KR950011065B1 - A character recognition method - Google Patents

Abstract

The method comprises the steps of obtaining the size of an inputted individual letter; performing a structure of the letter according to the size; extracting elements of the letter; obtaining the size of the extracted letter; comparing the size of the extracted letter with a reference value; storing the compared value in a memory; and recognizing the letter by comparing the extracted letter value with the input letter value.

Description

Character recognition method

1 is a diagram illustrating a general LAG technique.

2 (a) to 2 (e) are explanatory diagrams showing strokes and stroke extractions determined by a conventional character recognition method.

3 (a) and 3 (b) are explanatory views showing regions for the lateral stroke and the longitudinal / inclined direction stroke in the conventional character recognition method.

4 is a flowchart showing the overall operation of the character recognition of the present invention.

5 (a) to 5 (g) are explanatory diagrams showing an example of a trace component extracted by the present invention.

6 is a view showing a direction code defined in the present invention.

7 is a block diagram showing the overall configuration of a recognition device to which the character recognition method of the present invention is applied.

8 is a detailed block diagram showing the configuration of a candidate character classifier in the DSP section of FIG.

9 is a flowchart illustrating a process of classifying candidate characters according to the present invention.

10 (a) and 10 (b) are explanatory views showing an example of a mask used in the present invention.

11 (a) to 11 (i) are explanatory views showing examples of the trace components extracted in the present invention.

12 (a) to 12 (e) are explanatory diagrams showing the division of a mask and a feature region used in the present invention.

* Explanation of symbols for main parts of the drawings

11: host computer 12, 13: first and second DSP unit

14, 15: host interface unit 16, 17: local memory

18: Control Logistics Division 19: Mediation Logistics Division

20: Scanner 21: Scanner Interface

22: global memory 23: memory controller

31: controller 32: stroke extraction unit

33: stroke calculation section 34: reference data section

35: address decoder 36: candidate character registration register

The present invention relates to a character greeting device and a recognition method for selecting and recognizing a candidate character in a feature extraction step and a recognition step for recognizing various characters such as predetermined Hangul, English, Chinese characters and numbers. The present invention relates to a character recognition method for recognizing a character by extracting oblique strokes, vertical strokes, and horizontal strokes from a character without line drawing, and then obtaining the size of the extracted strokes and selecting candidate characters.

In general, in implementing the recognition of characters, in order to design the recognition method using the character's stroke information, it is necessary to extract the characteristics of the character's stroke, which is roughly classified into statistical and structural methods according to the character recognition approach.

In the character recognition method using a statistical method, a character which constitutes a character of a character is usually extracted from a character that recognizes a grid-shaped window called a mesh and does not extract different stroke components. It calculates the number of black pixels of the characters in the area and localizes it as the feature value of the mesh, and recognizes the character with the largest similarity and comparison with the reference data by using the feature value of each mesh as the feature amount of the letter. .

The recognition method by the structural method is generally recognized by using the stroke information of the character. The method of extracting the stroke information of the character is thinning the input character to be recognized and extracting a predefined direction code from the thin line wreath image. The method of selecting only the aromatic component of a specific direction is used.

Extracting the stroke without using the general method, that is, extracting the stroke without performing the thinning process of the input character, using the run-length technique and the line adjacenct graph (LAG) technique There is a run reims / LAG technique.

This run length / LAG technique uses a well-known run length technique. A black pixel that appears continuously while scanning an image is called a run, and the size of the run, that is, the number of black pixels is called a run length.

The LAG technique treats each run on a horizontal line extracted by the run length technique when they are connected up and down on their horizontal line. For example, as shown in FIG. Similarly, when two runs (1) and (2) are on a neighboring horizontal line, it is determined whether the conditional expression of A _x 1≤B _x 2 and the conditional expression of B _x 1≤A _x 2 are satisfied. It is determined that the two runs 1 and 2 are connected to each other.

The character recognition method of extracting a stroke in a specific direction of the character using the run length / LAG technique and recognizing the character by the extracted stroke is Korean Patent Application No. 86-597 (Notice No. 91-5385, Name; Stroke) Character recognition method using approximate straight line extraction).

According to this conventional technique, during scanning of an image of a character, as shown in FIG. 2 (a), pixels connected in a predetermined number or more in the lateral direction form an aggregate connected in the longitudinal direction over two or more scanning lines. Is detected, the lateral stroke is determined in the aggregate of the pixels, and as shown in FIG. 2 (b), the aggregated pixels are formed to extend in the longitudinal direction with only a smaller number than the predetermined number in the lateral direction. When it is detected that the aggregate of the pixels is determined as the longitudinal stroke, as shown in Figs. 2 (c) and 2 (d), the pixels connected with only fewer than a certain number in the lateral direction fall left. And when it is detected that the aggregate extends in the right downward slope direction, the aggregate of the pixels is moved to the left downward slope direction stroke and the right downward slope direction string. Chroman determined by visible instance of each extracted character stroke direction by the extracted by the extracting character stroke of a specific example of the direction the same as claim 2 (e) shown in Fig.

And in each of the divided regions of the region for the lateral stroke as shown in FIG. 3 (a) and the region for the longitudinal / inclined stroke as shown in FIG. 3 (b). By selecting the number of strokes extracted for each direction of the candidate, the candidate character was selected, and the character was identified. However, in the case of recognizing a character in which two or more phonemes are frequently connected to each other by using a conventional character recognition method of calculating a number of strokes and selecting candidate characters as in Korean, the character of the candidate character is frequently recognized. The number is increased, which causes a lot of time to recognize the character as well as the recognition rate of the character is lowered.

The present invention was devised to solve the above-mentioned problems, and it is possible to extract gradient strokes by extracting the traces of a desired direction by detecting the brightness change of the image in a specific direction corresponding to the traces to be extracted. Two gradient strokes of combing and down are extracted by logically combining the images of which brightness change in a specific three direction is detected, and the extracted comb and down from the image of detecting brightness in the vertical and horizontal directions of each image. Logically process the images of the two oblique strokes in to extract the respective strokes, divide the character area into eight specific areas, select the stroke in each feature area, normalize their length, and then measure the absolute value. Select the candidate character by this absolute value, use two specific quantities to recognize the characters, The object of the present invention is to provide a character recognition method for verifying a recognition result with another feature amount and displaying reliability of the candidate character selected by referring to FIGS. 4 to 12. It explains in detail.

4 is a flowchart showing the overall operation of recognizing characters according to the present invention. As shown in Fig. 5 (a), the individual characters to be recognized are cut out to obtain the size of the characters as shown in Fig. 5 (a), and the cutout characters are shown in Fig. 5 (b). As shown in FIG. 5, normalization is performed, and the trace component is extracted as a feature of the normalized characters as shown in FIGS. 5 (c) and 5 (g).

That is, in the present invention, the direction of the stroke is divided into eight directions of N, NE, E, SE, S, SW, W, and NW, as shown in FIG. 5 (c) As shown in Fig. 5 (e), the extraction of the stroke components in the directions of the SW direction, the S direction, and the SE direction is performed, and logically combining these extracted direction components, the fifth (f) As shown in FIG. 5 and FIG. 5 (g), two oblique strokes of the incidence and the down stroke are extracted, and the extracted incidence and the down stroke are extracted for the image generated by detecting the brightness change of the image in the vertical direction. The horizontal stroke is extracted by logically applying the image, and the vertical stroke is extracted by logically applying the image of the rain and down strokes to the image made by detecting the brightness change of the image in the horizontal direction.

In this way, when the stroke components for each direction are extracted, the size of the strokes for each direction extracted in the segmented text area is calculated, the candidate characters are selected by matching and comparing with the reference data, and the feature quantities extracted from these candidate characters. Compares the feature quantities of the input character with the character to recognize the character and save the result.

On the other hand, Figure 7 is a block diagram showing the overall configuration of the recognition device to which the character recognition method of the present invention is applied, as shown in the figure, the host computer 11 for controlling the overall operation of the character recognition, and operates independently of each other First and second DSP (Dingal Singal Processing) units 12 and 13 for selectively dividing a document, dividing individual characters and recognizing characters, and the host computer 11 and the first and second A host memory 14 and 15 for interfacing signals between the second DSP units 12 and 13, and a local memory dependent on the first and second DSP units 12 and 13 to store and output data; 16) (17), and the control logic unit 18 for controlling the first and second DSP units 12 and 13 to selectively perform the area division of documents, the process of cutting out individual characters, and the process of recognizing characters. The arbitration logic unit 19, a scanner 20 for inputting an image of a document, the scanner 20, and the first D Scanner interface unit 21 for interfacing signals between the SP units 12, a global memory 22 for storing the input document image, and a memory for controlling the operation of the global memory 22. It consisted of the controller 23.

In the character recognition apparatus configured as described above, an image of a document input from the scanner 20 through the scanner interface unit 21 is stored in the global memory 22 and is stored in the first DSP unit 12 or the second DSP unit 13. By this, character string separation and individual character extraction are performed, and the result is stored in a specific area of the global memory 22.

Here, the first and second DSP units 12 and 13 are adjusted by the control logic unit 18 and the arbitration logic unit 19 to perform different operations at the same time. For example, the first DSP unit 12 When performing the area division of the document and the process of cutting out the individual characters, the second DSP unit 13 performs the character recognition process, or the first and second DSP units 12 and 13 perform the above operations. You can also do the reverse.

Therefore, when the function of allowing the first DSP unit 12 to perform the character recognition process is granted, the second DSP unit 13 obtains the image information of the individual character from the specific region of the global memory 22 and then corresponds to the region. The image data of the globular 22 is transferred to the local memory 17 to perform feature extraction and candidate classification of the image data, and the finally recognized result is stored in the recognition result unit of the local memory 17.

FIG. 8 is a detailed block diagram showing the configuration of the candidate character classifier for classifying candidate characters in the first and second DSP units 12 and 13 of FIG. 7, and as shown therein, the overall operation of candidate character classification is controlled. The controller 31, the stroke extractor 32 for selecting the stroke in the corresponding region by the stroke extractors FS1-FS8 for each feature region, and the length of the stroke selected by the stroke extractor 32. A stroke calculation unit 33 for normalizing the absolute value of the stroke, a reference data unit 34 for storing and outputting the reference data to the stroke calculation unit 33, and outputting the stroke calculation unit 33 An address decoder 35 for decoding an address signal and a trailing-character registration register 36 for storing candidate characters output by the stroke calculator 33 in accordance with the output signal of the address decoder 35. The.

The candidate character classifier configured as described above operates according to the flowchart shown in FIG. 9, and under the control of the controller 31, the stroke extractor 32 corresponds to the stroke extractors FS1-FS8 classified for each feature region. The stroke of the area is selected, and the selected strokes are calculated by the stroke calculation unit 33 in the length normalization process, and the absolute value of the stroke is calculated, and the reference is output based on the corresponding characteristic area in the reference data unit 34 based on the calculated stroke absolute value. The class to which the input character belongs is classified by comparison and calculation with data, and the result is stored in the candidate character registration register 36 according to the output signal of the address decoder 35.

The character recognition method of the present invention using such a hardware configuration will be described in detail.

First, the basic principle used in the stroke extraction method of the present invention will be described. In the present invention, the stroke extraction method extracts a stroke in a desired direction by detecting a change in brightness of an image in a specific direction corresponding to the stroke to be extracted. do.

An image processing method in a spatial domain can be expressed by the following equation (I).

g (X, Y) -T {f (X, Y)} (I)

Here, f (X, Y) is an input image, and T is defined as an operator that converts a transform or mapping such as a pixel unit to an input image. Determined by a value of a predefined peripheral pixel, g (X, Y) represents an image obtained by applying a predetermined operation or operation T to the input image f (X, Y).

This approach is based on the use of masks, such as templates, filters, and windows, represented as two-dimensional arrays. Each element of the array that makes up the mask is an image. We have a predefined coefficient (coefficient or weight) to detect a given feature in.

FIG. 10 (a) shows an example of a 3x3 mask used in the present invention, where w ₁ -w ₉ represent respective coefficients of the mask, and FIG. 10 (b) corresponds to when the mask is applied to the input image. The pixel to be shown is shown.

Therefore, the image g (X, Y) obtained by using the mask may be generalized by the following equation.

g (x, y) = T [f (x, y)]

= w ₁ f (x-1, y-1) + w ₂ f (x-1, y) + w ₃ f (x-1, y + 1), w ₄ f (x, y-1) + w ₅ f (x, y) + w ₆ f (x, y + 1) + w ₇ f (x + 1, y-1) + w ₈ f (x + 1, y), w ₉ f (x + 1 , y + 1)

If the pixel value corresponding to this coefficient is expressed as a column vector,

Is equal to the inner product of these two vectors,

W ^T X = w ₁ x ₁ + w ₂ x ₂ +... + w ₉ x ₉

Can be obtained. That is, Formula (I) can be expressed by the following Formula (II).

G (x, y)

= T [f (x, y)]

= W ^T X (II)

Where x ₂ , x ₂ ,... , x ₉ represents the value of the pixel corresponding to the coordinate of the coefficient w in the mask, and W ^T represents the transpose of W.

In the present invention, the following formula is defined in order to detect the brightness change of the image for each specific direction by the above formula (II), and the appropriate vector coefficient is set to extract the trace component in the desired direction.

Gh = W ^T h X

Gv = W ^T v X

Gt = W ^T t X

Where Gh, Gv and Gt are gradients in the horizontal, vertical and oblique directions, respectively, and W ^T h, W ^T v and W ^T t are vectors of coefficients corresponding to each direction. In order to extract the stroke in the inclined direction, the direction change in three directions is detected. In this case, the three directions are referred to as W, SW, and SE, respectively.

Gt = W ^T s X (III)

= W ^T sw X (IV)

= W ^T se X (V)

Where each coefficient vector is

W ^T s = w ₄ + w ₆ + w ₇ + w ₈ + w ₉ -2w _5- (w ₁ + w ₂ + w ₃ )

W ^T sw = w ₁ + w ₄ + w ₇ + w ₈ + w ₉ -2w _5- (w ₂ + w ₃ + w ₆ )

W ^T se = w ₃ + w ₆ + w ₇ + w ₈ + w ₉ -2w _5- (w ₁ + w ₂ + w ₄ )

Specify with.

Vectors for horizontal and vertical stroke extraction are

Gh = W ^T h X (VI)

= W ^t v X (VII)

And the coefficient vector is

W ^T h = w ₅ -w ₂

W ^T v = w ₅ -w ₄

Specify with.

The following describes the extraction process for each stroke.

First, an image in which three brightness changes are detected by applying equations (III) to (V), which is an extraction formula of stroke components, is obtained.

(C)-(e) of FIG. 5 are images obtained by obtaining stroke components in the SW, S, and SE directions, respectively, obtained by the above-described stroke component extraction formulas of the formulas (III)-(V).

In the present invention, the following three methods are applied to the three images obtained through the above process to extract two oblique strokes and two horizontal and vertical component strokes.

[One. Rain Intrusion Extraction]

The non-obvious extraction extracts pixel values of the same coordinates for two images in the S and SW directions obtained by the above process by applying the following logical operation.

If the non-invasive image to be obtained is B (x, y),

B (x, y) = S (x, y) * SW (x, y)

It is expressed as follows, and a non-deviating image is obtained by applying the above equation.

Fig. 5 (f) is an example of rain capture extracted by the above method.

The above results are obtained from specific directions, i.e., directions of S, SW and SE, but the same result can be obtained from images of other constant directions. In FIG. 6, the combinations of N, NW and NE directions, W, The same result can be obtained in the combination of the NW and SW directions and the combination of the E, NE and SE directions.

[2. Descending stroke extraction]

Descending stroke performs the same process as the extraction process of rain shedding, except that the S and SE directions apply the same logical operation to the two images.

If the downstroke image to be obtained is N (x, y),

N (x, y) = S (x, y) * SE (x, y)

It is expressed as follows and by applying this equation, we get the down image.

An example of extracting the downstroke by the above method is shown in FIG. 5 (g).

The same effect can be obtained by applying the above method to two images of SW (x, y) and SE (x, y).

[3. Horizontal Stroke Extraction]

The horizontal stroke can be extracted by detecting a change in the brightness of the image in the vertical direction in the image, and the change in the brightness of the image in the y direction uses the above formula (VI) and its coefficient vector.

However, when detecting the brightness conversion in the vertical direction, the oblique stroke component is also extracted as shown in FIG. 11 (c), so that the horizontal stroke component can be clearly detected only by effectively removing it. First, in order to obtain the image H ^T (x, y) from which the brightness change in the y direction is detected from the equation (VI) and its coefficient vector, and to remove the non-invasive components contained in the image.

H "(x, y) = H ^T (x, y) * B (x, y)

By processing the image H ^T (x, y) extracted from the directional stroke component in the vertical direction as shown in FIG. The removed image H "(x, y) can be obtained.

[3. Horizontal Stroke Extraction]

The horizontal stroke can be extracted by detecting a change in the brightness of the image in the vertical direction in the image, and the change in the brightness of the image in the y direction uses the above formula (VI) and its coefficient vector.

However, when detecting the erectile in the vertical direction, since the eleventh (c) and the oblique stroke component are also extracted, the horizontal stroke component can be clearly detected only by effectively removing it.

First, in order to obtain the image H ^T (x, y) from which the brightness change in the y direction is detected from the equation (VI) and its coefficient vector, and to remove the non-invasive components included in the image

H "(x, y) -H ^T (x, y) * B (x, y)

By processing the image H ^T (x, y) extracted from the directional stroke component in the vertical direction as shown in FIG. The removed image H "(x, y) can be obtained.

However, the image H "(x, y) still has a downstroke component, which must also be removed, similar to the non-invasive component removal.

H (x, y) = H "(x, y) * N (x, y)

As a result, as shown in FIG. 11 (e), an image H (x, y) having only horizontal strokes from which even down strokes are removed is obtained.

[4. Vertical stroke extraction]

The vertical stroke extraction process is similar to the horizontal stroke extraction process except that equation (VII) and its coefficient vector are used to detect the change in brightness of the image in the horizontal direction from the image, and to detect the change in brightness in the horizontal direction. There is a difference.

As in the case of the horizontal stroke extraction, the gradient stroke components are extracted together as shown in FIG. 11 (f) when the brightness change in the horizontal direction is detected.

First, to obtain the image V '(x, y) in which the brightness change in the horizontal direction is detected from Equation (VII) and its coefficient vector, and to remove the invasive components included in the image.

V "(x, y) = V '(x, y) * B (x, y)

By processing the image V '(x, y) extracted from the directional stroke component in the horizontal direction as shown in FIG. The removed image V ″ (x, y) can be obtained.

However, the image V "(x, y) still has a downstroke component, so it must be removed as well.

V (x, y) = V '(x, y) * N (x, y)

As a result, as shown in FIG. 11 (h), an image V (x, y) having only vertical strokes from which even down strokes are removed is obtained.

Next, the process of classifying candidate characters will be described in detail.

[One. Area Division by Stroke]

First, the size of the input character is normalized to the size of the horizontal × vertical M × M as shown in Fig. 12 (a), and then the size of the region is divided into L equal parts, for example, by 5 equal intervals. The initial combing area is designated as shown in Fig. 12 (b), and the final combing area is designated as shown in Fig. 12 (c), and the horizontal and vertical stroke areas are As shown in FIGS. 12 (d) and 12 (e), eight feature areas exist.

[2. Stroke Extraction by Each Area]

This process of extracting only the strokes belonging to the corresponding area is selected by the stroke extractors FS1-FS8 of the stroke extracting section 32 assigned to each area. The designation of the corresponding area is made by the controller 31 which designates the area. The data of the selected strokes is transmitted to the stroke calculator 33.

[3. Normalize Stroke Length]

Since the length of the extracted stroke is not always constant, an error may occur when classifying the type, and there is a possibility that the accuracy of the constructed reference value may be lowered. Therefore, the stroke length is normalized to prevent this. It is normalized by multiplying by the weight.

[4. Calculation of absolute value of stroke]

The absolute value of the stroke is processed by the following equation (VIII) -expression (X).

Where i = 0, 1,... where n is the number of strokes in each feature region, d = 3 is the subregion in the horizontal and vertical stroke regions, and Wi is the coefficient corresponding to each stroke.

Equations (VIII)-(X) above are used as formulas for calculating the absolute values of the combing, horizontal and vertical strokes, respectively.

[5. Classification and Candidate Character Selection]

The type of the input character is determined by comparing the absolute value and the reference value calculated in the above process.

That is, by comparing the absolute value calculated for each feature region with the reference value corresponding to the region, a character belonging to the same type of input character for each region is selected as a candidate character, and the selected candidate characters are stored in the candidate character registration register 36. Save it.

At this time, the candidate character is stored as a classification diagram, which is represented by two values.

A duplicate distribution ratio indicating how duplicated the candidate characters are distributed in the eight types of feature areas and a classification probability value for predicting how accurately the candidate characters are classified are stored in the memory area of the candidate character of the candidate character registration register 36.

[6. recognition]

This step recognizes the candidate characters with the most similar feature as the input characters among the candidate characters as input characters, and performs two stages of verification to increase the accuracy of recognition. It is specified as a binary value of 1.

The reason for specifying the recognition reliability RF is to reduce the search time and convenience in the error correction of the recognition result by human or computer.

First, the use of the feature amount for comparison is as follows.

(1) Directional feature amount per unit area

After splitting the image of the three-way trace component extracted from the input characters into N × N mesh areas, the feature amount in each mesh is compared with the feature amount in the corresponding mesh area of the candidate characters. Candidate characters with the smallest difference are selected and designated as recognition results.

The feature amount used here is the number of black pixels in each direction per mesh.

Feature

Here, βh, βv, and βb are differences in the feature amounts in the horizontal, vertical, and comb directions, HR, VR, and BR are the feature amounts of candidate characters stored, respectively, and HI, VI, and BI are the feature amounts of the input characters.

Therefore, the difference Dm between the characteristic amounts of the predetermined candidate characters and the input characters is as follows.

Dm = βh + βv + βb

Where m = 1, 2,... , k is the number of candidate characters.

(2) Ratio of black pixel per unit area

The normalized input character is divided into N × N meshes to calculate the number of black pixels in each mesh, and the value divided by the total number of black pixels is designated as the feature amount of each mesh.

If the total number of pixels of the input character is Ti and the number of black pixels of each mesh is MI (i, j), the ratio of the black pixels of each mesh is

Ri = MI (i, j) / Tr

Similarly, the ratio of black pixels per mesh

Rr = MR (i. J) / Tr

It is expressed as

Here, MR (i, j) is the number of black pixels of each mesh of the candidate characters, and Tr is the total number of black pixels of the candidate characters.

Therefore, the difference of the feature amounts of the input characters from the candidate characters is defined by the following equation.

Where m = 1, 2,... denotes the number of k candidate characters.

The recognition process is performed by sequentially comparing the extracted feature quantities, obtaining the direction feature amounts Dm per unit area for each candidate character, sorting the candidate characters in order of the size of the Dm value, and having the smallest Dm. After selecting the candidate character, if the Dm of the selected candidate character is smaller than the threshold value α1, the candidate character is recognized as an input character, and the recognition reliability is set to RF = 1, and if Dm is equal to or larger than the threshold value α1, Obtain the ratio Rm of the black pixels per unit area, which is the second recognition feature, for the smallest candidate character Dm. If Rm is less than the threshold α2, the candidate character is recognized as the input character and the reliability of recognition is set to RF = 1. If Rm is equal to or greater than the threshold, the candidate character is recognized as an input character, but the reliability of recognition is set to RF = 0.

As described above, the present invention can simplify the configuration of the system by recognizing characters using a simple technique of detecting the change in brightness of the image in a specific direction, which not only reduces the processing time but also the existing method for noise. It is stronger and improves the accuracy of recognition by verifying the recognition result in two stages, and expresses the reliability of recognition, so that the retrieval time can be shortened and it can be conveniently modified.

Claims (10)

A first process of extracting the inputted individual characters to obtain a size thereof, a second process of normalizing the characters according to the size, a third process of extracting the trace components of each direction from the normalized character image, and the characters A fourth process of dividing a into feature regions and obtaining the size of the extracted stroke in each feature region, a fifth process of selecting candidate characters and storing the candidate characters in a memory by comparing the size of the strokes obtained in the fourth process with a reference value; And a sixth step of character recognition by comparing the feature amount of the candidate character extracted in the fifth step with the feature amount of the input character. The method of claim 1, wherein the third process extracts a directional stroke component by changing brightness of a normalized text image in a specific direction, and extracts a stroke component by combing, lowering, horizontal and vertical directions by logically combining the extracted directional stroke components. Character recognition method characterized in that. The character recognition method according to claim 2, wherein the comb direction trace component B (x, y) is extracted by the following logical expression. B (x, y) = S (x, y) * SW (x, y) Here, S (x, y) and SW (x, y) are pixel values of the same coordinates for the image in the S and SW directions, respectively. 3. The character recognition method according to claim 2, wherein the downcoming stroke component N (x, y) is extracted by the following logical expression. N (x, y) = S (x, y) * SE (x, y) Here, S (x, y) and SE (x, y) are pixel values of the same coordinates for the images in the S and SE directions, respectively. The character recognition method of claim 2, wherein the horizontal stroke component is extracted by multiplying the non-fraction image and the down stroke image by the image from which the vertical stroke component is extracted. The character recognition method of claim 2, wherein the vertical stroke component is extracted by performing an AND operation on the image of the horizontal stroke component extracted from the non-fraction image and the down stroke image. The method of claim 1, wherein the fourth process divides the input character region into eight feature regions, extracts only the strokes in each feature region, normalizes the length, and adds the normalized stroke lengths to add the absolute value of the feature region. Character recognition method characterized in that specified as. 8. The character recognition method according to claim 7, wherein the length is normalized by multiplying a stroke extracted from each feature region by a weight corresponding to its size. The character recognition method of claim 1, wherein the fifth process compares an absolute value obtained for each feature region with a reference value corresponding to the corresponding region, and selects and stores a character belonging to the same type as the input character for each region as a candidate character. Way. The method of claim 1, wherein the sixth step is a first step of obtaining a direction-specific amount Dm per unit area for each candidate character, sorting by size, and selecting candidate characters having the smallest Dm, and Dm of the selected candidate character being critical. A second step of recognizing this candidate character as an input character if it is smaller than the value α1, and specifying the reliability of the recognition RF = 1, and if Dm is equal to or greater than the threshold value α1, 2 for Dm is the smallest candidate character. The third step of calculating the ratio Rm of the black pixels per unit area, which is the first recognition feature amount, and the ratio Rm of the black pixels per unit area, the second recognition feature amount, if Rm is smaller than the threshold value α2, recognizes and recognizes this candidate character as an input character. And a fourth step of designating the reliability of RF = 1 and a fifth step of designating the candidate character as an input text if Rm is equal to or greater than a threshold.