JP3277977B2 - Character recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention performs a design process on a binary image in which an input image is represented by "0" or "1", particularly a document image to which noise such as blurring or crushing has been added, or a stripe pattern. The present invention relates to a method for recognizing characters in a document image, such as a document image, and a document image in which outline character images are mixed.

[0002]

2. Description of the Related Art Kanji OCR (Optical Character)
A character recognition processing device such as a Reader) first scans an input binary document image pattern in the horizontal or vertical direction, and projects the number of black or white pixels for each scanning line. The projection distribution of these amounts is obtained, and headings, child headings, character strings, and the like are extracted. Next, the extracted character string image is similarly scanned in the horizontal or vertical direction, and the number of black pixels or white pixels is projected for each scanning line, based on the projection distribution of these amounts. Cut out individual characters. Then, a feature for recognition is extracted in the form of a vector for each character cut out last, and a similarity or a difference between the standard pattern vector of each category in the standard dictionary created in advance is identified. A method is known in which a function is obtained and the most similar character or graphic category is used as a recognition result.

In this method, the character pattern is a black character,
In the case of a document image in which the background portion is composed of white pixels, the character distribution area can be determined from the document image or the individual character area can be determined from the character string image by looking at the projection distribution of the black or white pixels. Can be determined. However, such as newspaper headlines, headlines of magazines such as weekly magazines, etc., characters with black or white characters and a background with a pattern that has been subjected to design processing, such as halftone dots, and diagonal lines etc. For a document image or the like on which a design process has been performed, the projection distribution of the background design process is similar to the projection distribution of the character portion.

In order to solve this problem, the width of each line segment used in the background design process is smaller than the stroke width of the character pattern. There has been proposed a method of removing a background design processing portion by using image processing assuming the following (for example, Okamoto, Hayashi: "Character extraction from a headline with a ground pattern using expansion / contraction" IEICE) Technical Research Report, Pattern Recognition and Understanding Workshop PRU90-151, p
p.47-54,1991.). However, in actual document images, there are many exceptions that do not satisfy the above assumptions, which are clues for separating the design processing part from the character pattern part, and strokes important for recognition are missing due to image processing, line segments in the design processing part However, there is a problem in that the lingering ability remains without being removed, and the abilities of character recognition in the subsequent stage are reduced.

[0005]

As described above, the conventional OCR and other character recognition techniques use a character string image that has been subjected to design processing in the background, a character string image with blurred or crushed characters, and white characters. Means for correctly recognizing a character category from a character string image composed of mixed characters has not been sufficiently established.

[0006] The object of the present invention is to address such problems.
A method for extracting a character string area and individual character candidate images using a count value that reflects a difference between a character string area and a background area, even for noise such as blurring and dirt and a design-processed character string, A method of estimating the degree of noise using the count value for a candidate image of an individual character, and changing a threshold value of a discriminant function used for recognition / rejection determination according to the degree of noise. It is to realize character recognition with higher accuracy than the conventional method.

[0007]

In order to achieve the above object, according to the present invention, a character string and a background area are inputted to separate a character area and a background area from a character string image pattern which has been designed and processed, such as a newspaper headline. The value image pattern is scanned in a plurality of directions, and the number of black-black changes that change from black to black, the number of white-white changes that change from white to white, and the number of black-white changes that change from black to white are on the scanning line. The number of black-and-white changes from black to black is counted, and a character string area is extracted using these count values.

In the present specification, when the value of a binary image pattern does not change between adjacent units in scanning of a predetermined unit (for example, pixel unit) length, when the binary value is black, When the black value and the binary value are white, it is referred to as a white-white change.

Further, according to the present invention, scanning is performed in a plurality of directions on the basis of information on the image pattern of the extracted character string region, and the number of black-black changes on the scanning line, which changes from black to black, and from white to black. Count the number of white-white changes that change to white, the number of black-white changes that change from black to white, and the number of black-and-white changes that change from white to black,
These counts are used to extract candidate areas for individual characters, and the remaining unprocessed individual A method for extracting a character candidate area is provided.

Further, in the comparison between the individual candidate image pattern and the reference pattern in the dictionary, a discriminant function value is obtained by using a known discriminant function, for example. By performing recognition or rejection determination processing based on a determined threshold, a correct character category and its position are selected from many candidate image patterns.

[0011] The present invention is characterized in that a reference pattern of a dictionary and a threshold for recognition / rejection determination can be newly created / added / updated based on input binary image pattern information.

According to the present invention, an image pattern of a document or the like after the input binary image pattern is subjected to a pre-processing such as a skew correction processing and a shading correction processing of the entire document image is targeted. Is extracted, and the character category included in each character string is recognized by collating with the reference pattern of the dictionary. In addition, an image pattern of an erroneously recognized or unlearned character is additionally created or updated as a reference pattern in the dictionary.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to FIGS. FIG. 1 is a block diagram showing a configuration of a character recognition apparatus to which the method of the present invention is applied. In this figure, 1 is a binary image pattern storage circuit, 2 is a black and white counting circuit, 3 is a character string extraction circuit, and 4 is an individual A character extracting circuit, 5 is an individual character matching circuit, 6 is a reference pattern learning circuit, and 7 is an image input / output circuit. 2 is a flowchart for extracting a character string region of the character recognition apparatus of FIG. 1, FIG. 3 is a flowchart for extracting a candidate image of an individual character, and FIG. 4 is a flowchart for matching processing between an individual character candidate image pattern and a reference pattern. FIG. 5 is an explanatory diagram for explaining an operation example of the monochrome counting circuit 2, FIG. 6 is an explanatory diagram for explaining an operation example of the character string extracting circuit 3, and FIG. 7 is an explanatory diagram for explaining an operation example of the individual character extracting circuit 4. is there.

In FIG. 1, a binary image pattern storage circuit 1 stores a binary image pattern to be recognized. The binary image pattern is composed of M pixels, for example, white pixels are “0”,
Black pixels are indicated by "1". The black and white counting circuit 2
A binary image pattern composed of M pixels extracted from the value image pattern storage circuit 1 is input and scanned from a plurality of directions to distinguish between a character string region and a background region. Both the white-white change number (the number of changes from white to white) and the change from black to white, as well as the black-black change numbers (the number of changes from black to black), both of which are black. The number of changing black and white and the number of changing black and white from white to black are counted and output. The character string extraction circuit 3 distinguishes a character string region from a background region based on the black and white count value on each scanning line obtained by the black and white counting circuit 2, and outputs an image pattern of the character string region from the input image. .

The individual character extracting circuit 4 inputs the image pattern of the character string area obtained by the character string extracting circuit 3, selects a candidate image including an individual character to be collated by the individual character collating circuit 5, and outputs the candidate image. . The individual character matching circuit 5 receives the individual character candidate image pattern selected from the individual character extraction circuit 4, matches the image pattern with a reference pattern in a dictionary using a predetermined identification function, and determines the identification function value of each character category. A comparison is made with a threshold value determined by the degree of noise in the candidate image pattern to determine whether or not the candidate image pattern includes a recognized character category.
The reference pattern learning circuit 6 stores the image information output from the image input / output circuit 7 and the binary information of the binary image pattern storage circuit 1.
A value image pattern and a reference pattern of a dictionary in the individual character matching circuit 5 are input, a new reference pattern is created and updated, and the reference pattern of the dictionary of the individual character matching circuit 5 is added and updated.

The image input / output circuit 7 inputs information on recognition results such as a character category number, a discriminant function value, and a binary image pattern output from the individual character collating circuit 5, displays the information, and displays a mouse. Edit image information using a keyboard or keyboard. Further, the image input / output circuit 7 sends information necessary for newly creating or updating the reference pattern to the reference pattern learning circuit 6.

Next, referring to FIG. 1, a process for extracting a character string region from an input binary image pattern will be described. First, a binary document image pattern to be recognized is read out from the binary image pattern storage circuit 1,
Is input to Then, for an input binary image pattern composed of M pixels (horizontal _Nx pixels × vertical _Ny pixels), for example, N
_x> N _y if horizontal determines that N _x ≦ N _y if vertical. Characters are scanned in the horizontal direction for horizontal writing and in the vertical direction for vertical writing, and the total number of black-black change points a changing from black to black on each scanning line and the total number of white-white changing points changing from white to white e, the number of black and white changes b from white to black and the number of black and white changes c from black to white are counted.

Here, for example, it is assumed that a black point has a pixel value of 1 and a white point has a pixel value of 0. On each scan line, a, b, c, e
Using each count value of

[0019]

(Equation 1)

The value of p is determined. The a · e value of the numerator is the product of the total number of black-and-white transition points and the total number of white-and-white transition points on the scanning line. If many exist, p
The value takes a large value. On the other hand, the bc value is the product of the total number of black-and-white changes and the total number of black-and-white changes, and is the number of line segments traversing on a scanning line, that is, the amount corresponding to the complexity of the line segments. When the portion where the design processing is performed is scanned, the bc value becomes high and the p value becomes a small value. In addition, the p-value is −1 ≦ by the normalized term of the denominator.
It takes a value in the range of p ≦ 1.

The p value is projected on a projection axis corresponding to each scanning direction. If the p value obtained by horizontal scanning of the position y is defined as a projection value h _y , the projection distribution P _y =
(H ₁ , h ₂ ,..., H _y ,..., H _Ny ) are obtained. Assuming that a p value obtained by horizontal scanning at the position x is a projection value v _x , a projection distribution P _x = (v ₁ , v ₂ ,..., V
_x ,..., v _Nx ). Counting can be performed in other scanning directions. Here, a, b,
When obtaining c and e, it is also possible to obtain the total number of these for only a specific section.

Next, the character string extraction circuit 3 receives the projection distribution P _y and the projection distribution P _x obtained from the black-and-white counting circuit 2 and, if N _x > N _y , obtains by horizontal scanning. The obtained projection distribution P _y is divided into N equal parts so as to have a preset N section, and a new projection distribution Q = (q ₁ , q ₂ ,..., Q _j,.
q _N ). Otherwise, the projection distribution P _x obtained by scanning in the vertical direction is divided into N equal parts so as to have a preset N section, and a new projection distribution Q = (q ₁ ,
_{q 2, ..., q j,} ..., seek q _N).

Now, the case where Nx> Ny will be described. The new projection distribution Q is compared with a predetermined threshold value TQ, and j continuous sections [Js, Je] (js
The height Wy (= je-js + 1) of the character string is obtained from ≤j≤je. An image area surrounded by the height Wy pixels and the width Nx pixels is output as a character string image pattern W. Here, when the character string is composed of a plurality of lines, a plurality of continuous sections are output, so that a plurality of character string image patterns W are output .

Next, the individual character extraction circuit 4 normalizes the character string image pattern W of N _x × W _y pixels obtained by the character string extraction circuit 3 so that the height becomes N pixels. N _x
A character string image pattern W _N of × N pixels is obtained. For this normalized character string image pattern W _N , the projection distribution P _x of the p-value by the vertical scanning, which is obtained by the processing procedure of the monochrome counting circuit 2, is obtained, and the projection distribution P _{x is} set in advance. is compared with the threshold value T _x, v _j> T _x become x continuous interval [x _s, x _e] Request _{(x s ≦ x ≦ x e} ).

The k continuous sections [x _s1 , x _e1 ], [x _s2 ,
x _e2 ],..., [x _sk , x _ek ], z ₁ ≤x which falls within the range from z ₁ pixel to z ₂ pixel which is the range of the character width stored in the reference pattern. _ej− x _si ≦ z ₂ (i ≦
j) and the section j that satisfies j) are obtained, and the character width W _x (= x _ej −x _si +1) of the individual character candidate image pattern G is obtained. As a result, a plurality of individual character candidate image patterns G each composed of W _x N pixels are obtained. If in section i x _ei -x
_{When si} > z ₂ , the position x _si of the left frame of the candidate image pattern G is assumed on the assumption that the candidate image pattern G is W _x = N.
Are shifted in units of one pixel, and a plurality of Gs are created until the right frame of G is at the position _xei .

[0026]

The character pitch (interval) is estimated from the section of the individual character candidate image pattern G determined to be "recognized in a certain character category" by the individual character matching circuit 5, and the individual character candidate is estimated from the remaining area. A method of extracting the image pattern G is also applicable.

In the above description, a case where a horizontally long binary image is input has been mainly described. However, a vertically long binary image pattern can be similarly obtained by reversing the processing in the x and y directions. Can be processed.

Next, the processing of the individual character matching circuit 5 will be described in detail. First, an individual character candidate image pattern G composed of W _x × N pixels is input, and a normalized binary character pattern X = (x ₁ , x ₂ ,...) Of n (= N × N) pixels
x _i ,..., x _n ). At this time, translation is performed so that G is at the center of the image X. Using monochrome counting circuit 2 obtains a projection distribution P _y horizontal scanning of the normalized character pattern X, the sum H of h _y represented by the following formula

[0030]

(Equation 2)

Is obtained. Next, this normalized character pattern X is converted into a binary reference pattern T = (t ₁ , t ₂ ,.
For comparison with t _i ,..., t _n ), the similarity S (X, T) of the following equation, which is a known discriminant function, is used.

[0032]

(Equation 3)

Here, A is the inner product of X and T and is the total number where x _i = 1 and t _i = 1, T ′ is the total number of black spots of T, and X ′ is the total number of black spots of X. For the reference pattern of each character category m, the similarity S (X, T _m ) to X and the threshold value S _{m of the} reference pattern determined by the value of H are used for recognition or rejection as follows. Make a decision. This H and threshold S
The relation table of _m is learned in advance by the reference pattern learning circuit 6.

(Condition 1) If S (X, T _m )> S _m , it is determined that the character category m has been recognized. (Condition 2) Otherwise, it is determined that the character category is not m.

If the condition 1 is satisfied for a plurality of categories for all reference patterns, the category having the highest similarity is output as the recognition category. If there is no category that satisfies the condition 1, the individual character candidate image pattern G is rejected as having no recognition category.

Finally, a process of adding / updating a reference pattern of the dictionary of the individual character matching circuit 5 using the reference pattern learning circuit 6 and the image input / output circuit 7 will be described. The image input / output device 7 inputs classification character information such as a recognition character category number, a discrimination function value, and a binary image pattern output from the individual character matching circuit 5, displays the information, and displays a mouse or keyboard. Edit the image information by, for example, When there is a misrecognition in the recognition result or when a reference pattern of a new character category is created, the image input / output device 7
From the binary image pattern of input, add, edit, create a binary reference pattern T _m of recognition category m to be updated. The reference pattern learning circuit 6 uses the created binary reference pattern and the reference pattern of the dictionary of the individual character matching circuit 5 to
To learn this new n (= N × N) threshold S _m of recognition rejection judgment for the normalized binary reference pattern T _m including pixels.

Now, a binary random noise pattern R consisting of n (= N × N) pixels R = (r ₁ , r ₂ ,..., R _i ,.
Create a r _n), superimposed on each binary reference pattern T dictionary to create input image pattern X for learning. That is, the pixel value x _i = t _i vr _{i at} the position i is obtained (v = logical sum). The similarity S (X, T _m ) of the input image pattern X to the binary reference pattern T _m of the new category m is calculated, and among all the reference patterns, the maximum similarity is calculated for the category other than m. the value as S _m, the dictionary of individual characters matching circuit 5, and stores the threshold S _m a binary reference pattern T _m of a new category m.

Next, the flow of processing in each circuit described above will be described with reference to the flowcharts shown in FIGS.
Character recognition device shown in FIG. 1 is a process of extracting a character string area, as shown in FIG. 2, first, a binary document image pattern for recognition (M pixels = horizontal N _x pixels × vertical N _y pixels) Is input (step 102). Then, whether horizontal writing or vertical writing is determined from the horizontal and vertical sizes (step 103). In this determination, for example, when horizontal writing is determined, scanning is performed in the horizontal direction, and the number of white / black changes a, b, c,
The projection value h _y is obtained by counting e (step 104).

After this horizontal scanning is completed for the entire binary document image pattern (step 105), the p-value projection distribution P
_y = (h ₁ , h ₂ ,..., h _y ,..., h _Nx ) is divided into N equal parts so as to have a preset N section, and a new projection distribution Q =
(Q ₁ , q ₂ ,..., Q _j ,..., Q _N ) are obtained (step 106). The projection distribution Q, obtains the region greater than the threshold value T _Q set in advance as the character string image pattern W (step 107). If vertical writing is determined in step 104, scanning is performed in the vertical direction, the number of black and white changes a, b, c, and e on the scanning line is counted, and the projection value v is calculated.
_x is obtained (step 108).

After this vertical scanning is completed for the entire binary document image pattern (step 109), the p-value projection distribution P
_x = (v ₁ , v ₂ ,..., v _x ,..., v _Nx ) is divided into N equal parts so as to have a preset N section, and a new projection distribution Q =
(Q ₁ , q ₂ ,..., Q _j ,..., Q _N ) are obtained (step 110). The projection distribution Q, obtains the region greater than the threshold value T _Q set in advance as the character string image pattern W (step 107).

Next, in the process of obtaining a candidate image of an individual character from the character string image pattern W, first, as shown in FIG. 3, the character string image pattern W is input (step 202). For example, in the case of a horizontally written character string image pattern, a normalized character string image pattern W of horizontal _Nx × vertical _Wy pixels is subjected to normalization processing in a vertical and horizontal unity so that the height becomes N pixels, and n A character string image pattern W _N of _x × N pixels is obtained (step 203). For this normalized character string image pattern W _N, obtains the projection distribution P _x of projection values v _x by vertically scanning the processing procedure mentioned in steps 108 109 (step 204), with respect to the projection distribution P _x Thus, a continuous section of x greater than a preset threshold T _x [x _s ,
_xe ] ( _xs ≦ x ≦ _xe ) is obtained (step 205).

A character width W _x pixel suitable for collation is selected from the adjacent continuous sections, and an individual character candidate image pattern G is obtained (step 206). In the matching process between the individual character candidate image pattern and the reference pattern,
First, an individual character candidate image pattern G consisting of W _x N pixels is input (step 302). This G is subjected to position and size normalization processing, and n (= N × N)
Pixel normalized binary character pattern X = (x ₁ , x ₂ ,...,
x _i ,..., x _n ) are obtained (step 303). This X
The known similarity S (X, T _m ) is calculated between the reference pattern T _m and the reference pattern T m of the character category m (step 304). The cumulative value H of the scanning projection value h _y is obtained (step 305).

The threshold value for the cumulative value H is determined by referring to the relation table with reference to the threshold value S for the character category m.
_m is obtained (step 306). Similarity S (X, T _m) is compared with a threshold S _m (step 307), the similarity S
If (X, T _m ) is larger than the threshold value S _m , the character category m is determined to be a recognition category (step 308), otherwise it is determined to be rejected (step 309). The processing from step 304 to step 308 is performed for all the reference patterns (step 309), and the process ends.

Next, as a specific operation example of the monochrome counting circuit 2 which is a main part of the present invention, the binary image pattern storage circuit 1 stores 20 pixels horizontally × 14 pixels vertically (N _x
= 20, N _y = 14) in the case of a binary document image pattern. The figure shows an input image pattern in which a halftone dot design process is performed on the background of a character string image pattern in which “character” is written. 5-1 and 5
-2 is a projection axis for horizontal and vertical scanning, respectively.

Now, it is assumed that the binary document image pattern has been read out and input to the monochrome counting circuit 2. Since the width of the input binary image pattern is larger than the height of the input binary image pattern, it is determined that the image is written horizontally, and the character is scanned in the horizontal direction. Projected onto
In FIG. 5 as an example of scanning, horizontal scanning 5 at y = 2
At -3, a = 0, b = 7, c = 8, e = 4 are counted, and the p value at y = 2 is

[0046]

(Equation 4)

Is as follows. In the horizontal scanning 2-3 at y = 6, a = 5, b = 3, c = 4, and e = 7 are counted, and the p value at y = 6 is

[0048]

(Equation 5)

Is as follows. From this result, it is shown that the p-value of the background portion composed of halftone dots is smaller than the p-value obtained by scanning the character line portion, and the p-value represented by Expression 1 separates the character portion from the background portion. It turns out that this is an effective measure.

The projection distribution Py is obtained by horizontal scanning from y = 1 to y = 14. In the character string extraction circuit 3, y =
The projection distribution Py is input in the horizontal scanning from 1 to y = 14, and the projection distribution Py obtained by the horizontal scanning is divided into N equal parts so as to have N sections set in advance, and a new projection distribution Q = (Q1, q2, ..., qj, ..., qN). Here, when N = 14 pixels, in the projection distribution Q, advance
The section where the set threshold value TQ = 0.0 or more is
The section is from j = 2 to j = 12.
A character string image pattern W of 0 pixels × 11 pixels vertically is extracted.

The individual character extraction circuit 4 performs a normalization process on the character string image pattern W of 20 horizontal pixels × 11 vertical pixels obtained by the character string extraction circuit 3 so that the height becomes W = 11 pixels. A character string image pattern W _N of 20 × 11 pixels is obtained. In the example shown in FIG. 6, since W _x = N = 11 pixels in this case, W and W _N have the same image pattern. For this normalized character string image pattern W _N, as mentioned in the procedure of white counting circuit 2, and projects the p value by the vertical scanning from x = 1 to x = 20 to the projection axis 6-1 , Projection distribution P
_{Find x} . For this projection distribution P _x, when the threshold value T _X = -0.25 previously set, v _j> T _x become x
Is [x _s = 3, x _e = 20]. 6 in FIG.
-2 indicates an example of a vertical scanning line at x = 2.

Now, assuming that the character width is determined in the range of z ₁ = 2 pixels and z ₂ = 11 pixels in the reference pattern.
Since x _e −x _s > z ₂ is satisfied, the position x _s = 2 of the left frame of G is set as a candidate image pattern G of 11 × 11 pixel individual characters.
, Eight individual character candidate image patterns G are created until the right frame position of G becomes x _e = 20. 7-1 in FIG. 7 shows the case where the left frame of G is x = 3, and 7-2 shows the case where the left frame of G is x = 4.

[0053]

As described above, according to the first aspect of the present invention, a binary image pattern in which texture such as halftone dots is superimposed on noise or background such as blur or dirt, which is difficult to extract with the prior art. Even for a binary image pattern composed of white or black characters or white characters, the character string region can be easily and correctly extracted by scanning the image pattern and using a new count value.

According to the second aspect of the present invention, the new count value is applied to a character string image pattern containing noise,
Because individual character candidate images can be extracted, not only do the pre-processing of noise and design removal, which was performed by the conventional character recognition method, but also character strokes are mistakenly deleted for noise and design removal. There is an advantage that character recognition performed in the subsequent stage can be performed with high accuracy.

According to the third aspect of the present invention, in the process of collating between the individual character candidate image pattern and the reference pattern using a predetermined discriminant function and determining whether to recognize or reject, the individual character including noise By measuring the degree of noise of the candidate image pattern using the method according to the first aspect of the present invention, the threshold value of the discriminant function value can be set for each character category according to the degree of noise. There is an advantage that recognition determination can be performed correctly even for characters having noise from characters.

[0056] According to the fourth aspect of the present invention, a character string image pattern including noise is difficult to recognize under a noise background, and a character category such as a punctuation mark or a symbol having a small character width or a low character height is difficult to recognize. However, by paying attention to the area of the character string image excluding the image area for which recognition has already been completed and the information of those recognition results, comparing the unprocessed image pattern with the dictionary again, these punctuation marks And symbols can be recognized correctly.

According to the fifth aspect of the present invention, when a new character category is added to a character string image pattern including noise or a reference pattern is updated, a recognition determined in accordance with the degree of noise. Since the means for easily learning the threshold value of the judgment is provided, there is an advantage that the processing of new creation / addition / update is easy.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration example of a character recognition device according to the present invention.

FIG. 2 is a flowchart for extracting a character string region of the character recognition device of FIG. 1;

FIG. 3 is a flowchart of extracting an individual character candidate image of the character recognition device of FIG. 1;

FIG. 4 is a flowchart illustrating a process of comparing an individual character candidate image pattern with a reference pattern by the character recognition apparatus of FIG. 1;

FIG. 5 is an example of a monochrome counting process for an input binary image pattern according to the present invention.

FIG. 6 is an example of a character string extraction process according to the present invention.

FIG. 7 is an example of an individual character candidate image pattern according to the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Binary image pattern memory circuit 2 ... Black-and-white counting circuit 3 ... Character string extraction circuit 4 ... Individual character extraction circuit 5 ... Individual character collation circuit 6 ... Reference pattern learning circuit 7 ... Image input / output device 5-1 ... Horizontal scanning 5-3, 5-4: Example of horizontal scanning 6-1: Projection axis of vertical scanning 6-2: Example of vertical scanning 7-1, 7-2: Example of individual character candidate image pattern

Claims (5)

(57) [Claims] 1. An input binary image pattern,
By scanning in at least one direction or more in a plurality of directions, the number of black-black changes from black to black, the number of white-white changes from white to white, and the number of black-white changes from black to white on the scan line A black-and-white counting procedure for counting the number of black-and-white changes that change to black; information about the product of the count value for the black-black change number and the count value for the white-white change number obtained by the black-and-white counting procedure; A character string extraction procedure for extracting a character string area using information of a product of a numerical value and a count value regarding the number of black and white changes, and a candidate image of an individual character from the character string area
Individual character candidate image pattern to be extracted
A character recognition method comprising a separate character extraction procedure . 2. A character string image pattern corresponding to a character string region extracted by the character string extraction procedure is scanned in at least one direction in a plurality of directions based on information of the character string image pattern. From the black on the scan line
Counting the white number change changing from Shirajira change number and black-and-white change count and white change from black to white, which changes from black to white black discoloration number and white changes to black to black, the count values for said black black discoloration Number An individual character extraction procedure is provided for extracting an individual character candidate image pattern using information on the product of the count values regarding the number of white-white changes and information on the product of the count value regarding the number of black-white changes and the count value regarding the number of black-and-white changes. 2. The character recognition method according to claim 1, wherein: 3. An individual character candidate image pattern obtained by the individual character extraction procedure is collated by using a predetermined identification function between the individual character candidate image pattern and a reference pattern of a dictionary. Based on the function value and a threshold determined using the information of the individual character candidate image pattern,
3. The character recognition method according to claim 2, further comprising an individual character collation procedure for performing recognition or rejection determination processing. 4. The position of a character category determined to be recognized by the individual character matching procedure with respect to a character string image pattern corresponding to a character string area extracted by the individual character extracting procedure. 4. The character recognition method according to claim 3, further comprising an individual character extraction procedure for extracting an individual character candidate image pattern whose recognition judgment has not been processed in the character string image based on the information of the recognition result and the recognition result. 5. The input binary image pattern information and the individual character collating procedure and at least a character category and 2
Based on the information obtained by the image pattern input / output procedure for inputting / outputting and editing the value image pattern, a predetermined identification function is applied to perform recognition / 5. The character recognition method according to claim 3, further comprising a reference pattern learning procedure for newly creating, adding, and updating a rejection determination threshold value.