JPH0664629B2 - Character recognition method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a character recognition system for recognizing a character graphic, and more specifically, a character graphic pattern is divided into a plurality of sub-patterns and features of each sub-pattern. The present invention relates to a method of obtaining a matrix and identifying the feature matrix by comparing the feature matrix with a standard character mask in a dictionary.

(Prior Art) Conventionally, in a character / graphics recognizing device, a method of extracting strokes from a character / graphics pattern and recognizing them by using the positions, lengths, mutual relationships between the strokes, etc. are often adopted. Has been done. The method is as follows: (1) Calculate the curvature of the contour point sequence detected by tracing the contour of the character figure, divide the contour sequence with the point with a large curvature value as the dividing point, and combine the divided sequences. The strokes are extracted by combining them, or (2) the character / graphic pattern is thinned to form a skeleton, and the connectivity and skeleton pattern of the skeleton pattern are tracked to detect a sudden angle change point, etc. The strokes extracted and extracted in (1) and (2) above are identified by extracting geometrical features and the like. However, in the method (1), when the character / graphic pattern becomes large and the character / graphic pattern becomes complicated, the processing amount increases and the processing speed decreases. In the method (2), it is necessary to make the character / graphic pattern thin, and there is a problem that the thinning of the pattern causes distortion and whiskers in the pattern, making subsequent processing complicated.

In order to solve these problems, the applicant of the present invention has disclosed in
-23185 proposes the character recognition method disclosed. In this method, the character matrix of the sub-pattern extracted from the character pattern is divided into (N × M) regions to collate the extracted feature matrix of each stroke with the dictionary. High-speed and stable character recognition can be performed.

(Problems to be Solved by the Invention) However, the conventional character recognition method described above has the following problems.

For example, when recognizing the Gothic font and the Mincho font of printed characters, the vertical line width and horizontal line width of the Gothic line are almost constant, so the line width of the entire character is equal to the vertical line width or horizontal line width,
The characteristics are stable even if the line width of the entire character is normalized using the cell as a unit. On the other hand, the vertical and horizontal line widths of Mincho type are very different, so if you normalize by the line width of the entire character in units of cells, even a character line with the same line length as Gothic type will have a horizontal feature matrix compared to Gothic type Is small and the amount of the vertical feature matrix is large, and the feature is unstable.

Figure 2 shows examples of Gothic and Mincho subpatterns. The same phenomenon applies to handwritten characters, and even if a character line has the same line length depending on the state of the writing instrument and the force of the writer's ability to put it in the writing instrument, the features are different and the features are unstable. In order to solve this problem, it is necessary to increase the number of dictionaries for the identification, which eventually leads to a decrease in processing time.

Therefore, the present invention solves these problems and provides a fast and stable character recognition system.

(Means for Solving the Problems) According to the present invention, an original pattern of a digital signal represented by black bits and white bits is created by photoelectrically converting and quantizing a character graphic, and a line of the original pattern is created. The width is calculated, the original pattern is scanned in a plurality of directions to detect the number of consecutive black bits in each scanning row, and the number of consecutive black bits is detected and based on the calculated line width of the original pattern. A plurality of sub-patterns corresponding to each of the plurality of scanning directions are extracted, and the area within the character frame of the original pattern is divided into (N × M) areas (N and M are constants) for each extracted sub-pattern. Divide and count the number of black dots in the divided area in units of cells, calculate the feature amount based on the counted number of black points, normalize the calculated feature amount with the size of the character to form a feature matrix. Created and the created features Concerns a recognized character recognition system graphic character against the standard character masks prepared in advance character graphic pattern Torikusu.

In particular, according to the present invention, the line width is calculated for each extracted sub-pattern, and the feature amount is calculated by dividing the counted number of black dots by the line width for each extracted sub-pattern.

(Operation) The character pattern to be identified is binarized to create the original pattern of the digital signal. Next, the line width (average value) of this original pattern is calculated. On the other hand, a plurality of sub patterns corresponding to a plurality of scanning directions are extracted from this original pattern, and the line width of each extracted sub pattern is calculated. Further, the area within the character frame of the original pattern is divided into (N × M) areas (N and M are constants) for each extracted sub-pattern. Then, the feature amount is calculated for each extracted sub-pattern. The calculation of the feature amount is performed by dividing the number of black dots counted in units of cells in the divided area by the line width of each sub-pattern calculated for each extracted sub-pattern. The feature amount calculated in this way is normalized into a feature matrix, and the created feature matrix is collated with a standard character mask of a character / figure pattern prepared in advance to recognize a character / figure.

(Embodiment) Hereinafter, the present invention will be described in detail based on an embodiment with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the present invention.
In the figure, 1 is an optical signal input, 2 is a photoelectric conversion unit, 3 is a pattern register, 4 is a line width calculation unit, 5 is a character frame detection unit, 6 is a vertical sub-pattern extraction unit, and 7 is a horizontal sub-pattern extraction unit. , 8 is a right diagonal sub-pattern extraction unit, 9 is a left diagonal sub-pattern extraction unit, 10 is a vertical sub-pattern line width calculation unit, 11 is a horizontal sub-pattern line width calculation unit, 12 is a right diagonal sub-pattern line width calculation unit, 13 is a left diagonal sub-pattern line width calculation unit, 14 is a feature matrix extraction unit, 15 is an identification unit, and 16 is a character name output.

Next, the configuration of each unit will be described. The photoelectric conversion unit 2 converts the optical signal input of the original pattern into a binary quantized electric signal. The pattern register 3 stores this electric signal. At the time of this storage, the character is divided into, for example, 100 × 100 cells, and each binary code is stored in the pattern register 3. The line width calculation unit 4 uses the input pattern (original pattern)
Calculate the line width of. The vertical sub-pattern extraction unit 6 performs a vertical scan over the pattern register 3 and extracts a vertical sub-pattern (VSP) from the relationship between the continuous length of black bits and the line width calculated by the line width calculation unit 4. To do. Similarly, the horizontal sub-pattern extraction unit 7 detects the horizontal sub-pattern (HSP) by the horizontal scan and the right oblique sub-pattern extraction unit 8
The right diagonal sub-pattern (RSP) is scanned by the right diagonal (45 °) scan, and the left diagonal sub-pattern extraction unit 9 scans the left diagonal (45).
°) Extract left diagonal sub-pattern (LSP) by scanning. FIG. 3 shows an example of the original pattern and the sub-pattern. In Fig. 3, (a) is the original pattern, (b) is the VSP, (c) is the HSP,
(d) is RSP and (e) is LSP. The character frame detector 5 detects a character frame circumscribing the character pattern in the pattern register and sends the result to the feature matrix extractor 10. The feature matrix extraction unit 10 divides the area corresponding to the character frame of the original pattern in the sub-pattern register into (N × M) areas (N = M = 5 in the embodiment of the present invention). For example, the letter 10
It is divided into 0 × 100 cells, and when N = M = 5, each region has 20 × 20 cells.

Here, taking VSP as an example, a method for extracting a feature matrix will be described. The vertical sub-pattern line width calculation unit 10 calculates the line width (w _v ) of the vertical sub-pattern extracted by the vertical sub-pattern extraction unit 6 by the same process as the line width calculation unit 4. The feature matrix extraction unit 14 calculates the number of black dots (B _ij ) in the divided area described above in units of cells, and
Using the line width (w _v ) of the vertical sub-pattern based on the equation (1), the feature quantity expressing the stroke length of the sub-pattern in each area is calculated by the following equation (2) in units of cells. ,
A (N × M) -dimensional feature matrix is created.

Number of black dots = line length × line width (1) l _ij = B _ij / w _v (2) For the HSP, the horizontal sub-pattern line width calculated by the horizontal sub-pattern line width calculation unit 11 ( w _H ), and for RSP, the right diagonal sub-pattern line width calculator 12
Using the line width (w _R ) of the right diagonal sub-pattern calculated in, the left diagonal sub-pattern line width calculator 13 is used for LSP.
A feature matrix is created by using the line width (w _L ) of the left diagonal sub-pattern calculated in.

Next, the feature matrix extraction unit 14 normalizes the extracted feature matrix by the character size, and creates a normalized feature matrix. The method is as follows when one element of the feature matrix before normalization is l _ij , the element after normalization is L _ij , the horizontal length of the character frame is ΔX, and the vertical length is ΔY. Do.

(1) In case of vertical sub-pattern (VSP) matrix L _ij = l _ij / ΔY… (3) (2) In case of horizontal sub-pattern (HSP) matrix L _ij ＝ l _ij / ΔX… (4) (3) Oblique For sub-pattern (RSP, LSP) matrix Through the above processing, the feature matrix extraction unit 14 finally creates a {(N × M) × 4} -dimensional normalized feature matrix expressing the original pattern.

The discriminating unit 15 uses the distance (D) of the following formula (6) conventionally used between the standard character mask (f _m ) and the characteristic matrix (f _i ) extracted by the characteristic matrix extracting unit 14, That is, the category name of the standard character mask that gives the minimum value of the difference vector length of two vectors in the (N × M) × 4 dimensional feature space is output as the character name output 12.

Next, the operation will be described. The optical signal input 1 is photoelectrically converted by the photoelectric conversion unit 2, and the pattern register 3 and the line width calculation unit 4
Is supplied to. The line width calculation unit 4 calculates the line width of the original pattern and outputs it to each of the sub pattern extraction units 6 to 9. The vertical sub-pattern extraction unit uses the output of the pattern register 3 and the output of the line width calculation unit 4 to perform VSP according to the method described above.
To extract. Similarly, the horizontal sub-pattern extraction unit 7,
The right diagonal sub-pattern extracting unit 8 and the left diagonal sub-pattern extracting unit 9 extract HSP, RSP and LSP, respectively. The vertical sub-pattern line width calculation unit 10 calculates the VSP line width (w _v ). Similarly, the horizontal sub-pattern line width calculation unit 11, the right diagonal sub-pattern line width calculation unit 12 and the left diagonal sub-pattern line width calculation unit 13 respectively have an HSP line width (w _H ) and an RSP line width (w _R ) And the line width (w _L ) of the LSP. And
The feature matrix extraction unit 14 creates a feature matrix according to the equations (1) and (2), and further normalizes it with the size of the character output from the character frame detection unit 5 to create a feature matrix. The feature matrix created in this way is the identification unit
The category name of the standard character mask that gives the minimum distance is compared with the standard character mask at 11 and is output as the character name output 12.

(Effect of the Invention) As described in detail above, according to the present invention, the original pattern is scanned in a plurality of directions to extract each sub-pattern, and the line width is calculated and counted for each of the obtained sub-patterns. The feature matrix is calculated by dividing the number of black dots by the line width of each of the extracted sub patterns. Therefore, even if the horizontal line width and the vertical line width are significantly different (the line width of the horizontal line is thin and the line width of the vertical line is thick), for example, in the case of Mincho type, the horizontal feature depends on the line width of the horizontal sub-pattern. Since the matrix is calculated and the vertical feature matrix is calculated according to the line width of the vertical sub-pattern, a feature matrix similar to the case where the horizontal line width and the vertical line width are the same as in Gothic is obtained. That is, regardless of the typeface of the character, stable feature extraction and therefore stable character recognition can be performed even when the character is deformed or the character line width is changed.

Further, since the feature matrix is extracted by simple scanning for the sub-pattern extracted from the original pattern, it goes without saying that high-speed character recognition is possible.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing examples of Gothic and Mincho type subpatterns, and FIG. 3 is a diagram showing an example of extracting original patterns and subpatterns. . 1 ... Optical signal input, 2 ... Photoelectric conversion unit, 3 ... Pattern register, 4 ... Line width calculation unit, 5 ... Character frame detection unit, 6
...... Vertical sub-pattern extraction unit, 7 ... Horizontal sub-pattern extraction unit, 8 ... Right diagonal sub-pattern extraction unit, 9 ... Left diagonal sub-pattern extraction unit, 10 ... Vertical sub-pattern line width calculation unit, 11 ... … Horizontal sub-pattern line width calculator, 13 …… Left diagonal sub-pattern line width calculator, 12 …… Right diagonal sub-pattern line width calculator, 14 …… Feature matrix extractor, 15 …… Identifier, 16 …… Character name output.

Claims (1)

[Claims] 1. An original pattern of a digital signal represented by a black bit and a white bit is created by photoelectrically converting and quantizing a character pattern, and a line width of the original pattern is calculated. The number of consecutive black bits in each scanning row is detected by scanning in the direction of, and corresponding to each of the plurality of scanning directions based on the detected number of consecutive black bits and the calculated line width of the original pattern. Extracting a plurality of sub-patterns, dividing the character frame area of the original pattern into (N × M) areas (N and M are constants) for each of the extracted sub-patterns, and regarding the divided areas Black dots are counted in units of cells, a feature amount is calculated based on the counted number of black dots, the calculated feature amount is normalized by the size of a character to create a feature matrix, and the created feature matrix is set in advance. Prepared letter diagram In a character recognition method for recognizing a character figure by collating with a standard character mask of a pattern, a line width is calculated for each of the extracted sub patterns, and the counted number of black dots is divided by the line width for each of the extracted sub patterns. A character recognition method characterized in that the feature amount is calculated by the above.