogatory times -1 pixel count 1 9; φ's -- ”: '; S S 2 ±i;'
If the right nose base 0-''''g stomach β%B%2 ℃ Units 茗℃Tomoe Shigumo 22 22; 9 = N Q
:8 Tomoe F knee Cg Mi'j:, : I: CC(
p ℃2T 390;Nυir+Wku)fflKiminaSaku11Customer y52+3o+20=':':”'%'%'
%g'X"!"g℃? 299 o C (D: OCo
(+'l to(+ a3 r to 10 q h
M'lU)? (I +O?-1 ward dimension 0 0
! ! ~~ 0~tsu 1 9?
α α −+ Nutoφ’s Q −Ngin? Rido a) glance t goose wo field TI + e −ε ε ε Oε ε 0 to
10-Nnゞ0ゝTomoe? Hiroshi DeshiΦ■O-~Gin? 0t's expenses 00 〜〜 〉−〇 To Chiα-oj Niinij:e−−xxσ Q−Prayer−−
ko;9 snow 9;ρ2♂'ilL %tsu22 standingtt℃γt call 2? (I IQ tenia fi @ O-;
7 Z Z l: ; S % 2! 20 Ntg
%: q%”, T,: S%%~■w
? h(+Φ100ψ04fi+1'1〜〜〜ψ吟−吟〜〜−〇ζ L Dig [/] EI J・Dj//LO] [/]
- Bribery [3 losses, 3 dollars, 21 figures, 22 figures, master position! [-----------11 Yu 23 Vice Tea 24 Diagrams''・□- Sei 3 Ii April Cow Lj Special Correction Office J Kukan A' 1 Display of incident 1990 #j edition 1 head No. 3327 q7 No.: I March's famous ze・Funyukiji C) Kiri・Ko? 17r';, ;+, March 11th and Gu) Sekii system 4"1 Nido: 2.' r:a Iruuchi Higashi Shikikai 7th Sat "7c--1j -Industrial (沫 J(2?-) 14 Attorney--, 7877 Patent Attorney Yasugata Yu Date of the 35th Amendment Order Contents of the 7th Amendment

Claims (1)

[Claims] 1. An image input means for inputting character information, a density gradient information extraction means for determining the density gradient direction and density gradient strength for each pixel of the input image data, and information obtained by the density gradient information extraction means. a first density gradient direction distribution extraction means for extracting a density gradient direction distribution in the character cutting direction by adding the gradient direction along the character cutting direction perpendicularly to the character cutting direction; , a second density gradient that adds the gradient direction along the direction opposite to the character cutting direction perpendicularly to the character cutting direction to extract the density gradient direction distribution in the direction opposite to the character cutting direction. directional distribution extraction means; first character edge detection means for detecting one end of a character based on the distribution state extracted by the first density gradient direction distribution extraction means; and the second density gradient direction distribution. and second character edge detection means for detecting the other edge of the character based on the distribution state extracted by the extraction means, the character edge detected by the first character edge detection means and the second character edge detection means. A character cutting method characterized in that a character region is defined between the character end detected by the character end detecting means and the other character end detected by the character end detecting means.