JPS6330666B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character separation method, and in particular, when a plurality of characters are written in handwritten manuscripts, drawings, etc. in close proximity to each other so that they cannot be separated into individual characters with a mask of a certain shape. This invention relates to a character separation method that allows each character to be separated into individual characters even when the characters are different from each other.

For example, when freely handwritten characters are recognized by a character reading device, the first step is to separate and extract each character, and then to identify the extracted characters. When separating and extracting characters in this way, for example, a rectangular mask M as shown in Figure 1A was used, and this was scanned over the character string as shown in Figure 1B to extract individual characters. . In this case, as shown in Figure 1 B, if multiple characters are written apart from each other and in a certain size, they can be written using a mask of a certain shape such as mask M. Since each character can be separated, it is possible to separate and extract these characters from each other.

However, when multiple characters are written close to each other as shown in Figure 1C, it is no longer possible to accurately separate these characters using the method of simply scanning a mask as shown in Figure 1B. Therefore, in such a case, it is difficult for a character reading device to accurately recognize the characters.

Therefore, an object of the present invention is to provide a character separation method that can accurately separate each character even when a plurality of characters are written close to each other. To this end, the character separation method of the present invention includes an information holding means that holds image information of characters input from an input means, and an information holding means that scans the character information in a first direction to find black and white change points. a first feature extracting means for detecting a first region that is an area between the changing points; and scanning the image information in a second direction substantially orthogonal to the first direction to obtain the black and white changing points; a second feature extracting means for detecting a second region that is a region between the change points; and a second feature extraction means that sequentially scans in the first direction an intermediate region that is a region where the first region and the second region do not overlap. scanning means for obtaining a line segment; midpoint setting means for determining a midpoint in each scanned line segment of the intermediate area; and a horizontal line that is close to the image information of the character and generates a horizontal line in a first direction from the midpoint. The present invention is characterized in that it has a generating means, and character areas in which individual character information exists can be distinguished by the intermediate point string and the horizontal line.

Before explaining the present invention in detail below, the principle of the present invention will be explained based on FIGS. 2 to 14.

(a) First, image data m ₀ obtained by reading a manuscript or drawing with handwritten characters
Prepare memories m ₁ and m ₂ in which . Then, the memory m ₁ is
Scan in the direction. At this time, image data _m0 is obtained by setting black dots where characters are drawn as "1" and places where no characters are drawn as "0". and the second
As shown in Figure B, the point of change, i.e. "0" →
Find "1" and the point where "1" changes to "0". For example, as shown in Figure 2 (b), the line l
When scanning above, x ₁ changes from "0" to "1", x ₂ changes from "1" to "0", x ₃ changes from "0" to "1", x ₄ Then "1" → "0"
Therefore, all of these x ₁ to x ₄ are changing points. In this case, the change point P ₁ (“1” →
0) and P ₂ (“0”→“1”), and all areas between the change points P ₁ and _{P 2} on the memory m ₂ are set to “1”.

(b) Next, the image data m ₀ on the memory m ₁ is scanned in the y direction as shown by the chain line in Fig. ₂ direction)
and Q ₂ (“0” → “1”), and thereby the area on memory m ₂ between Q ₁ and _{Q 2} is changed to “1” if “0” and “1” if “ If it is "1", it is reversed to "0". As a result, as shown in FIG. 3, the double line area becomes "0", the single line area becomes "1", and the data shown in FIG. ₄ is written in the memory m2. At this time, the character part is usually a series of "1"s, so it remains as is. In this way, the data in the state where the shaded area in FIG. 4 is "1" is held in the memory _m2 .

(c) Next, clustering is performed on the areas in which the data in the state shown in FIG. 4 has been entered, and label symbols 1 to 8 are assigned to each area as shown in FIG.

Next, this clustering will be explained.

(c-1) 3×2 mask M ₂ as shown in Figure 6 A
is scanned in the x direction as shown in the same figure. When the mask M ₂ is scanned for the shaded area shown in FIG. 4, it is checked whether there is a point given a label symbol in the sections c to f of the mask M ₂ . If it exists, give that label symbol to the category a, but
If there is no point with a label symbol, a new label symbol, for example, label signal 1, is assigned to the segment f, and based on this, a label signal 1 is assigned to the segment a.
Grant. Then, when the mask M ₂ moves in the x direction, the previous section a of the section f will now be located, and this new section f is given the label symbol 1, so
Label symbol 1 is assigned to the new category a. In this way you can store memory m using this mask.
Scan everything with M ₂ and label it.

(c-2) However, in this case, the following problems arise. For example, as shown in FIG. 7A, even if the same section is scanned, different label symbols may be given depending on its shape. For example, when scanning an area α having a shape as shown in FIG. 7A, first the area α ₁ is scanned, and then the entire area α ₂ is scanned at a certain interval. In this case, the area α ₁
Since it cannot be determined at first that and α ₂ are the same, when scanning in the x direction, for example, label signal 11 is given to area α ₁ , and label symbol 12 is given to area α ₂ . . For this reason, after the scanning by mask _M2 is completed,
Scanning is performed again using a 3×3 mask M ₃ as shown in FIG. 7B. In this case, if different label symbols are given to the sections b to i surrounding the section α, these are unified into a small number of label signals, for example. As a result, the part near the dotted line in FIG. 7A is corrected to label symbol 1. This operation is repeated until there are no more changes in the label symbol. As a result, as shown in FIG. 7C, the same area is given the same label symbol.

(c-3) When label symbols in the same area are unified in this way, in most cases the label symbols are not assigned sequentially. For example, the label symbol is 1,
4, 9, and 12 are often given as .
This is corrected sequentially and the label symbols are aligned. Thus, 1, 4 above,
The numbers 9 and 12 will be corrected to 1, 2, 3, and 4, and the label symbols will be aligned.

(d) No. 5 labeled in this way.
The screen shown in the figure is scanned in the x direction as shown by the solid line in Figure 2A. And for each label symbol (Lm), the inflection point R ₁ ("0" → "Lm") is R ₂ ("Lm" →
"0") pairs are detected. Here, "Lm" is the area to which the label symbol Lm is assigned. Then, the locus of the midpoint M of the pair of change points R ₁ and R ₂ is stored. In this case, let the coordinates of R ₁ be (x ₁ , y ₁ ),
If the coordinates of R ₂ are (x ₂ , y ₂ ) and the coordinates of the midpoint M are (xm, ym), then xm = (x ₁ + x ₂ )/2,
ym= _y1 = _y2 . By storing the positions of the midpoints M ₁ (xm ₁ , ym ₁ ) and M ₂ (xm ₂ , ym ₂ ) obtained in this way in a table such as the one shown in FIG. The trajectory of can be memorized. The trajectory of such a midpoint is obtained for each label symbol area. As a result, the midpoint shown in FIG. 9 can be obtained.

(e) For the trajectory of the midpoint M of each label symbol area obtained in this way, draw a straight line (horizontal line) in the x direction between two points that overlap each other, as shown in Figure 10A. Pull. Then, the original image data _m0 set in the memory _m1 is searched on the straight line, and if there is no black point, that is, character information, the first
As shown in Figure 0B, mutual label symbols are given to each of the two points on the table. Once the other party's label symbol is assigned, this process between the two label symbol areas is completed. Then, such processing is performed for all label symbol areas.

(f) Next, find the distance ε before the position where the trajectory of the midpoint M found in (d) above touches the black dot. The magnitude of ε is determined appropriately by simulation. Therefore, as shown in FIG. 11A, the position ε apart from the position where the locus of the midpoint M contacts the black dot can be determined from the table as shown in FIGS. 11A and 11B. FIG. 11B shows, for example, the case of an area labeled with label symbol 1, and FIG.

(g) In this way, draw a horizontal line in the x direction from a position ε in front of you. When this horizontal line touches the black dot, the left end of the horizontal line is the starting point _x3 , and the right end is the end point _xE . Then, enter this into a table as shown in Figure 12A.

(h) For each horizontal line obtained in this way, draw a vertical line where the x-coordinate overlaps with another horizontal line, and search for the trajectory of that vertical line in the original image data _m0 set in memory _m1 . do. If there is no contact with the black dot and there is no mutual labeling between both label symbol areas (detected from the midpoint table), that vertical line will be used as one of the boundaries. In this way, it can be done by drawing lines as shown in FIG.

(i) Character information can be completely separated into individual pieces using the midpoint locus and the boundary line composed of horizontal lines, vertical lines, etc., as shown in FIG. 14 thus obtained. Therefore, individual character information can be obtained by extracting the original image data held in the memory _m1 using the boundary line information.

Next, one embodiment of the present invention will be described based on FIG. 15.

In the figure, 1 is an input section, 2 is a first image memory, 3 is a buffer, 4 is a first feature extraction section, 5 is a first address table, 6 is a first address generation section, and 7 is a second address generation section. , 8 is a buffer, 9 is a second feature extraction section, 10 is a second address table, 11 is an exclusive OR circuit, 12 is a second image memory, 13 is a control section, 14 is a third address generation section, 15 is a buffer, 16 is a first label processing section, 17 is a third image memory, 18 is a fourth address generation section, 19 is a buffer, 20 is a second label processing section, 21 is a fourth image memory, and 22 is a buffer. a label alignment circuit, 23 a fifth image memory, 24 a buffer,
25 is a fifth address generator, 26 is a midpoint extraction circuit, 27 is a first midpoint address table, 28 is a boundary line processing unit, 29 is a buffer, 30 is a second midpoint address table, and 31 is a horizontal line generator. processing section,
32 is a horizontal line table memory, 33 and 34 are buffers, 35 is a boundary line extractor, 36 is a boundary line table, 37 is a buffer, and 38 is an output memory.

The input unit 1 is an electrical signal generation unit that converts a handwritten manuscript or the like using, for example, a photoelectric conversion unit. The first image memory 2 is a memory in which image data input from the input section 1 is held. The buffer 3 is a working buffer memory to which the image data held in the first image memory 4 is sent and set, and is used to perform the processing as described in (a) above.

The first feature extraction unit 4 obtains a pair of change points P ₁ and P ₂ from the image data held in the buffer 3, as shown in (a) above, and sets the region of P ₁ to _{P 2} as "1". The first address table 5 is a table in which the addresses of the areas of "1" between the change points _P1 and _P2 are held.

The first address generating section 6 generates an address for scanning the image data held in the buffer 3 in the x direction as shown in (a) above.

The second address generating section 7 generates an address for scanning the image data sent from the first image memory 2 held in the buffer 8 in the y direction as shown in (b) above.

The second feature extraction unit 9 extracts the change points Q ₁ and Q ₂ from the image data held in the buffer 8 as shown in (b) above.
, and reads out the Q ₁ and Q ₂ areas as "1".
0 is a table in which addresses of the area of "1" between the change points Q ₁ and Q ₂ are held.

The second image memory 12 is a memory in which the image data shown in FIG. 4 obtained as a result of the above (a) and (b) is set.

The control unit 13 processes the image data input to the first image memory 2 according to the steps (a) to (i) above, and performs various controls for creating and extracting individual character areas. For example, it controls the first address generation section 6 for scanning the buffer 3 or the first feature extraction section 4.

The third address generation unit 14 scans the image data held in the buffer 15 using the mask shown in FIG. 6A, as shown in FIG. It is something that generates. The buffer 15 is for holding image data as shown in FIG. 4 transmitted from the second image memory 12.

The first labeling processing unit 16 scans the image data shown in FIG. 4 set in the buffer 15 with the mask shown in FIG. ) to add a label symbol to each area, and as a result, the image data to which the label symbol is assigned is set in the third image memory 17.

As explained in (c-2) above, the address generating section 18 generates an address for causing the buffer 19 to scan the mask _M3 shown in FIG. 7B. The buffer 19 is the third image memory 1
The image data transmitted from 7 is set here.

The second labeling processing section 20
Based on the data obtained as a result of scanning the mask _M3 , the processing described in (c-2) above is performed.

The fourth image memory 21 is a memory in which image data processed by the second labeling processing section 20 and having unified label symbols in the same area is set.

The label alignment circuit 22 sequentially aligns the label symbols as described in (c-3) above, and the resulting image data with label symbols as shown in FIG. 5 is set in the image memory 23. The image data set in the fifth image memory 23 is then sent to the buffer 24. Fifth address generation section 25
was set to buffer 24. This generates an address for scanning the image data shown in FIG. 5 in the x direction.

The midpoint extraction circuit 26 performs the process as described in (d) above to obtain the midpoint address, and the midpoint address obtained is the first midpoint address in a state as shown in FIG. - Set on table 27.

The boundary line communication processing section 28 performs the processing as described in (e) above, and the original image data necessary for this purpose is sent from the first image memory 2 to the buffer 29. The table obtained as a result and labeled with the label symbol shown in FIG. 10B is set in the second midpoint address table 30.

The horizontal line generation processing unit 31 performs the processes (f) and (g) above, and the original image data necessary for this purpose is sent from the first image memory 2 to the buffer 33. The resulting table as shown in Figure 12A is the horizontal line table,
It is set in memory 32.

The buffer 34 is set with the original image data sent from the first image memory 2, and is used to detect whether a black point exists in the overlap area of the horizontal line written in the horizontal line table memory 32, etc. It is used for.

The boundary line extracting unit 35 performs the processing as in (h) and (i) above, draws a vertical line for the boundary line shown in FIG. 13, and thereby creates a boundary line for each character as shown in FIG. 14. Then, the coordinate data of each boundary line is entered in the boundary line table 36.

The buffer 37 is to which original image data is sent from the first image memory 2 and set therein. And this original image data is border table 3
6 is extracted based on the coordinate data indicating the boundary line written in 6, and outputted to the output memory 38 as image data for each character.

Next, the operation shown in FIG. 15 will be briefly explained.

(1) First, image information such as a handwritten manuscript is converted into an electrical signal by the input unit 1, resulting in image data of "1" and "0", and is set in the first image memory 2. This image data is then sent to buffers 3 and 8 and set there as well.

(2) After the image data is set in the buffers 3 and 8 in this way, the control section 13
Address generation section 6 and second address generation section 7
A control signal is sent to the first address generating section 6 to scan the buffer 3 in the x direction, and to the second address generating section 7 to scan the buffer 8 in the y direction. control to generate an address for each. As a result, the first feature extraction unit 4 extracts the change point pair shown in (a) above.
P ₁ and P ₂ are detected and the area of this change point pair P ₁ to _{P 2} is set to ``1'', and this address area set to ``1'' is set in the first address table 5.
On the other hand, the second feature extraction unit 9 extracts the change points shown in (b) above.
The pair of Q ₁ and Q ₂ is detected and the area of this change point pair Q ₁ to _{Q 2} is set to "1", and this address area set to "1" is set in the second address table 10.

(3) In this way, the change points P ₁ ~ P ₂ and Q ₁ ~ Q ₂
After setting the area between
and performs an exclusive OR on the area "1" in the second address table 10, thereby obtaining image data as shown in FIG.
This is set in the second image memory 12.

(4) The image data of FIG. 4 obtained in this way is set from the second image memory 12 to the buffer 15. Then, the control section 13 controls the third address generation section 14 to scan the buffer 15 in the x direction, and based on the data thus read out, the first label processing section 16 performs the above (c-
As shown in 1), mask processing is performed in which data is read out using the mask _M2 shown in FIG. 6A. As a result, the image data to which label storage has been added is set in the third image memory 17.

(5) Next, the image data set in the third image memory 17 is sent to the buffer 19 and set therein. At this time, the control section 13 controls the fourth address generation section 18 to scan the buffer 19 in the x direction, and based on the data thus read out, the second labeling processing section 20 performs the above (c-
2), a masking process is performed in which the data is read using the mask _M3 shown in FIG. 7B. As a result, the same label symbol is given within the same area.

(6) The image data output from the second labeling processing unit 20 is set in the fourth image memory 4 and then processed by the label alignment circuit 22 (c-
After the process 3) is performed and the label symbols are sequentially arranged as shown in FIG. 5, they are set in the fifth image memory 23.

(7) The image data set in the fifth image memory 23 and provided with a label symbol as shown in FIG. 5 is sent to the buffer 24. At this time, the control section 13 sends a control signal to the fifth address generation section 25 to scan it in the x direction. Then, based on the data outputted as a result, the midpoint extraction circuit 26 performs the process as described in d above, and the trajectory of the midpoint as shown in FIG.
The ninth address entered in the midpoint address table 27
The trajectory of the midpoint shown in the figure is obtained.

(8) The trajectory data of this midpoint is the boundary line communication processing unit 2.
8. Further, the original image data sent from the first image memory 2 is set in the buffer 29, and the boundary line communication processing section 28 performs the process (e) above while referring to this original image data. As a result, a table with label symbols as shown in FIG. 10B is entered in the second midpoint address table 30.

(9) The data set in the second midpoint address table 30 is transmitted to the horizontal line generation processing section 31, where the above processes (f) and (g) are performed. At this time, the necessary original image data is set in the buffer 33, and the horizontal line generation processing section 31 performs the processes (f) and (g) above while referring to the original image data set in the buffer 33. The data obtained as a result is then set in the horizontal line table memory 32 in the state shown in FIG. 12A.

(10) Next, the boundary line extraction unit 35 performs the processing as in (h) and (i) above based on the data transmitted from the horizontal line table memory 32 and the original image data set in the buffer 34. . In this way, vertical lines for boundaries as shown in FIG. 13 are drawn, and based on these, boundaries for each character as shown in FIG. 14 can be created. The coordinate data of each boundary line shown in thick lines in FIG. 14 is then set in the boundary line table 36. Since the original image data is sent from the first image memory 2 to the buffer 37, the original image data can be extracted for each character based on this boundary line coordinate data and set in the output memory 38. can. In this way, the required character information can be obtained individually for each character from the output memory 38, so that handwritten characters can be recognized by inputting this information into a character reading device or the like.

As explained above, according to the present invention, even when a plurality of characters are handwritten randomly and close to each other, regardless of the writing position in the character frame etc., a border line can be created in the middle between the characters. and can accurately separate each letter. Therefore, since character information can be extracted individually, it is possible to provide a character separation means that is very effective in recognizing handwritten characters.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of a mask for character reading and cases in which character separation is possible and impossible using the mask, FIG. 2 is an explanatory diagram of the scanning state, and FIGS. 3 to 14 are diagrams illustrating the operation of the present invention. A state explanatory diagram, FIG. 15, is a configuration diagram of an embodiment of the present invention. In the figure, 1 is an input section, 2 is a first image memory, 3 is a buffer, 4 is a first feature extraction section, 5 is a first address table, 6 is a first address generation section, and 7 is a second address generation section. , 8 is a buffer, 9 is a second feature extraction section, 10 is a second address table, 11 is an exclusive OR circuit, 12 is a second image memory, 13 is a control section, 14 is a third address generation section, 15 is a buffer, 16 is a first label processing section, 17 is a third image memory, 18 is a fourth address generation section, 19 is a buffer, 20 is a second label processing section, 21 is a fourth image memory, and 22 is a buffer. a label alignment circuit, 23 a fifth image memory, 24 a buffer,
25 is a fifth address generator, 26 is a midpoint extraction circuit, 27 is a first midpoint address table, 28 is a boundary line processing unit, 29 is a buffer, 30 is a second midpoint address table, and 31 is a horizontal line generator. processing section,
32 is a horizontal line table memory, 33 and 34 are buffers, 35 is a boundary line extractor, 36 is a boundary line table, 37 is a buffer, and 38 is an output memory.

Claims (1)

[Claims] 1. An information holding means for holding image information of characters inputted from an input means, and a first area that scans the character information in a first direction to obtain points of black and white change, and is an area between the points of change. a first feature extraction means for detecting a black and white change point by scanning the image information in a second direction substantially perpendicular to the first direction; a second feature extraction means for detecting a region; a scanning means for sequentially scanning in the first direction an intermediate region where the first region and the second region do not overlap to obtain scanning line segments; and a scanning means for obtaining scanning line segments; intermediate point setting means for determining the intermediate point in each scanning line segment; horizontal line generating means for generating a horizontal line in a first direction from the intermediate point in proximity to the image information of the character; A character separation method characterized in that character areas in which individual character information exists can be distinguished by the horizontal lines.