JPH0147834B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention is characterized by normalizing the sum of black bits for each character frame division unit area in each subpattern representing a stroke in a specific direction of an input character pattern by the character line width and the size of the character frame corresponding to each subpattern. Extract the elements, perform a blurring process by weighted averaging each feature element of the input character pattern with its adjacent elements according to the complexity of the input character pattern, and refer to a dictionary corresponding to the blurring process. Concerning character recognition methods for recognizing characters. As a form of character recognition method, we create a feature matrix whose feature elements F (k, i, j) are the quantities expressed by the following formula, and measure the similarity with a standard character mask expressed in the same format. There is a character recognition method that recognizes characters. F(k, i, j)=Bk(i, j)/WL·WPk WL represents the character line width of the input character pattern. WPk represents WPh, WPv, WPr, WPl, etc. For example, WPh is the size of the character frame in the horizontal direction, and similarly, WPv, WPr, and WPl are the vertical direction, 45 degrees diagonally to the right, and 45 degrees diagonally to the left, respectively. The size of the character frame in the direction. Bk(i, j) is Bh(i,
j), Bv (i, j), Br (i, j), Bl (i, j), etc., for example, Bh (i, j) is the black color in the divided unit area (i, j) of the horizontal sub-pattern HSP. It is the sum of bits, where i is the number of the divided unit area of the character frame in the X-axis direction, and j is the number of the divided unit area of the character frame in the Y-axis direction. Similarly, Bv(i,
j), Br (i, j), and Bl (i, j) are vertical subpattern VSP and right diagonal 45 degree subpattern, respectively.
RSP is the sum of black bits in the divided unit area (i, j) of the 45-degree left diagonal sub-pattern LSP. The general tendency of this method is that it is easily affected by variations in the stroke position of characters due to individual differences, and the recognition rate improves by absorbing variations in the stroke position by weighting and averaging each feature element with its adjacent elements. do. On the other hand, characters with more complex shapes are less affected by stroke position fluctuations, and simpler characters are more susceptible to stroke position fluctuations. For example, if a feature matrix is created by dividing a character frame into seven parts in each direction of the A high recognition rate can be obtained by averaging around 12 strokes with a weight of around 1/6, and averaging around 16 strokes with a weight of around 1/7. Based on this point of view, the present invention calculates the sum of black bits of sub-patterns of an input character pattern normalized by the character line width and the size of the character frame in the stroke direction, or the sum of the number of strokes of each sub-pattern. is detected as complexity, determines a blurring coefficient according to this complexity, performs blurring processing on the feature elements of the feature matrix of the input character pattern, and performs character recognition by referring to a dictionary according to the blurring coefficient. Its purpose is to perform stable and accurate character recognition. FIG. 1 shows an embodiment of a character recognition device according to the present invention. A detailed explanation will be given below based on FIG. 1. 1 is optical input from the form. This optical input 1 is input to a photoelectric conversion section 2. The photoelectric conversion unit 2 decomposes one character area into 128 x 128 pixels and converts each pixel into a binary digital signal (hereinafter referred to as input character pattern). One character is represented by an input character pattern of about 60 x 60 bits. The pattern register 3 stores the input character pattern in a format capable of reproducing the X and Y coordinates of each pixel in the expected character area, and has a capacity of 128×128 bits corresponding to the expected character area. The character line width calculation unit 4 has a shift register configuration similar to a well-known filter circuit, and receives an input character pattern and calculates the number Q of states in which all pixels in a 2×2 shift register window are black bits. The black bit sum A of the input character pattern is counted, and the line width WL is calculated using the well-known formula shown below. WL=A/A-Q The character frame detection unit 5 detects the circumscribed frame of the character by expressing it in the left end coordinate Xl, right end coordinate Xr, upper end coordinate Yt, and lower end coordinate Yb in the pattern register, and further determines the size of the character frame. This is to detect. The size of the character frame in the horizontal direction is WPh=Xr−Xl
+1, as vertical size WPv=Yt−Yb+
1, is detected. The size in the 45 degree right diagonal direction and the 45 degree diagonal left direction is detected as WPr=WPl=WPh+WPv/2. The sub-pattern extraction unit 6 extracts horizontal, vertical,
This method extracts subpatterns HSP, VSP, RSP, and LSP that represent strokes in each direction (right diagonal, left diagonal), and is performed by extracting continuous black bits that are sufficiently longer than the line width corresponding to each direction. For example, the sub-pattern HSP is created by horizontally scanning the entire surface of the pattern register 3, detecting the number of consecutive black bits for each scanning line, and extracting the black bits whose length is greater than 2WL. Extract subpatterns. Similarly, vertical, diagonal right, and diagonal left subpatterns are extracted by scanning the pattern register 3 in the vertical direction, diagonal right direction, and diagonal left direction, respectively. An example of subpatterns extracted in this way is shown in Figure 2, where ORG in Figure 2 is the input character pattern, HSP is the horizontal subpattern, VSP is the vertical subpattern, RSP is the right diagonal subpattern,
LSP is a left diagonal subpattern. The character frame division determining unit 7 determines the number of divisions NX, NY in each direction of the X-axis and Y-axis determined by the division format,
Based on the character frame coordinates Xl, Xr, Yt, Yb and the input character pattern ORG, a preset maximum By finding each series of barycenter coordinates in a number sufficiently larger than the number of divisions, and selecting the barycenter coordinates at approximately equal intervals corresponding to the number of divisions NX and NY, respectively (NX-1),
(NY-1) division coordinates are determined. For example, to detect the X-axis barycenter coordinate series X (Mi) (where Mi is the barycenter number and i = 1 to 15), the input character pattern is projected onto the X-axis to obtain the black bit number distribution, and the Find the barycentric coordinates X (Mp) for the range Xl to Xr of the character frame, and then divide the range Xl to Xr of the character frame on the By repeating the process of finding the center of gravity coordinates three times, the other 14 center of gravity coordinates X (M1) ~
(M7), X (M9) to X (M15) are detected. The 15 barycenter coordinates obtained in this way are determined as candidate points for dividing coordinates based on the table below prepared in advance so that each divided section of the X-axis includes an approximately equal number of barycenters. .

[Table] Similarly for the Y axis, 15 barycenter coordinates Y
(M1) to Y(M15) are determined, and the division coordinates are determined by referring to the table according to the number of divisions. Note that the division format can be changed and set once it has been rejected, or it can be set according to the complexity of the input character pattern, but here, for the sake of simplicity, NX=NY It is assumed that the division format is set to =7. Therefore, the character frame division determining unit 7 determines that NX=NY=
Corresponding to the 7 division format, the division coordinates in the X-axis direction are X (M3), X (M5), X (M7), X (M9),
(M11),
(M11) and Y (M13) are determined. The feature matrix extraction unit 8 extracts division coordinates in the X-axis direction, both end coordinates Xl, X (M3), and
(M5), X (M7), X (M9), X (M11), X (M13),
Xr, division coordinates in the Y-axis direction, both end coordinates Yb, Y
(M3), Y (M5), Y (M7), Y (M9), Y (M11),
Y (M13), Yx and each subpattern HSP, VSP,
In response to RSP, LSP, character frame sizes WPh, WPv, WPr, WPl corresponding to each sub-pattern, and character line width WL, each sub-pattern is divided into multiple division unit areas determined from the division coordinates. Then, count the number of black bits Bk (i, j) for each divided unit area, and calculate the line width WL and the character frame size WPh, WPv, corresponding to each subpattern as shown in the following formula.
Normalized by WPr and WPl, the feature element F (k, i, j) =
Bk (i, j)/WL・WPk (k=h, v, r, l)
Extract and create a feature matrix. F (h, i, j) = Bh (i, j) / WL・WPh F (v, i, j) = Bv (i, j) / WL・WPv F (r, i, j) = Br (i , j)/WL・WPr F(l, i, j)=Bl(i, j)/WL・WPr Black bit sum Bk(i, j) for each divided unit area (i, j) of each subpattern The calculation is performed by reading each sub-pattern at the division coordinates corresponding to the division unit area (i, j) and adding up the number of black bits. For example, the black bit sum Bh (1, 1) corresponding to the feature element F (h, 1, 1) is the X coordinate Xl, X (M3) and Y
It is obtained by reading out the horizontal sub-pattern HSP in the range determined by the coordinates Yb, Y (M3) from the sub-pattern extraction section 6 and counting the number of black bits. The complexity calculation unit 9 calculates four types of sub-patterns,
HSP, VSP, RSP, LSP and character frame size
Receiving WPh, WPv, WPr, WPl and character line width WL, count the black bits in the character frame of each subpattern to find the black bit sum Ah, Av, Ar, Al, and use the following formula to calculate the character line width. The complexity G of the input character pattern can be calculated by normalizing it with the size of the character frame corresponding to each stroke direction, calculating the quantities Gh, Gv, Gr, and Gl corresponding to the character line density, and calculating the sum of these. calculate. G=Gh+Gv+Gr+Gl However, Gh=Ah/WL・WPh Gr=Ar/WL・WPr Gv=Av/WL・WPv Gl=Al/WL・WPl The blurring coefficient determination unit 10 determines the complexity level G of the input character pattern and the input character pattern. A table showing the correspondence with a blurring coefficient C used when performing blurring processing is provided, and in response to the complexity G of the input character pattern, the blurring coefficient C is determined by referring to the blurring coefficient table. In this embodiment, the contents of this table are as follows.

[Table] The blurring processing unit 11 receives the feature matrix of the input character pattern and the blurring coefficient, and sets the weight of the featured feature element to C/(C+2), and sets the weight of the adjacent feature element to 1/(C+2), The feature element of interest is blurred by weighted averaging of each feature element with two adjacent elements in the direction orthogonal to its stroke direction, and the
×7 × 7-dimensional blur feature element FB (k, i, j)
(However, k=h, v, r, l, i, j=1 to 7)
Create a blur feature matrix consisting of . That is, a blurring feature matrix is created by executing the following formula operation with different designations of adjacent elements for each subpattern. For horizontal sub-patterns, FB(h,i,j)=CF(h,i,j)+F(h,i
, j-1)+F(h,i,j+1)/C+2 For vertical sub-patterns, FB(v,i,j)=CF(v,i,j)+F(v,i
-1, j) + F (v, i + 1, j) / C + 2 For the right diagonal sub-pattern, FB (r, i, j) = C・F (r, i, j) + F (r, i
-1,j+1)+F(r,i+1,j-1)/C+2 For the left diagonal subpattern, FB(l,i,j)=C・F(l,i,j)+F(l,i
-1, j-1)+F(l, i+1, j+1)/C+2. Note that the feature elements with i-1=0 and j-1=0 are replaced with feature elements with i=1 and j=1, respectively, and the feature elements with i+1=8 and j+1=8 are replaced with i, respectively. =7 and j=7 to find the blurring feature element. The identification unit 12 includes a standard character mask FS (k, i,
j) is the same as the blur feature matrix in the input character pattern, the black bit sum of each divided unit area (i, j) for each subpattern of the standard character is calculated using the character line width and the size of the character frame corresponding to the subpattern. The feature matrix of the input character pattern is expressed by a blur feature that is normalized and further weighted and averaged according to a blur coefficient according to the complexity level, and a plurality of dictionaries are prepared corresponding to the plurality of blur coefficients. The degree of similarity between the standard character mask and the standard character mask is measured, and the most similar character code of the standard character mask is recognized as the input character pattern name. here,
A dictionary with a blurring coefficient C=3 stores standard character masks with a complexity G of G≦6, which corresponds to characters whose number of strokes is about 1 to 12. Blur coefficient C=
Dictionaries 4 and 5 contain standard masks with a complexity of 6<G≦8 and G>8, respectively, and correspond to characters with a number of strokes of about 4 to 15, and characters with a number of strokes of about 6 or more, respectively. . Note that even if standard character masks have the same category, those with large deformations are stored redundantly in two dictionaries as subcategories, and standard character masks whose complexity G is in the boundary area where G = 6 and 8 are also stored in two dictionaries. The identification unit 12 receives the blurring feature matrix FB (k, i, j) of the input character pattern ORG and the blurring coefficient C, selects a dictionary corresponding to the blurring coefficient C, for example, a dictionary corresponding to C=3, and selects a dictionary corresponding to the blurring coefficient C, for example, a dictionary corresponding to C=3, and identifies the input character pattern. Distance D between and standard character mask, D=√{(,,)−(,,)}
² is calculated, and the character code of the standard mask having the minimum value is output to the character name output terminal 13. As explained above, in this example,
3, 4, 5 depending on the complexity of the input character pattern
Since the blurring coefficient C is determined using the following three formats,
It has the advantage of being able to absorb stroke variations in handwritten characters by the writer and extract features that follow the complexity. Note that complexity itself is a characteristic element that can distinguish characters, and when preparing as many dictionaries as the above example, it is necessary to arrange and divide the storage array of standard masks for each dictionary in order of complexity. or by preparing a table that shows the correspondence between complexity and standard character mask categories, by specifying one dictionary and the classification or category name within it according to the complexity, one It is also possible to recognize characters simply by checking them against a standard character mask in a dictionary. Furthermore, the complexity of the input character pattern can be determined not only by the sum of the amounts corresponding to the character line density, but also by detecting the number of strokes in each subpattern HSP, VSP, LSP, and RSP, and calculating the sum of these. It can also be detected. Furthermore, although the present invention has been described as assuming that the predetermined division format NX×NY used in the character frame division determination section 7 is fixed, a plurality of division formats may be provided so that they can be used depending on the type of input character, the number of strokes, etc. You can also do that. In this case, dictionaries are also prepared to correspond to a plurality of division formats. As is clear from the above description, in the present invention, since the degree of blurring processing is changed according to the degree of complexity, there is an advantage that stability regarding individual differences can be maintained and a high recognition rate can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention;
FIG. 2 is a diagram showing sub-patterns. 1... Optical input, 2... Photoelectric conversion section, 3... Pattern register, 4... Character line width calculation section, 5... Character frame detection section, 6... Sub pattern extraction section, 7...
Character frame division determination unit, 8...Feature matrix extraction unit, 9...Complexity calculation unit, 10...Blur coefficient determination unit, 11...Blur processing unit, 12...Identification unit,
13...Character name output terminal.

Claims (1)

[Claims] 1. Obtained by normalizing the number of black bits of an arbitrary divided unit area in a subpattern representing a stroke in a specific direction of an input character pattern by the character line width and the size of a character frame corresponding to the stroke direction. A feature matrix is created by extracting the feature elements from the input character pattern for each of the plurality of sub-patterns with different stroke directions and for each of the divided unit areas obtained by dividing the character frame, using the quantity as the feature element. A blurring feature matrix is created by weighted averaging of each feature element with adjacent feature elements adjacent in the direction orthogonal to the stroke direction of each feature element according to the blurring coefficient, and the standard character mask is In a character recognition method that recognizes an input character pattern by referring to a dictionary described in the same format as the blur feature matrix, a table showing the correlation between character complexity and blur coefficient is prepared, and the blur coefficient Prepare a plurality of dictionaries with different values, and calculate the sum of the black bits of each subpattern normalized by the character line width and the size of the character frame corresponding to the stroke direction, or the summation of the number of strokes of each subpattern. is detected as the complexity, the blur feature matrix is created according to the blur coefficient corresponding to the complexity with reference to the table, and all standard character masks or A character recognition method that recognizes an input character pattern by specifying a group of standard character masks and measuring the degree of similarity between these standard character masks and a blur feature matrix of the input character pattern.