JP3375292B2 - Character recognition processing method and apparatus and recording medium recording the method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for recognizing character patterns, and more particularly to a method and apparatus for recognizing binary character patterns.

[0002]

2. Description of the Related Art As an example of a conventional character pattern recognition method and apparatus, features necessary for recognition are extracted in the form of a vector from a character pattern that has been binarized and normalized in position and size. For example, a character pattern is divided into coarse local areas, and a character part existing in each local area is observed from coordinate axes in a plurality of directions. And apparatus for recognizing characters by determining the degree of direction contribution (Japanese Patent Application No. 56-46659) "Norihiro Hagita, Seiichiro Naito, Isao Masuda,
"Handwritten Chinese Character Recognition Method Using Global / Local Direction Contribution Density Features", Theological Theory (D), J66-D, No. 6, p
p. 722-729 ”), a method of obtaining a discriminant function such as a distance value or a similarity with a standard pattern vector of each category in a standard dictionary created in advance and using the most similar character category as a recognition result. (Hereinafter, “first method”).

Further, the binary pattern itself of the input pattern is used as an element of the feature vector, and it is used together with the standard pattern vector (represented by a binary pattern like the feature vector) of each category in the standard dictionary created in advance. There is a method in which the most similar character category is used as the recognition result based on the simple similarity, which is an identification function based on the number of black pixels (hereinafter, "second method").

[0004]

In the first method of character pattern recognition, the feature vector elements of the prior art have a wide range of feature quantities reflecting character line structures such as character line directions, connection relationships, and positional relationships. Although it is used, there is a problem that it is difficult to recognize correctly for an image with strong background noise, because these feature values change greatly and the distance value or similarity changes greatly. is there.

Further, the second method of character pattern recognition has a problem that an image with a lot of background noise is erroneously recognized as a category having a large number of black pixels in the vector of the standard pattern. Further, since the determination is made by the number of overlapping pixels, there is a problem that the recognition performance is significantly deteriorated even if the position of superimposition of the input pattern and the standard pattern is slightly shifted.

The present invention has been made in view of the above-mentioned drawbacks, and its problem is to extract a feature required for recognition in the form of a vector from a character pattern whose binarization, position and size are normalized. Then, in the method that finds the discriminant function such as the distance value or the similarity with the standard pattern vector of each category in the standard dictionary created in advance and makes the most similar character category the recognition result, When there is noise in the background, the feature value extracted from the noise is superimposed on the feature value extracted from the character line in the local area where the noise exists, so the value of the feature vector obtained from the local area fluctuates greatly. As a result, the distance value or the similarity value obtained by the conventional discriminant function greatly changes, making it difficult to correctly recognize the character type. Tsu and thus solve the problem, even in situations where background noise as described above exists, it is to provide a correctly recognizing method and apparatus letters.

[0007]

In the character recognition processing method of the present invention for solving the above problems, a character pattern cut out for each character and binarized is normalized with respect to position and size. Processing is performed, the normalized character pattern is divided into local regions, and each feature is divided into feature vector vectors.
It performs a process of extracting a form, also separately from the processing of extracting the feature, the and the local region the normalized character pattern
The number of black pixels existing in the local region is measured for each local region of the character pattern divided by the dividing process , and the number of black pixels determined by the measurement is calculated in advance. A black pixel number difference value is calculated from the created standard number of black pixels of each category of the standard dictionary of black pixels, and the extracted feature vector is created in advance for each local region. The distance value or the similarity value obtained with the standard feature vector of the corresponding category of the feature standard dictionary is adjusted according to the magnitude of the black pixel number difference value, and the adjusted value of all local regions is adjusted. It is characterized in that the most similar character category is used as the recognition result of the character pattern by accumulating and using the distance value or the similarity.

Alternatively, for each of the local regions , the extracted
The distance value or the similarity value obtained between the generated feature vector and the standard feature vector of the corresponding category of the feature standard dictionary created in advance is adjusted according to the magnitude of the black pixel number difference value. In the processing stage, when the black pixel number difference value is 0, the distance values or the similarities calculated for all the dimensions of the feature vector are accumulated, and when the black pixel number difference value is larger than 0, It is characterized in that the distance value or similarity of each dimension of the feature vector is divided by the difference value of the number of black pixels and accumulated, and the accumulated value is used as the adjusted distance value or similarity of the local region.

Alternatively, for each of the local regions , the extracted
The distance value or the similarity value obtained between the specified feature vector and the standard feature vector of the corresponding category of the feature standard dictionary created in advance is adjusted according to the magnitude of the black pixel number difference value. In the processing step, the distance value or similarity of each dimension of the feature vector is subtracted by the black pixel number difference value, and the subtracted distance value or similarity when the subtracted distance value or similarity is 0 or more. The degree is accumulated, and the accumulated value is used as an adjusted distance value or similarity of the local area.

Further, a program for causing a computer to execute the processing steps in the above character recognition processing method is recorded in a computer-readable recording medium.

Similarly, the character recognition processing apparatus of the present invention for solving the above problem is a preprocessing for normalizing the position and size of a character pattern that is cut out for each character and binarized. Means, and the normalized character pattern is divided into local regions and each feature is a feature vector
Feature extracting means for extracting in the form of, the normalized character pattern is divided similarly to the local region, wherein each local region of the divided character pattern, the number of black pixels present in the local region A black pixel number extraction means for measuring the number of black pixels, and a black pixel number difference value between the black pixel number obtained by the measurement and the standard black pixel number of each category of the black pixel number standard dictionary created in advance. A distance value or a similarity is calculated between the black pixel number difference value calculating means and the extracted feature vector and the standard feature vector of the corresponding category of the feature standard dictionary created in advance for each local region. and the distance values or similarity calculation means, for each said local region, the value of the calculated distance values or similarity, and adjusting means for adjusting in accordance with the size of the black-pixel count difference value, all the local regions the adjusted distance value or The use by accumulating acetonide,
And an identification unit that uses the most similar character category as a recognition result of the character pattern.

According to the present invention, the character pattern is divided into local areas, the number of black pixels existing in each local area is measured, and each of the categories of the black pixel number standard dictionary created in advance is measured. When calculating the distance value or the similarity between the feature vector obtained from the local region of the input pattern and the standard feature vector of the corresponding category of the feature standard dictionary, the difference value from the standard black pixel number is calculated. A value proportional to the black pixel number difference value is adjusted to reduce the distance value or the similarity value obtained between the local areas, so that the distance value obtained from the local area containing a lot of noise or A solution to the problem is to reduce the variation of the similarity and reduce the variation of the distance value or the similarity obtained from the entire character pattern.

The feature of the present invention is that in the character recognition method using the conventional feature, the black pixel number difference value calculation process / means, the black pixel number difference dictionary normalization process, the black pixel number difference value normalization process / means, the distance value. -The point is that the similarity adjustment processing / means are used together. By using this means / processing together, a feature vector due to noise (dirt etc.) locally superimposed on the divided areas, which has occurred when recognizing the characteristics of the character for each area where the conventional character to be recognized is divided The negative impact on
Even in such a case, the character recognition rate is improved.

[0014]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings.

[Embodiment 1] FIG. 1 is a block diagram for explaining a first embodiment of a character recognition method and apparatus according to the present invention.

The preprocessing section 1-2 calculates the width and height of the input character pattern 1-1 by using, for example, a position normalization processing method that has been known so far. Is calculated, and the translation processing of the entire input character pattern is performed so that the center comes to the center position of the character frame. Further, for example, by using a conventionally known normalizing method of a size, an input character pattern is enlarged / reduced so that the width and height of the character become the same as the width and height of the character frame. I do.

FIG. 2 shows an example in which the input character pattern is normalized by the normalization process of the preprocessing unit in the character "pressure". FIG. 2- (a) is an example of the input character pattern 1-1. FIG. 2B is a character pattern after the position and size of the input character pattern 1-1 are normalized by the preprocessing unit 1-2.

The feature extracting unit 1-3 uses the character pattern normalized by the pre-processing unit 1-2 to identify the features necessary for recognition such as the direction of the character lines, the connection relation, and the positional relation in the form of a vector. The process of extracting is performed.

FIG. 3 is a block diagram showing an example of the configuration of a device that implements the feature extracting section 1-3. 1-3 here
Reference numeral -1 is a feature value extraction unit that extracts a feature value required for recognition from the character pattern, and 1-3-2 is a feature value calculation unit that calculates the feature value in the form of a vector.

The black pixel number extraction unit 1-4 includes a preprocessing unit 1-2.
In, the character pattern subjected to the normalization process is input, the character pattern is divided into coarse local regions, and the process of measuring the number of black pixels included in the local region from each local region is performed.

FIG. 4 is a block diagram showing an example of the configuration of an apparatus for implementing the black pixel number extraction unit 1-4. Where 1
4-1 is a character pattern division unit that divides a character pattern into a plurality of coarse local areas, and 1-4-2 is a pixel detection unit that detects a pixel existing in the local area and determines whether the pixel is white or black.
1-4-3, when the pixel detection unit determines that the pixel is a black pixel,
A black pixel number measuring unit for counting the number 1-4-4 is a black pixel number calculating unit for outputting the number of black pixels in the local region.

The identification unit 1-5 forms the main part of the present invention. The value of the feature vector obtained by the feature extraction unit 1-3 and the number of black pixels obtained by the black pixel number extraction unit 1-4. Using the feature standard dictionary 1-6 for each character and the black pixel number standard dictionary 1-7 that have already been created, the feature vector and the standard feature of each category of the feature standard dictionary 1-6 are used. In the calculation for each local area in the discriminant function of the distance value or the similarity degree that has been already known between the vector and the vector, the number of black pixels in each local area of the input pattern and the black pixel number standard dictionary 1- 7, the difference value of the standard black pixel number of each category is calculated, and the distance value or the similarity value calculated for each local area is reduced according to the magnitude of the black pixel number difference value. By performing processing on all local regions By obtaining the distance values or similarity between the category of the input character pattern and the standard dictionary, and identifies the character pattern.

FIG. 5 is a block diagram showing an example of the configuration of a device that implements the identifying unit 1-5. Where 1-5-1
Between the feature vector obtained from the input pattern and the standard feature vector of each category of the feature standard dictionary 1-6 already stored, using the distance value or the similarity degree already known, The distance value / similarity calculation unit 1-5-2, which calculates for each local region, includes the number of black pixels obtained from the input pattern for each local region and the black pixel number standard dictionary 1- The black pixel number difference value calculation unit that obtains the difference value of the black pixel number from the standard black pixel number of each category of 7; 1-5-3 calculates the black pixel number difference value by the distance value / similarity calculation. Black pixel number difference value normalization unit that normalizes to a numerical value according to the distance value or similarity used in the unit 1-5-1, 1-5
-4 is for the distance value or similarity for each local region,
A distance value / similarity adjustment unit that adjusts the distance value or the similarity value based on the normalized difference value of the number of black pixels.

In the distance value / similarity adjustment unit 1-5-4, the normalized black pixel number difference value obtained in the black pixel number difference value normalization unit 1-5-3 and the distance value / similarity calculation unit 1 Input the distance value or the degree of similarity obtained in -5-1 and output the value obtained by reducing the value of the distance value or the degree of similarity by the normalized difference value of the number of black pixels by using subtraction or subtraction. By doing so, it is an object to obtain the distance value or the similarity according to the value of the difference between the number of black pixels of the input pattern and the number of standard black pixels of the standard black pixel dictionary in each local area.

FIG. 6 shows a flowchart of the processing in the distance value / similarity adjusting unit 1-5-4. FIG. 6 shows a processing example in which division is used as the distance value adjustment method. In this example, when the normalized black pixel number difference value is 0, the distance value or the similarity for each feature vector dimension obtained by the conventional method is accumulated in all dimensions, and the accumulated value is directly used for each local region. Output as the distance value of. When the normalized black pixel number difference value is larger than 0, the distance value or the similarity degree of the conventional method obtained for each feature vector dimension is divided by the normalized black pixel number difference value, and is divided for all dimensions. The distance value or the similarity is accumulated, and the accumulated distance value or the similarity is output as the distance value or the similarity in each local area.

Returning to FIG. 5, 1-5-5 is a distance value / similarity calculation in which the distance value or similarity for each local region is added to calculate the distance value or similarity for each character in the standard dictionary. Department,
1-5-6 is a sorting unit that sorts the results in the order of the character type having the highest similarity from the distance value or the similarity value obtained from all the character types, and 1-5-7 outputs the obtained identification result. It is an identification result output unit.

Next, as a specific embodiment of the character recognition processing method and apparatus of the present invention, as a characteristic for recognition, a character pattern is divided into coarse local regions and black pixels in each local region are predetermined. In the case of multiple directions, for example, 8 directions, 0 °, 45 °, 90 °, 135 °, 1
80 °, 225 °, 270 °, 315 ° (1, respectively,
The number of black pixels connected in each direction is counted, and the distribution status of each direction component of the black pixels is represented. A character pattern is identified by calculating the distance value by dividing the distance contribution value (Japanese Patent Application No. 56-46659) by the value corresponding to the difference in the number of black pixels using the Euclidean distance as the identification function. A case will be described.

FIG. 7 is a diagram showing a flow chart for explaining it.

First, the input character pattern (step S1) cut out as an area for each character is processed by the preprocessing section 1-2.
Sent to. The preprocessing unit 1-2 normalizes the position and size of the input character pattern (step S2). The N × N mesh normalized character pattern (step S3) obtained by the pre-processing is the character pattern of the feature value extraction unit 1-3-1 of the feature extraction unit 1-3 and the black pixel number measurement unit 1-4. It is sent to the dividing unit 1-4-1.

The character pattern division unit 1-4-1 of the black pixel number extraction unit 1-4 equally divides the normalized character pattern into K coarse local regions, for example, square local regions (step S4). In order to calculate the number of black pixels in each local area of the locally divided character pattern, the character pattern divided into local areas is sent to the pixel detection unit 1-4-2 (step S5).

A concrete processing flow for calculating the number of black pixels in each local area is shown in FIG. The pixel detection unit 1-4-2 is
Pixels in each local area in the character pattern divided into local areas are detected (step S5-1). If the detected pixel is a black pixel (step S5-2), the counter for the number of black pixels is incremented by 1 (step S5-3). The detection and the number of black pixels are counted for all the pixels in the local area, and the number of black pixels in the local area is obtained (step S5-4). The number of black pixels obtained as described above is sent to the black pixel number difference value calculation unit 1-5-1 of the identification unit 1-5.

Feature value extraction unit 1-3 of feature extraction unit 1-3
1 is a character pattern division unit 1-of the black pixel number extraction unit 1-4.
The character pattern is divided by the same method as in 4-1 (step S6). Next, a feature value is calculated in each local area of the character pattern divided into local areas (step S7).

FIG. 9 shows a specific processing flow for calculating the feature value in each local area. The feature value extraction unit 1-3-1,
Pixels in each local area in the character pattern divided into local areas are detected (step S7-1). If the detected pixel is a black pixel (step S7-2), the counter for the number of black pixels is incremented by 1 (step S7-3). Next, a process for obtaining the direction contribution is performed on the detected black pixel (step S7-4).

A concrete processing flow for calculating the direction contribution in each pixel is shown in FIG. Feature value extraction unit 1-3
1 uses the detected pixel as a reference point (step S7-4-
1) Extend the tentacle in each direction to detect adjacent pixels (step S7-4-2). If the pixel adjacent in the scanning direction is black, the connection length counter is incremented by 1 (step S7).
-4-3), using the adjacent pixel as a new reference point (step S
7-4-4), and the scanning process is repeated. The scanning process is
It is performed for the entire normalized character pattern without being limited to the pixels in the block in which the detected pixel exists.

If the adjacent pixel is a white pixel or there is no adjacent pixel, the scanning is ended (step S7-4-5).
The above processing is performed for all eight directions (step S7-4).
-6). The black pixel connection length cumulative value is obtained from the black pixel connection lengths in the eight directions obtained for each black pixel, for example, by using the simple sum or the square root of the sum of squares (step S7-4-).
7). The direction contribution degree is obtained by dividing the black pixel connection length in each direction by the black pixel connection length cumulative value (step S7-4-8).

The direction contribution f of each black pixel is f = (α1, α2, α3, α4, α5, α6, α7, α
8) is represented by the following eight-dimensional vector. Where α1, α2,
.. .alpha.8 is a direction contribution component in each of eight directions, and using black pixel connection lengths li (i = 1, 2, ..., 8) obtained by extending the tentacles from the black pixels in eight directions, As an example, when the square root of the sum of squares is used as the black pixel connection length cumulative value, it is expressed by αi = li / √ (Σ _{j = 1} ⁸ lj ² ).

The directional contribution f thus obtained is obtained for all black pixels in each local area and accumulated for each direction (step S7-5). The accumulated direction contribution value and the number of black pixels are sent to the characteristic value calculation unit 1-3-2. The feature value calculation unit 1-3-2 averages the accumulated values of the direction contribution according to the number of black pixels in each local region to calculate the feature value in each local region (step S7-6). The feature value fk obtained in the k-th (1, 2, ..., K, ..., K) local region is represented by fk = (αk1, αk2, ..., αk8). Here, αk1, αk2, ..., αk8 are
The elements of the direction contribution vector obtained by accumulating the direction contribution vectors of all the black pixels existing in the k-th local area for each direction component are averaged by the number of black pixels.

The feature vector f (step S8) of the character pattern in each local area thus represented is
It is sent to the distance value / similarity calculation unit 1-5-1 of the identification unit 1-5.

FIG. 11 shows a specific processing flow for calculating the distance value in each local area.

The black pixel number difference value calculation unit 1-5-2 calculates the number of black pixels sent from the black pixel number extraction unit 1-4 and the number of standard black pixels of each category of the black pixel number standard dictionary 1-7. In between, the black pixel number difference value is calculated (step S9-1). The black pixel number difference value is sent to the black pixel number difference value normalization unit 1-5-3. The black pixel number difference value normalization unit 1-5-3 normalizes the black pixel number difference value to a value matched with the Euclidean distance (step S9-2). The number of black pixels of the input pattern in the k-th local area is ak, and the number of black pixels is in the standard dictionary 1-7
The number of standard black pixels in each category i (1 ≦ i ≦ M) of
Then, the black pixel number difference value C1k is represented by C1k = (ak-bik).

As an example of the normalization of the black pixel number difference value C1k according to the Euclidean distance, the normalized black pixel number difference value C2k is C2k = 1 when C1k <= 0, and C2k = when C1k> 0. (C1k / W) +1 (1 <= W <= the number of pixels included in the local area). The normalized black pixel number difference value is sent to the distance value / similarity degree adjusting unit 1-5-4.

The distance value / similarity calculation unit 1-5-1 combines the feature vector sent from the feature extraction unit 1-3 and the standard feature vector of each category of the feature standard dictionary 1-6 created in advance. In between, the square value of the difference is calculated (step S9-3). The above processing is performed for all eight directions (step S9-4). The feature vector of the input character pattern in the kth local region is fk = (α1k, α2k, α
k3, αk4, αk5, αk6, αk7, αk8), and the standard feature vector of each category i (1 ≦ i ≦ M) of the feature standard dictionary 1-6 is si = (βk1, βk2, βk3, βk
4, βk5, βk6, βk7, βk8), the squared value of the difference between the feature values is (αk1-βk1) ² , (αk2-βk2) ² , ..., (α
It is represented by k8−βk8) ² . The calculation result is the distance value / similarity adjusting unit 1-5.
-4.

The distance value / similarity adjustment unit 1-5-4 uses the square value of the difference sent from the distance value / similarity calculation unit 1-5-1 for the black pixel number normalization unit for all eight directions. It is divided by the normalized number of black pixels sent from 1-5-2 and accumulated (step S9).
-5). A value accumulated for all eight directions is calculated as a distance value (step S9-6). The value obtained by dividing the squared value of the difference for each direction by the number of normalized black pixels is (αk1-βk1) ² / C2k, (αk2-βk2) ² / C2k, ..., (αk8-βk8) ² / C2k ... 1) and the value Dik obtained by accumulating the above values is: Dik = (αk1-βk1) ² / C2k + (αk2-βk2) ² / C2k + ... + (αk8-βk8) ² / C2k (2) Can be expressed as

The distance value calculated for each local area is sent to the distance value / similarity calculator 1-5-5. The distance value / similarity calculation unit 1-5-5 accumulates the distance values calculated from all the local regions, and sends them to the sorting unit 1-5-6 as the distance value in each category (step S10).

The distance value Di obtained between the input character pattern and each category i (1≤i≤M) of the feature standard dictionary 1-6 is the distance value obtained from all the 1st to Kth local areas. Cumulatively, it can be expressed as Di = Di1 + Di2 + ... + Dik + ... + DiK.

In the sorting unit 1-5-6, the distance values of all categories obtained by performing the above series of processing for all categories of the standard dictionary are sorted in ascending order (depending on other distance values / similarities. Rearrange (in descending order) (step S11). The sorted result is the identification result output unit 1
-5-7.

The identification result output unit 1-5-7 outputs the character having the smallest distance value (in order of increasing distance value / similarity) as an identification result (step S1).
2).

In the above description, the distance value / similarity adjusting unit 1-
In 5-4, as a method of adjusting the distance value / similarity, the distance value is divided by the normalized black pixel number difference value.
Any other method that can reduce the distance value / similarity according to the black pixel number difference value, such as subtraction, can be applied.

The effect of this embodiment example in a specific character pattern will be described with reference to FIG. 17- (a) shows a normal character pattern, and FIG. 17 (b) shows an example of a character pattern having noise in the background. K-th local area 17- (a)-
k, 17- (b) -k, when the direction contribution feature described in the embodiment is used as the feature, 17- (a)
The characteristic value Fa obtained from −k and the characteristic value Fb obtained from 17- (b) −k are shown in Table 1, and the number of black pixels B in each local region is
a and Bb are shown in Table 2, and the distance value D when the conventional Euclidean distance D1 as the discriminant function and the method of the present embodiment is used
2 is shown in Table 3.

[0050]

[Table 1]

[0051]

[Table 2]

[0052]

[Table 3]

From Table 1, it can be seen that the feature amount obtained when noise is present varies from the feature amount obtained when no noise is present. From Table 3, in the above situation, the Euclidean distance, which is the conventional discriminant function, gives a value as large as 120/1000 as the distance value obtained between local regions, but the method of the present invention gives 8/100.
It can be seen that the variation of 0 and the distance value is suppressed. By using the distance values obtained from all the local areas by the same process below, the original Euclidean distance is the area where noise is not present, and the distance that is reduced according to the amount of noise is the local area where noise is present. By obtaining the value, the variation in the distance value obtained from the entire character pattern can be reduced even when noise is present in the background, and correct recognition can be achieved.

[Embodiment 2] As a second embodiment of the present invention, the same characteristic as that of the first embodiment is used, and the difference in the number of black pixels is obtained by using the city block distance as the discrimination function. When the character pattern is identified by subtracting the distance value by the value according to and calculating the distance value,
FIG. 12 shows a specific processing flow of calculating the distance value of each local area.
Shown in.

The black pixel number difference value calculation unit 1-5-2 calculates the black pixel number sent from the black pixel number extraction unit 1-4 and the standard black pixel number of each category of the black pixel standard dictionary 1-7. In between, a black pixel number difference value is calculated (step S9-11). The black pixel number difference value is sent to the black pixel number difference value normalization unit 1-5-3. The black pixel number difference value normalization unit 1-5-3 normalizes the black pixel number difference value to a value matched with the city block distance (step S9-12). For example, the number of black pixels of the input pattern in the k-th local area is dk, and the number of black pixels of each character i (1 ≦ i ≦ M) of the black pixel number standard dictionary 1-7 is ei.
If k, the black pixel number difference value G1k is expressed by G1k = (dk-eik).

As an example of the normalization of the black pixel number difference value G1k according to the city block distance, the normalized black pixel number difference value G2k is: G1k <= 0 G2k = 0 G1k> 1 G2k = G1k / V (1 <= V <= the number of pixels included in the local region). The normalized black pixel number difference value is sent to the distance value / similarity degree adjusting unit 1-5-4.

The distance value / similarity calculation unit 1-5-1 is arranged between the feature vector sent from the feature extraction unit 1-3 and the feature standard vector of each category of the feature standard dictionary 1-6 created in advance. Then, the absolute value of the difference is calculated (step S9).
-13). The above processing is performed for all eight directions (step S9-14). The feature vector of the input character pattern in the k-th local area is hk = (γ1k, γ2k, γ
k3, γk4, γk5, γk6, γk7, γk8) and the feature vector of each character i (1 ≤ i ≤ M) of the feature standard dictionary 1-6 is ti = (δk1, δk2, δk3, δk4, δk).
5, δk6, δk7, δk8), the absolute value of the difference between the feature values is | γk1-δk1 |, | γk2-δk2 |, ..., | γk
It is represented by 8-δk8 |. The calculation result is the distance value / similarity adjusting unit 1-5.
-4.

The distance value / similarity adjustment unit 1-5-4 uses the absolute value of the difference sent from the distance value / similarity calculation unit 1-5-1 for all eight directions in the black pixel number normalization unit 1 The absolute value of the value subtracted by the normalized black pixel number sent from -5-2 is accumulated (step S9-15). A value accumulated for all eight directions is calculated as a distance value (step S9-16).

The value obtained by subtracting the absolute value of the difference for each direction by the number of normalized black pixels is: || γk1-δk1 | -G2k |, || γk2-δk2 | -G2k |, ..., || γk8-δk8 | -G2k | ... can be expressed by equation (3), and the value Jik obtained by accumulating the above values is Jik = (|| γk1-δk1 | -G2k |) + (|| γk2-δk2 | -G2k |) + ... + (|| k8- [delta] k8 | -G2k |) Equation (4) is obtained, and the distance value in each local region is obtained.

Distance value / similarity adjusting unit 1 in the present embodiment
-5-4 uses subtraction as a distance value / similarity adjustment method. A specific processing example in this case is shown in the flowchart of FIG. In this example, the subtraction is performed between the distance value or the similarity degree obtained by the conventional method for each feature vector dimension and the normalized black pixel number difference value. When the obtained distance value or similarity is 0 or more, the subtracted distance value or similarity is accumulated. When the obtained distance value or similarity is less than 0, the distance value or similarity is not accumulated and the calculation of the next dimension is started. Do the above for all dimensions,
The cumulative value finally obtained is output as a distance value or similarity in each local region.

In the above description, the distance value / similarity adjusting unit 1-
In 5-4, as a method of adjusting the distance value / similarity, the distance value is subtracted by the normalized black pixel number difference value,
Of course, any other method, such as division, that can reduce the distance value / similarity according to the black pixel number difference value can be applied.

In FIG. 14, the character pattern is composed of K local areas 14-1, 14-2, ..., 14-k, ...
The figure when it divides into 4-K is shown.

FIG. 15 shows eight directions (15−15) as the directions for extending the tentacles in order to obtain the black pixel connection length of the character pattern.
1, 15-2, 15-3, ..., 15-7, 15-8).

FIG. 16 shows the k-th local area 1 of FIG.
In order to obtain the direction contribution in a 4-k black pixel,
It shows how the tentacles are extended to obtain the black pixel connection length.

Table 4 shows the effects of the second embodiment of the present invention. Table 4 uses the same characteristics as those of the present embodiment as the characteristics of the conventional method, uses the city block distance (conventional method 1) and the weighted city block distance (conventional method 2) as the discrimination function of the conventional method, and uses the feature dictionary. The average value and the standard deviation of each category are used as the input patterns, and the top 1 and 2 of the 3449 patterns that include a lot of background noise as the input pattern.
The cumulative classification rates of the 5th, 10th, and 10th ranks are calculated.

[0066]

[Table 4]

For a pattern containing a lot of background noise,
Since the conventional method 2 using the weighted city block distance cannot deal with the fluctuation of the feature value due to noise, the recognition rate is lower than that of the conventional method 1 using the city block distance. However, the distance value depends on the noise amount. According to the second embodiment of the present invention that performs the adjustment, it is possible to effectively suppress the fluctuation due to noise, so that it is possible to significantly improve the classification rate as compared with the method using the conventional discriminant function.

The feature of the present invention described above is that the method using the direction contribution density feature which is a conventional character recognition method (distance value / similarity calculator 1-5-1, feature standard dictionary) is used. 1-6, a method using the distance value / similarity calculation unit 1-5-5 as a constituent element), the black pixel number difference value calculation unit 1
-5-2, the black pixel number standard dictionary 1-7, the black pixel number difference value normalization unit 1-5-3, and the distance value / similarity degree adjustment unit 1-5-4 are used together. By using these together, noise that is locally superimposed on the divided areas that occurred when recognizing the character features for each area where the character for recognition of the direction contribution density feature was divided in the conventional way (dirt etc.) It is possible to suppress the adverse effect on the feature vector due to, and to improve the character recognition rate in such a case.

To further supplement the above-mentioned features of the present invention, the fact that the divided areas have a lot of noise means that the black pixel number standard dictionary 1-7 and the divided areas are defined by the black pixel number difference value calculation unit 1- It is judged by comparing in 5-2. If it is judged to be noisy,
The adjustment processing is performed in the black pixel number difference value normalization unit 1-5-3, the difference value C2k is obtained as an output, and C2k is input to the distance value / similarity degree adjustment unit 1-5-4. The central processing of the present invention is the processing of the distance value / similarity adjusting unit 1-5-4. As a specific method of the processing, when division is used as the distance value adjusting method shown in the first embodiment (FIG. 6), and subtraction is used as the distance value adjusting method shown in the second embodiment. May be used (FIG. 13). Details of “division of distance value (or similarity) of each dimension by normalized black pixel number difference value” in FIG. 6 are described in Expression (1). For details of “accumulate divided distance values” in FIG.
It is described in. The conventional method is used for the process of “accumulating distance values of each dimension” on the right side in FIG. Further, details of “subtraction of distance value (or similarity) of each dimension by normalized black pixel number difference value” in FIG. 13 are described in Expression (3). Figure 1
For details of “accumulate divided distance values” in 3, refer to equation (4).
It is described in.

It should be noted that some or all of the means shown in FIGS. 1, 3, 4, and 5 can be made to function using a computer, or the steps of the processing shown in FIGS. Needless to say, the program can be executed by a computer, or a program that causes the computer to function as the means or a program that causes the computer to execute the steps of the process can be read by the computer. (Floppy disk), MO, ROM, memory card, CD, DVD, removable disk, etc. can be recorded, provided, and distributed.

[0071]

As described above, according to the present invention,
The number of black pixels in each local area is measured, the difference from the number of black pixels in a standard dictionary created in advance is obtained, and the value of the distance value or the similarity is calculated using the difference value of the number of black pixels. By adjusting, if there is noise in the background, the feature value will change greatly depending on the amount of noise, and the distance value will change greatly accordingly, whereas the distance will change depending on the amount of noise. The value can be reduced, and by reducing the variation of the distance value obtained from the entire character pattern, it becomes possible to correctly recognize even a pattern that would be erroneously recognized by the conventional identification function.

Further, in the present invention, when no noise is present, the same distance value / similarity as that of the conventional discriminant function can be obtained by normalizing the appropriate difference value of the number of black pixels. It is possible to obtain the same discrimination ability as the conventional one without adversely affecting a character pattern that does not exist.

Further, the present invention is not particularly limited to a peculiar feature extraction method, and can be applied as long as it is a method of dividing a character pattern into coarse local regions and extracting a feature used for recognition from each of the local regions. Therefore, it is applied to various feature extraction methods that have been proposed as character recognition methods,
The performance of various methods can be improved by making it possible to recognize a character pattern containing noise.

In addition, in the conventional method of recognizing the degree of overlap of black pixels of a character pattern, the position of the pixels at the time of superposition may be slightly displaced, or the character font may be different,
Although the recognition performance is significantly deteriorated, in the present invention, by using the method of performing the feature extraction from the character pattern, the recognition performance is not so much improved even with respect to the pixel position shift in the range of the coarse local area and the character font difference. It is possible to recognize a pattern in which noise exists without reducing the pattern.

[Brief description of drawings]

FIG. 1 is a first part of a character recognition method and device according to the present invention.
It is a block diagram which shows the example of embodiment.

FIGS. 2A and 2B are diagrams showing a normalization process in a preprocessing unit according to the first embodiment.

FIG. 3 is a block diagram illustrating an example of a configuration of an apparatus that implements a feature extraction unit of the first exemplary embodiment.

FIG. 4 is a block diagram illustrating an example of a configuration of an apparatus that implements the black pixel number extraction unit of the first embodiment.

FIG. 5 is a block diagram illustrating an example of a configuration of an apparatus that implements the identification unit according to the first exemplary embodiment.

FIG. 6 is a diagram showing a processing example in the case where division is used as a distance value adjustment method in the distance value / similarity adjustment unit of the first embodiment.

FIG. 7 is a flowchart for explaining processing in the first embodiment example.

FIG. 8 is a flowchart for explaining a specific processing example of calculating the number of black pixels in a local area in the first embodiment.

FIG. 9 is a flowchart for explaining a specific processing example of feature value calculation in each local area in the first embodiment.

FIG. 10 is a flowchart for explaining a specific example of processing for calculating the direction contribution in each pixel in the first embodiment.

FIG. 11 is a flowchart for explaining a specific process of calculating a distance value in each local area in the first embodiment.

FIG. 12 is a flowchart for explaining a specific processing example of distance value calculation of each local area in the second embodiment of the present invention.

FIG. 13 is a diagram showing a processing example when subtraction is used as the distance value adjustment method in the distance value / similarity adjustment unit of the second embodiment.

FIG. 14 is a diagram showing how a character pattern is divided into coarse local regions in a feature extraction unit according to an embodiment of the present invention.

FIG. 15 shows a direction in which a tentacle is extended in order to obtain a black pixel connection length in the feature extraction unit according to the embodiment of the present invention.
It is a figure which shows the case where it makes 8 directions.

FIG. 16 is a diagram showing how a feature pixel extraction unit obtains a black pixel connection length according to the embodiment of the present invention.

17A and 17B are diagrams for explaining a character pattern when noise is present in the background as an effect of the present invention.

[Explanation of symbols]

1-1 ... Input character pattern 1-2 ... Preprocessing unit 1-3 ... Feature extraction unit 1-4 ... Black pixel number extraction unit 1-5 ... Identification processing unit 1-6 ... Feature standard dictionary 1-7 ... Black pixels Number standard dictionary 1-8 ... Identification results 14-1, 14-2 to 14-k to 14-K ... Local areas 15-1 and 15- when character patterns are divided into coarse K local areas 2, 15-3, 15-4, 15-5,
15-6, 15-7, 15-15 ... Direction to extend tentacles

Continuation of the front page (56) References Japanese Unexamined Patent Publication No. 10-27215 (JP, A) Japanese Unexamined Patent Publication No. 5-166008 (JP, A) Japanese Unexamined Patent Publication No. 11-96302 (JP, A) Image telop that is robust against deterioration Character Recognition, IEICE Technical Report PRMU98-154, Japan, December 18, 1998, Vol. 98 No. 490, pp. 33-40 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (5)

(57) [Claims] 1. A character pattern cut out for each character and binarized is subjected to normalization processing for position and size, and the normalized character pattern is divided into local regions.
Characterized performs processing for extracting in the form of feature vectors of and separately from the processing of extracting the feature performs a process of dividing the normalized character pattern as with the local region, in the process of the divided For each local region of the divided character pattern, the number of black pixels existing in the local region is measured, and the number of black pixels obtained by the measurement and the number of black pixels in the standard dictionary created in advance are calculated. A process of calculating a black pixel number difference value between the standard black pixel number and the standard black pixel number is performed, and the extracted feature vector and the standard feature vector of the corresponding category of the feature standard dictionary created in advance are calculated for each local region. The distance value or the similarity value calculated between the two is adjusted according to the magnitude of the black pixel number difference value , and the adjusted distance value or the similarity value of all local regions is accumulated.
A character recognition processing method characterized in that the most similar character category is used as a recognition result of the character pattern by being accumulated and used. 2. A distance value or a similarity value obtained between the extracted feature vector and a standard feature vector of a corresponding category of a feature standard dictionary created in advance for each of the local regions, At the stage of the process of adjusting according to the magnitude of the black pixel number difference value, when the black pixel number difference value is 0, the distance values or the similarities calculated for all the dimensions of the feature vector are accumulated, and When the difference value of the number of black pixels is larger than 0, the distance value or similarity of each dimension of the feature vector is divided by the difference value of the number of black pixels and accumulated, and the accumulated value is adjusted in the local region. The character recognition processing method according to claim 1, wherein the distance value or the similarity is used. Wherein for each of the local region, the distance values or similar determined between the standard feature vector of the corresponding category of the features standard dictionary that has been created in advance with the extracted serial JP <br/> FEATURES vector In the process of adjusting the degree value according to the magnitude of the black pixel number difference value, the distance value or similarity of each dimension of the feature vector is subtracted by the black pixel number difference value, and the subtraction is performed. 2. The distance value or the degree of similarity subtracted when the distance value or the degree of similarity is 0 or more is accumulated, and the accumulated value is used as the adjusted distance value or the degree of similarity of the local region. Character recognition processing method described. 4. Preprocessing means for normalizing the position and size of a character pattern cut out for each character and binarized, and the normalized character pattern is divided into local regions. Each
Feature extracting means for extracting a feature in the form of a feature vector, the normalized character pattern is divided similarly to the local region, for each local region of the divided character pattern here,
A black pixel number extraction means for measuring the number of black pixels existing in the local area; and the number of black pixels obtained by the measurement and the standard number of black pixels of each category of the standard dictionary of black pixels created in advance. Between the black pixel number difference value calculating means for obtaining the black pixel number difference value between the two, and between the extracted feature vector and the standard feature vector of the corresponding category of the feature standard dictionary created in advance for each local region. A distance value or similarity calculating means for calculating a distance value or similarity , and adjusting the calculated distance value or similarity value for each local region according to the magnitude of the black pixel number difference value. and adjusting means for, the adjusted distance values or similarity of all local regions cumulative
A character recognition processing device comprising: an identification unit that uses the most similar character category as a recognition result of the character pattern by accumulating and using. 5. A program for causing a computer to execute the processing steps in the character recognition processing method according to claim 1, is recorded on a computer-readable recording medium. A recording medium on which a characteristic character recognition processing method is recorded.