JPH0562019A - Character recognizing device - Google Patents

Abstract

(57) [Summary] [Purpose] Accurately recognize characters with a high recognition rate closer to human judgment. An appearance frequency calculation unit 11 calculates an appearance frequency of each pixel of a pattern portion and a background portion of each character, and at the appearance frequency, each pixel of a pattern portion and a background portion of each character is calculated. Each appearance frequency of the portion corresponding to is weighted, weighting values obtained by assigning different signs to the weighting values of the pattern portion and the background portion are calculated by the weighting unit 13 and the normalizing unit 15, and the weighting values are Each character stored corresponding to the pixel is stored in the dictionary 17 as a dictionary image, and the correlation calculation between the processed image of the recognition target character supplied via the replacing unit 19 and the dictionary image of each character stored in the dictionary 17 is calculated. Is performed in the similarity calculation unit 21,
The degree of similarity is judged from each degree of correlation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device used in an optical character reading device.

[0002]

2. Description of the Related Art As a method applied to a character recognition apparatus of this type, there is a conventional method such as a composite similarity method (Japanese Patent Application No. 44-881).
09)), a matching method, a structural analysis method, and the like. The composite similarity method and the matching method basically recognize a character based on the degree of correlation between the target character and the dictionary (reference character). In particular, the composite similarity method is to be strong against displacement. The feature is that the blurring process is performed.

Further, the structure analysis method analyzes what kind of features a character is composed of and how, describes the structure based on the analysis, and matches this description with the description of each character concept. Is a method of recognizing.

[0004]

In each of the above-mentioned methods, the character to be recognized retains its characteristics, but if a part of the character is missing or stains are overlaid, a false determination is made. There are problems such as reading and reading.

The present invention has been made in view of the above,
It is an object of the present invention to provide a character recognition device which has a high recognition rate closer to human judgment and recognizes characters accurately.

[0006]

In order to achieve the above object, a character recognition device of the present invention provides a recognition target character group in which the appearance frequency of each pixel of a pattern portion and a background portion forming each character of the recognition target character group is determined. Appearance frequency calculating means for calculating the whole, and weighting each appearance frequency of the pattern portion of each character and the portion corresponding to each pixel of the background portion in the appearance frequency of the entire recognition target character group calculated by this appearance frequency calculating means Then
Weighting value calculating means for calculating weighting values by giving different signs to the respective weighting values of the pattern portion and the background portion, and each character having the weighting value calculated by this weighting calculation means corresponding to each pixel. A storage unit for storing as a dictionary image, a correlation calculation between the input processed image of the recognition target character and the dictionary image of each character stored in the storage unit, and the similarity for determining the similarity from each correlation The gist is to have a judgment means.

[0007]

In the character recognition device of the present invention, the appearance frequency of each pixel of the pattern portion and the background portion of each character is calculated for the entire recognition target character group, and the pattern portion and the background portion of each character are calculated at the appearance frequency. Each appearance frequency of the portion corresponding to each pixel is weighted, a weighting value is calculated by assigning a different sign to each weighting value of the pattern portion and the background portion, and the weighting value is provided for each pixel. Each character is stored as a dictionary image, a correlation calculation is performed between the input processed image of the recognition target character and the stored dictionary image of each character, and the degree of similarity is determined from each degree of correlation.

[0008]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a character recognition device according to an embodiment of the present invention. The character recognition device shown in the figure extracts a feature point of each character, creates a dictionary in which the weight of the feature point is increased, performs a correlation calculation between this dictionary and the recognition target character, and calculates the similarity from each correlation degree. Judge,
Characters are identified from the largest one.

In FIG. 1, the recognition target character image signal imaged by the camera 1 is supplied to the analog-digital converter 5 of the processing device 3 and converted into a digital signal. The image signal of this digital signal is further binarized by the binarization processing unit 7 and supplied to the character cutting unit 9 to cut out the target character portion.

The character information cut out by the character cutout unit 9 includes the appearance frequency calculation unit 11, the weighting unit 13, and the normalization unit 1.
5 and the dictionary 17 may be supplied to the similarity calculation unit 21 via the dictionary creation unit or may be supplied to the similarity calculation unit 21 via the replacement unit 19. First, by supplying all the reference characters of the recognition target character group to the dictionary creation unit including the appearance frequency calculation unit 11, the weighting unit 13, the normalization unit 15, and the dictionary 17, the reference dictionary image of each character is obtained. After being stored in the dictionary 17, the recognition target character is replaced by the similarity calculation unit 2 via the replacement unit 19.
1, the similarity calculation unit 21 calculates the similarity with the dictionary image that serves as a reference for each character from the dictionary 17, and recognizes the recognition target character.

Therefore, first, it is assumed that the camera 1 captures the reference character, and the image signal of the reference character is supplied to the character slicing unit 9 through the analog-digital converter 5 and the binarization processing unit 7. A case will be described in which the reference character information cut out from the character cutout unit 9 is supplied to the appearance frequency calculation unit 11, the weighting unit 13, and the normalization unit 15 to create a dictionary.

Each reference character image signal cut out from the character cutout unit 9 is supplied to the appearance frequency calculation unit 11, where the appearance frequency of each pixel of the pattern portion and the background portion forming each character of the recognition target character group. Is calculated for the entire recognition target character group. A more specific description will be given with reference to FIGS. 2 and 3.

FIG. 2 shows a character example in which the entire recognition target character group is represented by numbers 1 to 0, and the numbers 1 to 0 are composed of horizontal 5 × vertical 9 pixels. FIG. 3 shows each pixel of horizontal 5 × vertical 9 shown in FIG. 2 surrounded by each frame, and shows the frequency of appearance of each pixel in the pattern portion of each character shown in FIG. 2 in each pixel. .. The frequency of this appearance is m in the horizontal position.
And Gm, n where n is the vertical position, G1,1 = 5, G1,2 = 9, etc., for example, G1,2
= 9 indicates that a number other than 4 which is a recognition target number has a pixel at the position of G1,2, and when there is no pixel at this position, it is understood that it is 4. Also, it can be seen that there is one number having a pixel at the position of G2,3, which is the number 1.

The appearance frequencies calculated by the appearance frequency calculating unit 11 as shown in FIG. 3 are supplied to the weighting unit 13 and the normalizing unit 15, and are weighted and normalized for each character. FIG. 4 shows the result of weighting and normalization for the numeral 1 shown by enclosing the pattern portion with a thick line. In FIG. 4, k is the normalization coefficient of the pixel of the pattern portion of each character, and k ′ is each It is a normalization coefficient of pixels in the background portion without a character pattern. The weighting and normalization are values obtained by multiplying the reciprocal of the appearance frequency of each pixel by the normalization coefficient k for each pixel in the pattern portion of each character, that is, the numeral 1 in FIG. 4, and for each pixel in the background portion without the pattern portion. For a pixel, the reciprocal of the appearance frequency of the pixel of the background portion (that is, the appearance frequency obtained by subtracting the appearance frequency of the pixel of the pattern portion shown in FIG. 3 from the overall appearance frequency of 10) is multiplied by a normalization coefficient, The sign is negative, that is, the value is −. Specifically, in FIG. 4, for the appearance frequency G1,1 = 5 (see FIG. 3) of the upper left pixel of the pattern portion of the character 1, the reciprocal of 5 is taken, and this reciprocal is multiplied by the normalization coefficient k. For the appearance frequency G1,5 = 7 (10-3) of the upper right pixel which is the background part, the reciprocal of 7 is taken, and this reciprocal is multiplied by the normalization coefficient k ', The value is −k ′ / 7 where the sign is −. FIG. 4 shows the normalized weighting value of the character 1 with respect to the entire appearance frequency of the recognition target character group shown in FIG. 3, but similar processing is performed for each character of the entire recognition target character group other than the numeral 1. To do. The normalized weighting value obtained for each of the characters 1 to 0 of the recognition target character group in this manner is stored in the dictionary 17 as a dictionary image of the reference character for each character.

Note that FIG. 4 shows a value obtained by multiplying the reciprocal of the appearance frequency shown in FIG. 3 by a normalization coefficient as the normalization weighting value, but FIG. 5 shows another normalization weighting value.
The normalization weighting value shown in FIG. 5 is a value obtained by subtracting the appearance frequency shown in FIG. 3 from the overall appearance frequency 10 for the pattern part and a normalization coefficient. 3 (that is, the appearance frequency obtained by subtracting the appearance frequency of the background portion from the overall appearance frequency 10) is multiplied by a normalization coefficient.

When the dictionary image for each character is stored in the dictionary 17 as described above, the character to be recognized next is read by the camera 1, and this character is stored in the dictionary for the reference character dictionary. The character recognition is performed by calculating the degree of similarity with the image. This processing will be described.

As described above, the recognition target character imaged by the camera 1 is converted into a digital signal by the analog-to-digital converter 5, binarized by the binarization processing unit 7, and targeted by the character slicing unit 9. When the character portion to be cut out is cut out, the image signal of the cut-out recognition target character is supplied to the replacing section 19. The image signal of the recognition target character is given as a binary signal of 1 and 0 from the binarization processing unit 7 to the density information as shown in FIG. 6A, that is, to the pixels of the pattern unit of each character. For that, 1 is given,
0 is given to the pixel of the background part which is not the pattern part, but the value of 0 for the pixel of this background part is "-1".
The replacement unit 19 performs the process of replacing the character string with the character string, and supplies it to the similarity calculation unit 21 as the recognition target character image P (m, n). As described above, by setting 0 to -1, the multiplication in the correlation calculation performed by the similarity calculation unit 21 is enabled. That is, in the correlation calculation, the multiplication process is performed, but 0 is no. Therefore, in order to prevent this, the replacement process in which 0 is changed to -1 is performed.

The similarity calculation unit 21 calculates the similarity between the recognition target character image P (m, n) supplied from the replacement unit 19 and the dictionary image Bi (m, n) of each character read from the dictionary 17. It is calculated from the correlation calculation shown in the following formula.

[0020]

[Equation 1]

Here, i is the number of character types in the character group to be recognized and is 1 to 10. M and N are image sizes for each character, and specifically, the numbers of horizontal pixels and the number of vertical pixels as shown in FIGS.

FIG. 7 illustrates the similarity calculation, but the recognition target character image P (m, n) and the dictionary 17 are shown.
The similarity calculation of each character read from the dictionary image Bi (m, n) is calculated according to the above correlation calculation formula, and the character of the dictionary having the maximum similarity is output as the answer to the recognition target character.

In the above-described embodiment, the binarized image is targeted, but the multi-value grayscale image of the character recognition device of the present invention can be processed. The appearance frequency Gm, n in this case uses the added value of the densities of the respective pixels, and the "-1" sign in the replacement section 19 is provided with a bias of a certain density, and a positive or negative sign is given based on this density value. Just attach it.
For example, with respect to the multi-valued input image density as shown in FIG. 8A, a bias having an almost middle value as shown in FIG. 8B is provided, the upper side of this bias is +, and the lower side is Can be-.

As described above, the character recognition apparatus extracts the feature points of each character, creates a dictionary in which the weight of the feature points is increased, and calculates the similarity between the dictionary and the recognition target character. However, since the characteristic points are emphasized, a result closer to human recognition can be obtained. In addition, since the feature points consider not only the pattern part that constitutes the character but also the background part without the pattern part, even if the character part is missing in the pattern part, the part of the less important part of each character is missing. It is blunt to, and has a great effect on the similarity of the missing part of the character that is of high importance, and a similar effect is obtained for the stain around the character in the background part. A higher correct answer rate can be obtained with respect to dirt and the like.

[0025]

As described above, according to the present invention,
The appearance frequency of each pixel of the pattern portion and the background portion of each character is calculated for the entire recognition target character group, and the appearance frequency of the portion corresponding to each pixel of the pattern portion and the background portion of each character in the appearance frequency is calculated. Weighting is performed, weighting values are calculated by assigning different signs to the weighting values of the pattern portion and the background portion, and each character having the weighting value corresponding to each pixel is stored as a dictionary image and input. Correlation calculation between the processed image of the recognition target character and the stored dictionary image of each character is performed, and the similarity is determined from each correlation, so even if each character is missing or has stains, Character recognition can be performed accurately with high accuracy.

[Brief description of drawings]

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

FIG. 2 shows the number 1 as an example of the entire recognition target character group.
FIG. 3 shows a diagram in which each of the numbers 1 to 0 in the case of ˜0 is composed of horizontal 5 × vertical 9 pixels.

FIG. 3 is a diagram showing the frequency of appearance of pixels in the pattern portion of each character shown in FIG.

FIG. 4 is a diagram showing a normalized weighting value of character 1 with respect to the appearance frequency shown in FIG. 2;

5 is a diagram showing another normalized weighting value of character 1 with respect to the appearance frequency shown in FIG. 2;

6 is a diagram showing an input image density of 2 in the character recognition apparatus of FIG.
It is explanatory drawing which shows a value conversion process.

7 is an explanatory diagram showing a correlation calculation process in a similarity calculation section used in the character recognition device in FIG.

FIG. 8 is an explanatory diagram showing a coding process in a replacement unit of the character recognition device in FIG. 1 for an input image of a multi-value grayscale image.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 camera 3 processing device 7 binarization processing unit 9 character cutout unit 11 appearance frequency calculation unit 13 weighting unit 15 normalization unit 17 dictionary 19 replacement unit 21 similarity calculation unit

Claims (2)

[Claims] 1. An appearance frequency calculation means for calculating the appearance frequency of each pixel of a pattern portion and a background portion forming each character of the recognition target character group for the entire recognition target character group, and calculation by the appearance frequency calculation means. In the appearance frequency of the entire recognition target character group, each appearance frequency of the pattern portion of each character and the portion corresponding to each pixel of the background portion is weighted, and a different code is given to each weight value of the pattern portion and the background portion. Weighting value calculating means for calculating the weighted value, storage means for storing each character having the weighting value calculated by the weighting calculating means corresponding to each pixel as a dictionary image, and inputting the recognition target character And a similarity determination unit that determines the similarity from each correlation by performing a correlation calculation between the processed image and the dictionary image of each character stored in the storage unit. Character recognition device. 2. The input processed image of a recognition target character is an image in which a pattern portion and a background portion of each character are given different codes.
Character recognition device described.