JP2875330B2 - Character recognition method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a character recognition method used for identifying a postal code or the like, and more particularly to a character recognition method which is fast and has excellent recognition accuracy.

(Prior Art) Conventionally, a technique in this field is disclosed in Japanese Patent Application Laid-Open No. 57-2318.
There was one described in No. 5 gazette and the like. Hereinafter, description will be made with reference to the drawings.

FIG. 2 is an explanatory view for explaining a conventional character recognition method described in the above-mentioned document, and FIGS. 3 (a) and 3 (b) are views showing an example of an information medium.

In this character recognition method, first, characters on an information medium are photoelectrically converted to obtain a character pattern A (step 1). From the character pattern A, sub-patterns A1 to A4 representing line element components (stroke components) in each direction are extracted (step 2). Further, each of the sub-patterns A1 to A4 is divided into a plurality of areas A1a of N × M (for example, N = M = 5) (step 3). After that, the number of black spots in each of the divided areas A1a is counted, and a feature amount representing the stroke length of the sub-pattern in each of the areas A1a is extracted. Create a feature matrix. Then, characters are identified by comparing the obtained feature matrix with a dictionary mask prepared in advance.

By the way, the postal codes such as postcards and postcards shown in FIGS. 3 (a) and 3 (b) are printed in advance together with the address as shown in FIG. 3 (a). As shown in FIG. When an individual fills in the form, it is generally handwritten.However, when the form is printed, it is found in a reply postcard of direct mail, an application postcard inserted in a reply envelope or a magazine book, etc. The number is huge. The same applies to a case where handwritten characters and printed characters are mixed in slips such as an order slip, an application form, and a request slip.

In such a case, a large-capacity dictionary memory corresponding to both handwritten characters and printed characters is prepared to perform character recognition.

(Problems to be Solved by the Invention) However, in the above-described character recognition method, it is necessary to perform matching with all dictionary masks in a large-capacity dictionary memory corresponding to both handwritten characters and printed characters. For this reason, there is a problem that the number of collation errors during collation is likely to occur correspondingly, which causes a reduction in recognition accuracy and an increase in the time required for collation.

An object of the present invention is to provide a character recognition method that solves the problems of the prior art, that is, a reduction in recognition accuracy and an increase in collation time.

(Means for Solving the Problems) In order to solve the above problems, the present invention extracts the characteristics of a character from a multi-valued image obtained by photoelectrically converting a character on an information medium, and prepares the extraction result in advance. In the character recognition method for identifying the character by comparing the plurality of dictionary masks with each other, different thresholds for recognition and different thresholds for dictionary selection are set for the multi-valued image. A binary pattern for recognition and a dictionary corresponding to the binarization threshold for recognition and the binarization threshold for dictionary selection based on the binarization threshold for recognition and the binarization threshold for dictionary selection After each of the selection binary patterns is obtained, the feature of the dictionary selection binary pattern is detected, and a predetermined dictionary mask is selected from the plurality of dictionary masks based on the detection result. is there.

Further, as the plurality of dictionary masks, masks for handwritten characters and printed characters may be used.
The feature of the value pattern may be the sum of the number of black points in the binary pattern, the character dose, the line width, or the distribution state of the black points.

(Operation) According to the present invention, since the character recognition method is configured as described above, a binary threshold for recognition and a dictionary for a multivalued image of an information medium such as a postcard are determined based on the density and the like of the multivalued image. A binarization threshold for selection is set, binarization is performed based on the binarization threshold for recognition and the binarization threshold for dictionary selection, and the binarization threshold for recognition and the binarization threshold for dictionary selection are set. After obtaining the corresponding binary pattern for recognition and the binary pattern for dictionary, the features such as the sum of the number of black points, the character dose, the line width, or the distribution state of the black points of the binary pattern for dictionary selection are detected. Based on the detection result, a predetermined dictionary mask is selected from a plurality of dictionary masks for handwritten characters, printed characters, and the like, and collation is performed using the predetermined dictionary mask. Thereby, the dictionary mask to be compared can be limited.

 Therefore, the above problem can be solved.

(Embodiment) FIG. 1 is a block diagram showing a configuration of a character recognition apparatus for implementing a character recognition method according to an embodiment of the present invention.

This character recognition device uses an optical signal L input from an input terminal 10.
A photoelectric conversion unit 11 such as a CCD sensor for converting
It has an image memory 12 for storing digital signals obtained by photoelectric conversion in the photoelectric conversion unit 11 as multi-valued image data.

The density of the multivalued image data is represented by a multivalued value such as 4 bits per pixel. For example, in the case of 4 bits, the most white is represented by 0 and the most black is represented by 16 levels. Image memory
Reference numeral 12 denotes a storage format having a capacity of, for example, 64 × 64 × 4 bits for storing multi-valued image data for one character, and capable of reproducing a signal of each pixel corresponding to two-dimensional coordinates on an information medium. , RAM (random access memory) and the like.

On the output side of the image memory 12, binarization processing is performed on the multi-valued image data in the image memory 12 using the binarization threshold T1 for recognition, and the binarization processing corresponding to the binarization threshold T1 for recognition is performed.
A first binarizing unit 13 for obtaining a value pattern, and a second binarizing unit 14 for similarly performing a binarizing process using a binary threshold for dictionary selection T2 to obtain a binary pattern for dictionary selection And are respectively connected. The threshold value T1 for recognition and the threshold value T2 for dictionary selection are determined by the characteristics of the photoelectric conversion unit 11. The values “1” and “0” of the binary pattern for recognition and the binary pattern for dictionary selection represent the black point and white point of the binary pattern for recognition and the binary pattern for dictionary selection, respectively. I have.

Further, on the output side of the first and second binarization units 13 and 14, a first pattern register 15 for storing the binary pattern for recognition obtained in the first binarization unit 13 is provided. 2 of 2
A second pattern register 16 for storing the dictionary selection binary pattern obtained by the value conversion unit 14 is connected to each other.

The output side of the second pattern register 16 has a register 16
Scan the entire surface of the register 16 and count the number of black points in this register 16,
A character presence determining unit 17 that outputs a dictionary selection signal DS according to the counting result, and a handwritten character dictionary unit having a printed character dictionary unit having a dictionary mask for printed characters and a dictionary mask for handwritten characters, The dictionary memory 18 for selecting one of the dictionary units by the dictionary selection signal DS is sequentially connected. The dictionary memory 18 is composed of a RAM or the like, and outputs a dictionary mask for both the handwritten character dictionary unit and the printed character dictionary unit when a dictionary selection signal DS of logic “1” is input, and outputs a logic “0”. In the case of the dictionary selection signal DS, it is a memory for outputting a dictionary mask of the print character dictionary unit.

The dictionary memory 18 and the first pattern register 15 are connected to the recognition unit 19. The recognizing unit 19 performs feature extraction on the binary pattern for recognition stored in the first pattern register 15 to create a feature matrix fi, and generates the feature matrix fi and the dictionary mask output from the dictionary memory 18. And output the resulting character name
This is a circuit to output to 20. Here, the dictionary mask is composed of a dictionary matrix gi having the same format as the feature matrix fi.

The operation of the character recognition device configured as described above will be described.

An optical signal L is transmitted from an information medium such as a postcard, a letter, a form, or the like on which characters, graphics, etc. are written through a vidicon camera (not shown).
Is input from the input terminal 10, the photoelectric conversion is performed in the photoelectric conversion unit 11, and as a result, the obtained digital signal is stored in the image memory 12 as multi-valued image data. Thereafter, the multi-valued image data is binarized by the first binarizing unit 13 and stored in the first pattern register 15 as a binary pattern for recognition.

This binarization process is performed as follows. Recognition binary pattern f1 stored in first pattern register 15
(X, y) is obtained by the following equation.

f1 (x, y) = 1 when g (x, y) ≧ T1 f1 (x, y) = 0 when g (x, y) <T1, where g (x, y); multi-valued image T1 which is data and has a value of 0 to 15; binarization threshold value for recognition (1) In the second binarization unit 14, similarly to the first binarization unit 13,
Binarization processing is performed on the multi-valued image data based on the following equation, and the obtained binary pattern for recognition f2 (x, y) is stored in the second pattern register 16.

f2 (x, y) = 1 when g (x, y) ≧ T2, f2 (x, y) = 0 when g (x, y) <T2, where T2; 2) The character presence determination unit 17 scans the entire second pattern register 16 and counts the number of black points BL in the second pattern register 16 as follows.

Based on the number of black points BL, the dictionary selection signal DS created by applying the following equation is output to the dictionary memory 18.

DS = 1 when BL ≧ BTHL DS = 0 when BL <BTHL However, BTHL; threshold value (4) The dictionary selection signal DS output from the character existence determination unit 17 is
The data is input to the dictionary memory 18. For example, when the dictionary selection signal DS is input with (DS = 1), the dictionary mask for handwriting and the dictionary mask for printing are output to the recognition unit 19. When the dictionary selection signal DS is input with (DS = 0), only the printing dictionary mask is output to the recognition unit 19.

 The recognizing unit 19 performs the following processing.

First, the line width of a character pattern which is a binary pattern for recognition in the first pattern register 15 is calculated. Next, the character pattern is scanned in a plurality of directions to detect the continuous number of black spots in each scanning row. On the basis of the detected continuous number and the line width, four types of sub-patterns are obtained by scanning in, for example, four directions of horizontal, vertical, diagonal right and diagonal left.

Subsequently, the area within the character frame of the character pattern is divided into N × M areas (where N and M are positive integers) for the sub-pattern.
Divided into The number of black points is counted in the divided area, and the feature amount is calculated based on the counted result and the line width. Further, the feature amount is normalized by the size of the character to create a feature matrix fi. Note that the dimensions are N × M × 4 dimensions. For example, if N = 4 and M = 4, the dimensions are 64 dimensions.

Then, the feature matrix fi is compared with the dictionary mask output from the dictionary memory 18. That is, the dictionary matrix gi (i = 1, 2,..., 64) and the feature matrix fi (i =
, 64) to calculate the distance D, and output the character name corresponding to the dictionary matrix gi having the smallest distance D to the output terminal.

Next, a case where a zip code is read using the above-described character recognition method will be described.

FIG. 4 is a diagram showing the density when reading a black handwritten postal code, and FIG. 5 is a diagram showing the density when reading a red printed postal code.

As shown in FIGS. 4 and 5, for example, zip code 3
When reading 1,32, the entry box 30 is usually printed in red. Reply postcards that are enclosed in direct mail or application postcards bound in magazines have a pre-printed postal code 32, and the postal code 32 is the same color as the entry box 30 for convenience of printing. Often red. On the other hand, the postal code 31 entered by a general handwriting is usually black or blue.

In FIG. 4, when reading a postal code 31 written in black by hand in a red writing box 30, first, as shown in this figure, the medium is scanned in the x direction by the photoelectric conversion unit 11. Since the density of red is lower than the density of black, the density of the entry frame 30 printed in red is lower than the density of the postal code 31 of black. Here, the value of the multi-valued image data is red
Assuming that about 10 and about 15 are obtained in black, the binarization threshold T1 for recognition is 8 and the binarization threshold T2 for dictionary selection is 12
Set to. As a result, in the first binarization unit 13 in which the binarization processing is performed based on the binarization threshold value T1 for recognition, the recognition is performed in which the entry frame 30 and the postal code 31 are binarized by the above equation (1). And a first pattern register is obtained.
Stored in 15.

On the other hand, the second binarization unit 14 that performs the binarization processing based on the dictionary selection binarization threshold T2 uses the entry frame by the above equation (2).
A binary pattern of 30 is not obtained, and a binary pattern of only the postal code 31 is obtained. The binary pattern of this zip code 31 is
After being stored in the second pattern register 16, the number of black points BL is counted in the character presence determination section 17 by the above equation (3). Here, if BTHL = 10, then BL ≧ BTHL as shown in the above equation (4), so that DS = 1, and the dictionary selection signal DS of logic “1” is output to the dictionary memory 18. As a result, in the dictionary memory 18, both the handwritten character dictionary mask and the printed character dictionary mask are output to the recognition unit 19. Then, the dictionary mask and the recognition 2 in the register 15 are used.
Matching with the value pattern is performed, and the obtained character name is output
Output to 20.

On the other hand, as shown in FIG.
In the case of reading the postal code 32 written in red print characters, the first binarizing unit 13 scans the medium in the x direction by the photoelectric conversion unit 11 as described above. )
A binary pattern in which the entry frame 30 and the postal code 31 are binarized is obtained by the formula, but the second binarization unit 14 uses the binary pattern of the entry frame 30 and the postal code 31 according to the above formula (2). Is not obtained. Then, since BL <BTHL as shown in the above equations (3) and (4),
DS = 0, and the character selection unit 17 outputs a dictionary selection signal DS of logic “0” to the dictionary memory 18. As a result, in the dictionary memory 18, only the print character dictionary mask is selected and output to the recognition unit 19. Subsequently, the dictionary mask is compared with the binary pattern for recognition in the register 15 in the same manner as described above, and the obtained character name is output to the output terminal 20.

 This embodiment has the following advantages.

When writing a zip code, the printed characters are usually red or black, and the handwritten characters are black and not red. Utilizing this, the binarization threshold is set so that the character pattern printed in red does not appear, and only the character pattern printed or written in black is reproduced, so if the character pattern appears, , And are compared with the dictionary masks for printed characters and handwritten characters.
On the other hand, if the character pattern does not appear, it is assumed that the character pattern is written in red, and the character pattern is compared with the print character dictionary mask. Therefore, the dictionary mask to be collated can be limited, so that the time required for collation can be reduced, and improvement in recognition accuracy can be expected.

Note that the present invention is not limited to the illustrated embodiment, and various modifications are possible. For example, there are the following modifications.

(B) In the above embodiment, the binarization threshold T1 for recognition and the binarization threshold T2 for dictionary selection are each set as a single threshold. However, the binarization threshold T1 for recognition and binarization for dictionary selection are used. Threshold T2
Both can be set as a plurality of thresholds.

(B) In the above embodiment, the feature of the binary pattern for dictionary selection is detected by the sum of the number of black spots of the binary pattern, but this is detected by using, for example, the character dose, the line width, or the distribution state of the black spot. It may be detected.

(C) Although the dictionary mask of the above embodiment targets handwritten characters and printed characters, the present invention is not limited to this, and it is possible to target various other masks.

(D) The first and second binarizing units 13 and 14, the first and second pattern registers 15 and 16, and the character presence determining unit 1 of the above embodiment.
7 and the recognition unit 19 may be configured as individual circuits, or may be executed by program control by a computer.

(E) In the above embodiment, when BL ≧ BTHL, the dictionary selection signal DS
Is set to logic “1”, and when BL <BTHL, it is set to logic “0”.
Conversely, the logic may be set to “1” when BL <BTHL, and may be set to logic “0” when BL ≧ BTHL.

(F) In the dictionary memory 17 of the above embodiment, a dictionary mask for print characters and handwriting characters is output by the dictionary selection signal DS of logic "1", and a dictionary for print characters is output by the dictionary selection signal DS of logic "0". Although only the mask is output, the dictionary mask for print characters and handwritten characters is output by the dictionary selection signal DS of logic “0”, and the dictionary for print characters is output by the dictionary selection signal DS of logic “1”. Only the mask may be output.

(G) In the above embodiment, the values “1” and “0” of the binary pattern for recognition and the binary pattern for dictionary selection are used for the binary pattern for recognition.
Although the black point and the white point of the value pattern and the dictionary selection binary pattern are represented respectively, the opposite black point may be represented by “0” and the white point may be represented by “1”.

(H) The density of the multi-valued image data is represented by 4 bits per pixel and represented by 16 levels of density, but is not limited to this, and is multi-valued with n (positive integer) bits per pixel. Can be expressed in the stage. In that case, the capacity of the image memory 12 also needs to be adjusted according to the density level.

(Effects of the Invention) As described in detail above, according to the present invention, a binarization threshold for recognition and a binarization for dictionary selection are applied to a multivalued image obtained by photoelectrically converting characters on an information medium. After obtaining the binary pattern for recognition and the binary pattern for dictionary selection using the threshold value, respectively, the feature of the binary pattern for dictionary selection is detected, and based on the detection result, a plurality of dictionary masks are selected.
Since a predetermined dictionary mask is selected, the dictionary mask to be collated can be limited, the time required for collation can be shortened, and improvement in recognition accuracy can be expected.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of a character recognition apparatus according to an embodiment of the present invention, FIG. 2 is an explanatory view of a conventional character recognition method, and FIGS. 3 (a) and 3 (b) show an example of an information medium. 4 and 5 are diagrams showing the density at the time of reading the postal code. 11: photoelectric conversion unit, 12: image memory, 13, 14: first
, A second binarization section, 15, 16... First and second pattern registers, 17... A character presence determination section, 18... A dictionary memory, 19... A recognition section, DS. T1: binarization threshold for recognition, T2: binarization threshold for dictionary selection.

Claims (3)

(57) [Claims] 1. A method for identifying a character by extracting a characteristic of the character from a multi-valued image obtained by photoelectrically converting a character on an information medium, and comparing the extraction result with a plurality of dictionary masks prepared in advance. In the character recognition method, different binarization thresholds for recognition and binarization for dictionary selection are set for the multi-valued image, and the binarization threshold for recognition and binarization for dictionary selection are different from each other. After performing binarization based on a threshold to obtain a binary pattern for recognition and a binary pattern for dictionary selection corresponding to the binary threshold for recognition and the binary threshold for dictionary selection, respectively, the dictionary selection is performed. A character recognition method comprising: detecting a feature of a binary pattern for use; and selecting a predetermined dictionary mask from the plurality of dictionary masks based on the detection result. 2. The character recognition method according to claim 1, wherein the plurality of dictionary masks use masks for handwritten characters and printed characters. 3. The character recognition method according to claim 1, wherein a characteristic of the binary pattern for dictionary selection is a sum of the number of black points of the binary pattern, a character dose, a line width, or a distribution state of black points. Method.