JPS62125485A - Character recognizing system - Google Patents

Abstract

PURPOSE:To obtain the same feature matrix even in the Ming style and the Gothic style and to execute the stable character recognition to the deformation of a character and the change of a character line width by dividing the character frame area into the area of (NXM) concerning the subpattern extracted from an original pattern, and forming the feature matrix based upon the black point in the area and the line width of respective subpatterns. CONSTITUTION:A line width calculating part 4 calculates the line width of an input pattern. A vertical subpattern extracting part 6 extracts the vertical subpattern from the relation of the continuous length and the line width of a black bit. A vertical subpattern line width calculating part 10 calculates the line width of the vertical pattern and in the same way, calculates the horizontal and right and left slanting subpattetrn. A feature matrix extracting part 14 forms the feature matrix from these, normarizes it by the size of the character from a character frame detecting pat 5 and prepares the characteristic matrix. The feature matrix is compared with a standard feature matrix at an identifying part 15, and the category name of the standard character matrix to give the minimum distance is given as a character name output 16.

Description

Detailed Description of the Invention (Industrial Field of Application) The present invention relates to a character recognition method for recognizing character figures, and more specifically, a character figure/ξ turn is divided into a plurality of sub-patterns, and each sub-noise turn is The present invention relates to a method of obtaining a feature matrix for each character and comparing this feature matrix with a standard character mask in a dictionary for identification.

(Prior Art) Conventionally, in character/figure recognition devices, strokes are extracted from character/figures/turns, and recognition is performed using the positions, lengths, and mutual relationships between strokes of the extracted strokes. There are nine people who are often employed.

The method is (1) Calculate the curvature of the contour point series detected by tracing the contour of the character figure, divide the contour series using points with large values of curvature as dividing points, and assemble the divided series. (2) Extract the strokes by combining them, or (2) perform thinning processing on the character/figure pattern to create a skeleton, trace the connectivity of the skeleton pattern and detect sudden angle changes, etc., and extract the strokes. The strokes extracted from (1) and (2) above are identified by extracting their geometric features. However, in method (1), as the character/graphic pattern becomes larger and the character/graphic/pattern becomes more complex, the processing band increases, resulting in a reduction in processing speed. Also (2)
In this method, it is necessary to thin the character shapes and ε-turns,
Further, due to the thinning of the wire, there are problems such as the occurrence of distortion and hides at the /-O turn, which complicates subsequent processing.

In order to solve these problems, the applicant has proposed
A character recognition method disclosed in Japanese Patent No. 7-231-85 is proposed. This method (a) creates an original pattern of a digital signal represented by black bits and white pits by photoelectrically converting and quantizing the character shape, and (b) creates the original pattern of the entire character shape of this original ε turn. Calculate the line width (for example, average value), (c) scan this original pattern in multiple directions to detect the number of consecutive black bits in each scanning line, and compare this number of consecutive black bits with the line width. A plurality of subpatterns corresponding to each of the plurality of directions are extracted based on the width, and (d) the area within the character frame of the previous original pattern is extracted for the subpattern (N×
M) regions (N and M are constants), further calculate the feature amount representing the character line length in each region using the line width of the original pattern in units of cells, and (e) Normalize the feature amounts by the font size and create a feature matrix),
f) This feature matrix is compared with a standard character mask of character and graphic patterns prepared in advance to recognize characters and figures.

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

For example, when recognizing Gothic fonts and Mincho fonts in printed characters, since the vertical line width and horizontal line width of Gothic fonts are almost constant, the line width of the entire character is equal to the vertical line width or horizontal line width, and the cells are The characteristics remain stable even when normalized using the line width of the entire character as a unit. On the other hand, since Mincho fonts have large differences in vertical and horizontal line widths, if we normalize the line width of the entire character in units of cells, even if the character lines have the same line length as Gothic fonts, the horizontal feature width will be significantly different compared to Gothic fonts. ) The amount of IJ soup is small,
The amount of vertical feature matrix becomes large and is unstable as a feature.

Second, examples of subpatterns for Gothic and Mincho fonts are shown.

A similar phenomenon applies to handwritten characters, and even character lines of the same length can have different feature quantities, making the characteristics unstable, depending on the state of the writing instrument and the force applied by the writer to the writing instrument.

In order to solve this problem, it is necessary to increase the number of dictionaries for identification, which in turn leads to a reduction in processing time.

Therefore, it is an object of the present invention to solve these problems and provide a high-speed and stable character recognition method.

(Means for Solving the Problems) The present invention includes the following components (a) to (f).

(,) By photoelectrically converting and quantizing character figures, an original pattern of a digital signal represented by black bits and white bits is created.

(b) Next, calculate the line width of the original pattern.

(c) Next, the original pattern is scanned in a plurality of directions to detect the number of consecutive black bits in each scanning row, and based on the number of consecutive black bits and the line width, the number of consecutive black bits is detected in the plurality of scanning directions. A plurality of sub-patterns corresponding to each pattern are extracted.

(d) Next, sub-ε
The turn is divided into (N x M) areas (N and M are constants), and based on the result of counting black points in units of cells within the divided area and the line width of each sub-ε turn. Calculate features.

(e) Next, the feature amount is normalized by the character size to create a feature matrix.

(f) Then, the IJ mask is compared with a standard character mask of a character/figure pattern prepared in advance to recognize the character/figure.

(Function) The component (a) has the function of converting characters and figures to be identified into binary digital signals to create an original pattern.

Component (b) functions to calculate the line width (average value) of the original pattern.

Component (c) functions to extract a plurality of sub-patterns from the original pattern.

Component (d) functions to create a feature matrix for each sub-turn. In particular, calculating the feature amount based on the number of black dots and the line width of each sub-pattern has the effect of absorbing variations in character width and the like.

Component (e) acts to normalize the feature matrix.

Component (f) has the function of identifying an input character figure by comparing the obtained feature matrix with a standard character mask of a character figure pattern prepared in advance.

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

FIG. 1 is a block diagram showing one embodiment of the present invention.

In the figure, 1 is an optical signal input, 2 is a photoelectric conversion unit, 3 is a turn noster, 4 is a line width calculation unit, 5 is a character frame detection unit, 6 is a vertical/turn extraction unit, and 7 is a horizontal pattern extraction unit. Part, 8 is right diagonal sub-J? 9 is a left diagonal sub-pattern extraction section; 10 is a vertical sub-turn line width calculation section; 11 is a horizontal sub-turn line width calculation section; 12 is a right diagonal sub-pattern line width calculation section; Left diagonal sub iJ? 14 is a feature matrix extraction section, 15 is an identification section, and 16 is a character name output.

Next, the configuration of each part will be explained. The photoelectric conversion unit 2 converts the optical signal input of the original pattern into a binary quantized electric signal. A' Turnle mask 3 Stores the electrical signal of the cigarette. During this storage, the characters are divided into, for example, 100×100 cells, and each binary code is stored in the pattern renoster 3.

The line width calculation unit 4 calculates the line width of the input turn (original pattern). The vertical sub-pattern extraction unit 6 is
The mask 3 is vertically scanned all times to extract a vertical sub-pattern (vsp) based on the relationship between the continuous length of black bits and the line width calculated by the line width calculation unit 4. Similarly, the horizontal sub-pattern extractor 7 extracts the horizontal sub-pattern (I (SP)) by horizontal scanning, and the right diagonal sub-pattern extractor 8 extracts the right diagonal sub/'P turn (H8P) by scanning right diagonally (45°). , the left diagonal sub-pattern extraction unit 9 extracts a left diagonal sub-pattern (LSP) by performing a left diagonal (45°) scan.

FIG. 3 shows an example of an original pattern and a sub-pattern. Third
In the figure, (a) is the original pattern, (b) is the vsp, (c) is the H8P, (d) is the H8P, and (e) is the LSP. The character frame detection unit 5 detects a character frame that circumscribes the character pattern in the p-turn soster and sends the result to the feature matrix extraction unit 10. The feature matrix extraction unit 10 is a sub ie p
- For Nrenostar, the area corresponding to the original 78-turn character frame is an area of (N×M) (in the embodiment of the present invention, N=M
=5). The character is divided into 100x100 cells, and if N=M=5, each region has 20x20 cells.

Here, a method for extracting a feature matrix will be explained using VSP as an example. The vertical sub-pattern line width calculation unit 10
The line width (Wv) of the vertical sub-pattern extracted by the vertical sub-pattern extraction section 6 is calculated by the same process as the line width calculation section 4. The feature matrix extraction unit 14 calculates the number of black points (Bij) of the divided area described above in units of cells, and calculates the line width (w) of vertical sub-turns based on the following formula (1).
sub iJ in each region according to the following equation (2) using
? A feature amount representing the length of the stroke of a turn is calculated in units of cells, and a (N×M) dimensional feature matrix is created.

Number of sunspots = line length x line width ... (1) ti
j = Bij/Wv
- (2) Similar processing is performed for H3P by using the horizontal sub-pattern line width ("o) calculated by the horizontal sub-pattern line width calculation section 11, and for H8P by using the right diagonal sub-pattern line width calculation section 12. The calculated line width (wR) of the right diagonal sub-pattern is used, and for LSP, the line width (wL) of the left diagonal sub-pattern calculated by the left diagonal sub-pattern line width calculation unit 13 is used to generate the feature matrix. create.

Next, the feature matrix extraction unit 14 extracts the extracted feature matrix.
- Normalize IJ Kus by character size and create a normalized feature matrix. The method is to use the feature matrix before normalization)
One element of IJ is 11j, and the element after normalization is Li1.
, the horizontal length of the character frame is ΔX, and the vertical length is Δ
If Y is selected, the following processing will be performed.

(1) For vertical subpattern (vsp) matrix L
i j = ti j/ΔY...
・(3) (2) In case of horizontal sub-pattern (RSP) matrix L1j=tjj/ΔX ・・
・(4)(3) Diagonal sub-patterns (RSP, LSP
) For matrix, Li j ”” 'i j/(ΔX)2+(Δy)
2 (5) Through the above processing, the feature matrix extraction unit 14 finally expresses the original pattern ((N×
M) Create a feature matrix whose dimensions are normalized by X4'.

The identification unit 15 calculates the distance (D) between the standard character mask (fm) and the feature matrix (fl) extracted by the feature matrix extraction unit 14 according to the formula (6) below.
That is, the category name of the standard character mask that gives the minimum length of the difference vector between the two vectors between the (N×M)X4-dimensional feature is output as the character name output 12.

D=7 penalty - D17 (6) Next, the operation will be explained. The optical signal input 1 is photoelectrically converted by a photoelectric converter 2 and supplied to a pattern renoster 3 and a line width calculator 4. The line width calculation section 4 calculates the line width of each original turn and outputs it to each subpattern extraction section 6 to 9. Vertical sub-
The turn extraction section extracts vsp based on the output of the pattern renostar 3 and the output of the line width calculation section 4 using the method described above. Similarly, the horizontal sub-turn extraction section 7,
The right diagonal subpattern extraction unit 8 and the left diagonal subpattern extraction unit 9 extract RSP, RSP, and LSP, respectively. The vertical sub-pattern line width calculation unit 10 calculates the line width (Wy) of vsp. Similarly, the horizontal sub-pattern line width calculation section 11, the right diagonal sub-pattern line width calculation section 12, and the left diagonal sub-pattern line width calculation section 13 each perform an RSP.
line width (W, , ), RSP line width (W□) and LSP
Calculate the line width (WL) of Then, the feature matrix extraction unit 14 creates a feature matrix according to equations (1) and (2) above, and further normalizes it with the character size output from the character frame detection unit 5 to create a feature matrix. . The feature 7 matrix created in this manner is compared with the standard character mask in the identification unit 11, and the category name of the standard character mask that provides the minimum distance is output as the character name output 12.

(Effect of the invention) As explained above, the character frame area is divided into (N×M) areas for the subpattern extracted from the original pattern,
Furthermore, a feature matrix is created based on the black dots within the divided area and the line width of each sub-pattern, and a feature matrix is extracted by normalizing this with the font size, so even Mincho and Gothic fonts can be used. The same feature matrix is obtained, and stable character recognition is possible despite character deformation and character line width variations. Furthermore, since the feature matrix is extracted by simple scanning of subpatterns extracted from the original pattern, high-speed character recognition is possible.

Ru.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing examples of sub-patterns of Kosic typeface and Mincho typeface, and FIG.
The figure shows an example of extraction of original patterns and sub-patterns. 1... Optical signal input, 2... Photoelectric conversion section, 3...
Zeturn 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 Part, 9... Left diagonal sub pattern extraction part, 10...
Vertical sub-pattern line width calculation section, 11...Horizontal sub-pattern'
? Turn line width calculation unit, 12... Right diagonal sub/p turn line width calculation unit, 14... Feature matrix extraction unit, 15.
...Identification part, 16...Character name output. Patent applicant: Oki Electric Industry Co., Ltd. Patent application agent: Patent attorney Megumi Yamamoto - Goshi Noah
Leap

Claims (1)

[Claims] (a) Create an original pattern of a digital signal represented by black bits and white bits by photoelectrically converting and quantizing character figures; (b) Calculate the line width of the original pattern; (c) scanning the original pattern in a plurality of directions to detect the number of consecutive black bits in each scanning row, and scanning the plurality of scans based on the number of consecutive black bits and the line width of the original pattern. extracting a plurality of sub-patterns corresponding to each direction; (d) dividing the region within the character frame of the original pattern into (N×M) regions (N and M are constants) for the sub-pattern;
A feature amount is calculated based on the result of counting black points in each cell in the divided area and the line width of each sub-pattern, and (e) the feature amount is normalized by the character size to create a feature matrix. (f) A character recognition method characterized by recognizing a character figure by comparing the feature matrix with a standard character mask of a character figure pattern prepared in advance.