JPH0719276B2 - Character recognition device and method by stroke approximate line extraction - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention belongs to the field of character recognition, in which a stroke approximation straight line is calculated by a unique method without normalizing and thinning the size of a character image. This relates to a method of recognizing a character by extracting it.

B. Conventional Technology In the conventional technology for extracting a black region of a character, that is, a stroke for character recognition, a process of predicting and normalizing the size of a character image and a process of thinning it are necessary.

Therefore, the conventional technique has the following problems.

(1) When the size of the character image is normalized, the character characteristics may be lost.

(2) When processing such as thinning, a beard occurs,
Chips and separation of intersections may occur.

(3) In conventional character recognition by geometrical feature extraction, the number of characters to be recognized is on the order of several tens at most.

C. Problems to be Solved by the Invention (1) To accurately capture the information necessary for character recognition without normalizing the size of the character image and thinning it.

(2) To increase the reading accuracy and reduce the false recognition rate.

D. Means for Solving the Problems The unique stroke approximation straight line extracting means of the present invention is (a) a fixed number in a horizontal direction during scanning of a pattern memory in which a character to be recognized is scanned for scanning, quantized and stored. When it is detected that the connected pixels form a block that is vertically connected over two or more scanning lines, the XY coordinate of the leftmost pixel of the block is used as the starting point information, and the XY of the rightmost pixel is set. The coordinates are stored as end point information in the memory together with the horizontal stroke identifier symbol, and (b) it is detected that the vertical connectivity of pixels connected to a fixed number or less in the horizontal direction extends in the vertical direction. Then, the XY coordinate of the middle point of the pixel group in the uppermost scanning line is used as the start point information,
The XY coordinate of the middle point of the pixel group in the bottom scanning line is used as the end point information, and stored in the memory together with the vertical stroke identification symbol,
(C) When it is detected that the connectivity in the vertical direction of a certain number or less of pixels connected in the horizontal direction forms a cluster with a tendency toward leftward or rightward downward, XY of the middle point of the pixel group in the uppermost scanning line is detected. The coordinates are used as the start point information, the XY coordinates of the middle point of the pixel group in the bottom scanning line are used as the end point information, and they are stored in the memory together with the leftward or rightward diagonal stroke identification symbol, and stored in advance in the device. It is now possible to compare with a candidate table of reference characters to be recognized.

E. Embodiment FIG. 2 is a block diagram of a character recognition device to which the present invention is applied, and includes a character reading unit, a character recognition processing unit, and a recognition result utilization unit. Here, the character recognition processing unit is composed of a stroke extraction unit 1, a character identification preprocessing unit 2, a character identification unit 3, a candidate table 4 and a memory 5 according to the present invention. (1) Character reading unit The character reading unit may be a conventionally known one, and the character to be recognized is read and scanned by a character / image reader, and the obtained information is quantized and stored in a pattern memory as binary pixel data. Store. The quality of the character image data is usually sufficient for the OCR (optical character reader). Small circles in FIGS. 1 (a) and 1 (b) represent pixels stored in the pattern memory after being read and scanned for the handwritten characters 0 and 5, respectively.

(2) Stroke Extraction Section The straight lines in FIGS. 1 (a) and 1 (b) represent stroke (line segment) approximate straight lines extracted according to the present invention. This stroke approximate straight line is extracted by processing the pixels received from the character reading unit ((pattern memory) in the example of FIG. 2 in the stroke extraction unit. This is done by taking out each scan line and checking the connectivity of the geometric position with the pixel of the immediately preceding scan line.When at least 4 pixels are connected, there is connectivity (4 connections), and one block is The stroke is obtained as a unit and approximated by the following straight line having four directions.

1. Horizontal direction (HORZ) 2. Vertical direction (VERT) 2. Left falling (INC) 4. Right falling (DEC) This four-way detection is processed in parallel by one scan. This is also a feature of the present invention.

(A) Extraction of horizontal (HORZ) strokes A certain number of horizontal (HORZ) strokes are set in the horizontal direction (for example,
If two or more pixels arranged in a row are concatenated continuously in two or more rows, the rightmost pixel and the leftmost pixel in the cluster are connected to obtain a linear approximation.
For example, as shown in FIGS. 3 (a) and 3 (b). The strokes in the horizontal direction are extracted independently of the strokes in the vertical direction and the oblique directions (downward leftward, downward rightward), but their relationship must be kept down. It has been confirmed that "12 or more" shown as an example of connecting a certain number or more in the horizontal direction is an example of a handwritten character, and half or less of the above may be sufficient in the case of a print character (print type).

(B) Extraction of Vertical (VERT) Stroke If the vertical connectivity of a block of pixels satisfies one of the following conditions, a linear stroke is approximated as a vertical stroke.

1. The position where the pixel first appears on each scan line (that is, the start pixel position) is the same position as that of the immediately previous scan line. FIG. 4 (a).

2. The position where the last pixel appears on each scan line, that is, the start pixel position is the same position as that of the previous scan line.
FIG. 4 (b).

3. The start pixel position and the end pixel position on each scan line are the same as the start pixel position and the end pixel position of the immediately preceding scan line. FIG. 4 (c).

4. When the start pixel position starts on the left of the previous one and the end pixel position ends on the right of the previous one. FIG. 4 (d).

5. When the start pixel position starts to the right of the previous one and the end pixel position ends to the left of the previous one. FIG. 4 (e).

(C) Extraction of diagonal (INC / DEC) strokes If the vertical connectivity of pixels satisfies the following conditions, linear approximation is performed as a diagonal (INC / DEC) stroke.

1. Left downward stroke: The start pixel position is on the left side of the start pixel position in the previous scan line and is on the left side of it, and the end pixel position is on the left side of the end pixel position in the previous scan line. When in. FIG. 5 (a).

2. Right-down stroke: When it corresponds to the state where "left side" is replaced with "right side" in the explanation of the left-down stroke. FIG. 5 (b).

3. However, when the start and end pixel positions are the same as the immediately preceding position, the scan line ends. The central part of FIG.

(D) Processing of stroke start point A case where a pixel exists in the scan line and a pixel does not exist in the corresponding range of the immediately previous scan line is treated as a stroke start point. When a certain number (for example, 12) of pixels in one scanning line are continuously arranged, it is treated as a candidate for a horizontal stroke. If the number is less than a certain number (for example, 11 or less), remember that there was a pixel for the time being, and continue processing as a candidate for a stroke in the vertical direction / oblique direction. FIG. 1 (b).

(E) Determination of directionality The following processing is performed for strokes that do not yet have directionality. If the following conditions are not met, the decision is left to the next line.

-Conditions for horizontal strokes A certain number of consecutive pixels in the scan line (for example, 12 pixels)
When more than one line up. However, when it can be considered that the number of strokes is a certain number or more due to the intersection of a plurality of strokes, the strokes are not lateral.

-Conditions for a vertical stroke The start pixel position in each scan line must be the same, and the end pixel position must end within one digit to the right of the previous position.
Or, the end pixel position is the same position, and the start pixel position ends within one digit to the left of the immediately preceding position.

-Conditions for a stroke in the downward-left (downward-right) direction When both the start pixel position and the end pixel position are located on the left side (right side) of the position of the immediately preceding line.

(F) Processing of intersection If two or more strokes intersect during processing, the following processing is performed.

-When one of them is a lateral stroke, the lateral stroke is processed without being aware of the other strokes. The other strokes are in consecutive pixels in the current line,
The brush width up to now (the number of connected pixels in the direction perpendicular to the stroke) is added to the judgment target, and the processing is continued while maintaining the angle and directionality.

-In other cases, each stroke continues processing in the continuous pixels in the current line while maintaining the stroke width, angle and directionality up to now.

(G) Processing of branch point Each stroke continues processing in continuous pixels in the current line while maintaining the stroke width, angle and directionality up to now. In this case, when there is a pixel on the left or right of the stroke (or on both sides) that is sufficiently wider than the current stroke width, processing is started as the start point of a new stroke. FIG. 6 shows an example of processing at a branch point.

(H) End point processing As the end point of the stroke in the vertical direction / oblique direction, the above (2)
When the conditions shown in items (b) and (c) are not satisfied,
Alternatively, it is assumed that the stroke ends when there are no pixels in the range. As the end point of the horizontal stroke, when the number of pixels in the scan line in the range becomes equal to or less than a certain number (including the case where no pixel exists), the stroke is considered to have ended.

(I) About the start point / end point of the approximate line of the stroke For the stroke in the vertical direction / oblique direction, the middle point of the pixel group in the uppermost scanning line is treated as the starting point in the cluster of extracted pixels, and the pixel in the lowermost scanning line is treated. Treat the midpoint of the group as the end point. When the number of pixel groups in the scanning line is an even number, either one of the middle points, which is closer to the right or left, is adopted. Regarding the stroke in the horizontal direction, the rightmost position and the leftmost position of the extracted pixel block are treated as the start point and the end point.

(J) Processing of noise Regarding noise, preprocessing is generally performed to remove it. However, in the case of this method, when the end point is detected before the directionality during the start point processing can be recognized (that is, When the cluster of connected pixels is relatively small), the cluster of pixels can be treated as noise and removed.

The strokes extracted by the above processing are information such as stroke identification symbols indicating the directionality (HORZ, VERT, INC, DEC) for each stroke, XY coordinate values of the start and end points, and connectivity (intersection point position) if necessary. Is temporarily stored in the memory.

(3) Character Identification Preprocessing Section The information extracted by the stroke extraction section and stored in the memory is organized to obtain the final information required for character identification.

(A) Combining strokes Among continuous strokes having the same vertical and diagonal directions, those having similar angles are combined as one stroke (of the stroke information stored in the memory, the start point and The XY coordinate value of the end point is updated.). FIG. 7. Strokes with the same direction but different angles are treated as different strokes.

(B) Calculation of Number of Strokes in Each Direction After the stroke combining process, the number of strokes in each direction is calculated and stored in the memory. If necessary, the number of intersections is also calculated and stored in the memory. These pieces of information are used together with the following zone information when selecting a candidate.

(C) Extraction of zone information There are two types of zone division methods: division into strokes in the horizontal direction and strokes in the vertical direction / oblique direction.

・ Zone for horizontal stroke As shown in Fig. 8, the zone for horizontal stroke is defined by the three parts (a, b, c) divided by lines from the top 1/4 and bottom 1/4 of the pixel. And Check for lateral strokes in each of these zones.

・ Zone for vertical / diagonal strokes Of the nine parts divided by the line 2/5 from the top and bottom of the pixel and the line 1/4 from the left and right, the six shown in Fig. 9 (d, e, f , g, h,
Let i) be the zone for vertical / oblique strokes. Similarly, it is checked whether or not there is a stroke in the vertical direction / oblique direction in each zone.

(4) Character identification section A candidate is selected from a character group (called a category) to be recognized based on the number of strokes / intersection points in each direction and zone information. The character identification processing will be described below with reference to FIG.

(A) Selection of Candidates by Number of Strokes Each character in the category is stored in the memory in a table with a certain range of the number of strokes having each directionality (Table 1). Using this table, candidates are selected for what character image is being recognized.

Example: In the case of the character '5', the total number of strokes is between 5 and 9, the number of horizontal strokes is 1 to 3, the number of leftward strokes is 0 to 2, and the number of rightward strokes is 0. It shows that it is within the range from 3 to 3.

(B) Selection of Candidate Based on Zone Information Similarly, a candidate is also selected based on zone information. In the case of zone information, for each stroke, the table shows for each stroke whether it is'must exist 'or'must not exist', so that information can be used to select a candidate. it can.

Table 2 shows zone information about numbers and symbols as an example.

Example: For character '5', the horizontal stroke must be at the top (a) of the character image (symbol 1), and the vertical / oblique strokes must be at the upper left and lower right (d) of the character image. , i)
Must be in Further, symbol 2 indicates that the upper right portion (g) of the character image must not have a stroke in the vertical direction / oblique direction.

(C) Character identification processing It is investigated whether the candidates obtained by the stroke number and zone information satisfy necessary and sufficient conditions. If there are extra strokes (and if there is no intersection when the intersection information is also adopted), remove it from the candidates. After performing the above processing, when the candidate is uniquely determined (that is, the first
In the case of '= 1' in Fig. 10), this is the recognition result. When there are two or more candidates (that is, when '>1'), it is treated as unrecognizable (REJECT).

When there are no candidates (that is, when '= 0')
Changes (or relaxes) the selection condition, and performs the above-described processing with the candidates obtained by the table and the candidates obtained by the number of strokes, for example, excluding the information that "must not exist" in the zone information. For example, it is assumed that when the selection table corresponds to the categories “0” and “8”, it is determined by the check of the recognition logic that neither condition of the characters “0” or “8” can be satisfied. At this time, for example, if the condition of zone (b) is relaxed, only the character "0" remains as a candidate. If no candidate is found, the above process is repeated only for the candidate obtained from the stroke information.
(As a result, if it is not found, there is no candidate (REJEC
T). ') (5) Application to Katakana, alphabet, and kanji In the candidate selection table based on the number of strokes shown in Table 1 and the candidate selection table based on the zone information shown in Table 2, the recognition target character group (category) is extended and given. By doing so, recognition of katakana, alphabets, and kanji can be performed similarly. Examples of stroke approximation straight lines extracted from the connection of each pixel are shown in FIGS. 11 to 13.

F. Effect of the Invention The character recognition method of the present invention does not perform normalization and thinning of the size of a character image, detects four directions in one scan, and considers the stroke width to approximate strokes and characteristic points. Since it is a novel processing method for extracting characters, it is possible to realize character recognition with high accuracy and low erroneous recognition rate with a relatively simple configuration.

[Brief description of drawings]

FIGS. 1 (a) and 1 (b) are diagrams showing pixels stored in a pattern memory by scanning handwritten characters 0 and 5 and stroke approximate straight lines extracted according to the present invention, and FIG. 2 applies the present invention. Schematic view of the character recognition device, FIG. 3 (a),
(B) and (c) are examples of lateral strokes of handwritten characters, FIGS. 4 (a) to (e) are examples of vertical strokes of handwritten characters, and FIGS. 5 (a) and 5 (b) are An example of extraction of oblique strokes of handwritten characters, FIG. 6 is an example of processing branch points, FIG. 7 is an example of combining strokes, FIG. 8 is a diagram showing zones for horizontal strokes, and FIG. FIG. 10 is a diagram showing zones for oblique strokes, FIG. 10 is a diagram showing character identification processing, FIGS. 11 (a) and 11 (b) are examples of pixels and stroke approximate straight lines of handwritten katakana, and FIG. 12 (a) and FIG. 13B is a diagram showing an example of a pixel and a stroke approximation line of a handwritten alphabet, and FIGS. 13A and 13B are diagrams showing examples of a pixel and a stroke approximation line of a handwritten Chinese character. 1 ... Stroke extraction section, 2 ... Character identification preprocessing section, 3
...... Character identification unit, 4 ... Candidate table, 5 ... Memory.

Claims (5)

[Claims] 1. A unit for reading and scanning a character to be recognized, a unit for quantizing the scanning result and storing it as binary pixel information in a pattern memory, and a unit for scanning the pattern memory to connect pixels to each other. Means for extracting information of a stroke approximation straight line which is a straight line approximating a part or all of the strokes of the character to be recognized, a memory for storing the information of the extracted stroke approximation straight line, and the storage in the memory And a means for selecting a candidate character based on the information of the stroke approximate straight line, the means for extracting the information of the stroke approximate straight line is: (a) scanning the pattern memory It is detected that pixels connected by a first fixed number or more in the middle horizontal direction form a block connected by a vertical fixed number over a second fixed number or more of scan lines. When, the XY coordinates of the leftmost pixel of the cluster as the start point information, the XY coordinates of the rightmost pixel as the end point information, the start point information and the end point information are stored in the memory together with the lateral stroke identification symbol, (B) When it is detected that the connectivity in the vertical direction of the pixels connected in the horizontal direction in the first fixed number or less forms a cluster extending in the vertical direction, the middle point of the pixel group in the uppermost scanning line. (However, when the number of pixel groups is an even number, it is set to either the left or right predetermined adjacent pixel position. The same applies below). As the end point information, the start point information and the end point information are stored in the memory together with the vertical stroke identification symbol, and (c) the diagonal connectivity of the first fixed number of pixels connected in the horizontal direction decreases to the left or Form a lump with a downward sloping tendency When it is detected that the start point information is the XY coordinate of the middle point of the pixel group in the uppermost scanning line, and the XY coordinate of the middle point of the pixel group in the lowermost scanning line is the end point information. The information is stored in the memory together with the slanting stroke identification symbols that are descending to the left or to the right, and the means for selecting the candidate character is based on the stroke approximate straight line information stored in the memory. The number of stroke approximate straight lines and their total number are calculated, and compared with the number in the candidate table indicating the number of stroke approximate straight lines in each direction of the reference character to be recognized and the total number of them, which are stored in advance in the device. Character recognition device by stroke approximation straight line extraction, which selects the corresponding candidate character by 2. A means for reading and scanning a character to be recognized, a means for quantizing the scanning result and storing it in a pattern memory as binary pixel information, and a pattern memory for scanning the pattern memory to connect pixels to each other. Means for extracting information of a stroke approximation straight line which is a straight line approximating a part or all of the strokes of the character to be recognized, a memory for storing the information of the extracted stroke approximation straight line, and the storage in the memory A character recognition device having a means for selecting a candidate character based on the information of the stroke approximated straight line, reads and scans the character to be recognized, quantizes the scanning result, and stores it in the pattern memory as binary pixel information. Information of a stroke approximate straight line which is a straight line which is obtained by scanning the pattern memory and approximating a part or all of the strokes of the character to be recognized from the connectivity between pixels. A character recognition method by stroke approximation straight line extraction for extracting and storing in the memory, and selecting a candidate character based on the stroke approximation straight line information stored in the memory, wherein the stroke approximation straight line information is extracted. (A) It is detected that pixels that are connected in a horizontal direction in a first fixed number or more during scanning of the pattern memory form a block in which the pixels are connected in a vertical direction over a second fixed number or more of scan lines. At this time, the XY coordinate of the leftmost pixel of the cluster is used as start point information, the XY coordinate of the rightmost pixel is used as end point information, and the start point information and the end point information are stored in the memory together with the lateral stroke identification symbol, b) Pixels in the uppermost scanning line when it is detected that the vertical connectivity of the pixels connected in the first fixed number or less in the horizontal direction extends in the vertical direction. Of the middle point of the pixel group in the bottom scan line The XY coordinates are used as end point information, and the start point information and the end point information are stored in the memory together with a vertical stroke identification symbol, and (c) the diagonal connectivity of the first fixed number of pixels connected in the horizontal direction is determined. When it is detected that a lump with a tendency of falling to the left or falling to the right is formed, the XY coordinate of the middle point of the pixel group in the uppermost scanning line is used as the start point information, and the XY of the middle point of the pixel group in the lowermost scanning line is used. The start point information and the end point information are stored in the memory together with the leftward or rightwardly slanting direction stroke identification symbol using the coordinates as the end point information, and the step of selecting the candidate character is performed in the memory. Of stored stroke approximation line The number of stroke approximate straight lines in each direction and the total number thereof are calculated from the information, and the number of stroke approximate straight lines in each direction of the reference character to be recognized and the total number thereof are stored in the device in the candidate table. Character recognition method by stroke approximate straight line extraction, which selects the corresponding candidate character by comparing with the number of. 3. The pixel connectivity is any one of the following conditions (i) to (v): (i) The position where a pixel first appears on each scanning line (hereinafter, referred to as a start pixel position) is immediately before. When the position where the pixel appears at the end of each scan line (hereinafter referred to as the end pixel position) is the same position as the end pixel position of the immediately preceding scan line (Iii) When the start pixel position and the end pixel position on each scan line are the same as the start pixel position and the end pixel position of the immediately previous scan line (iv) The start pixel position on each scan line is the last scan line When starting from the left side of the starting pixel position and ending at the ending pixel position on the right side of the ending pixel position of the immediately preceding scanning line (v) the starting pixel position on each scanning line is the starting pixel position of the immediately preceding scanning line From the right side It is detected that the connectivity of the pixels forms a cluster extending in the vertical direction when the end pixel position satisfies the condition that the end pixel position ends on the left side of the end pixel position of the previous scan line. 3. The character recognition method according to claim 2, wherein the stroke in the vertical direction is linearly approximated. 4. The connectivity of pixels detected to form a lump extending in the vertical direction satisfies any one of the following conditions (i) to (ii): (i) start pixel position on each scanning line Is at the same position as or on the left side of the start pixel position of the immediately preceding scan line, and the end pixel position is at the same position or at the left side of the end pixel position of the immediately previous scan line (ii) Start pixel on each scan line If the position is at the same position as the start pixel position of the immediately preceding scan line or at the right side thereof and the end pixel position is at the same position as the end pixel position of the immediately previous scan line or at the right side thereof, connection of the pixels is satisfied. The character recognition method by stroke approximation straight line extraction according to claim 3, wherein the strokes in the oblique direction are linearly approximated by detecting that the character forms a group of downward-sloping or downward-sloping groups. 5. The character by stroke approximate straight line extraction according to claim 4, wherein when the start and end pixel positions are the same as those of the immediately preceding scan line, the approximate straight line is ended at the scan line. Recognition method.