JPH0576664B2 - - Google Patents

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention relates to printed character recognition, particularly kanji, hiragana,
This paper relates to a classification method for recognizing printed Japanese characters such as katakana, alphabets, and numbers.

(Prior art and its problems) Conventional printed kanji recognition has problems such as a large number of character types, complexity of character shapes, and large number of similar character pairs when using a single font, and even more when using a multi-font. Changes have been considered a problem. In general, the target character types for kanji recognition are at least about 2,000 commonly used kanji to about 3,000 JIS first-level characters, and in some cases there is a need to read more character types. Therefore, as a countermeasure against the large number of character types, a broad classification that narrows down the candidates to about 100 character types through relatively simple processing,
It is used in most cases. A recognition method that takes into account the complexity of the character shape is applied to the candidate character types selected in this way, and if the recognition output cannot be narrowed down to one character type and similar character pairs remain, detailed recognition processing is performed. can be done. In an example of such processing, a combination of complexity index and four-sided code method as a major classification, a composite similarity method as a recognition process, and a mixed similarity method as a detailed discrimination process are applied to the recognition of approximately 2000 single-font printed Kanji characters. It is known that (Nikkei Electronics 1977.10.31 issue
pp.102−128 Kenichi Mori, Kunio Sakai, “2000 character types to 100
(Development of OCR for printed kanji that can be read in characters per second) However, in this type of broad classification processing, there is a problem in classification performance when the recognition target is multi-font, and in order to send the correct candidate character type to the subsequent recognition processing, Increasing the number of character types increases the burden of recognition processing, creating a problem in that it becomes difficult to maintain recognition performance and processing speed.

In addition to this example, we targeted approximately 2,000 types of multi-font printed kanji, such as commonly used kanji, and made the classification process more efficient, processing only candidates with similar shapes using detailed identification, and processing most of the characters. There have also been reports of cases in which recognition processing occurs during the recognition process. (IEICE Journal Vol. J67-D, No. 8, Shinichi Meguro,
(Michio Umeda, “Multi-font printed kanji recognition device”) In this classification process, two types of feature quantities are used as shown below. One is to define the horizontal sides of the circumscribed rectangle of the character pattern, which is expressed as an m×n two-dimensional array in which the character part is binarized with black dots and the background part is white dot, by M.
Divide the vertical direction into N parts, and divide this circumscribed rectangle into M
It divides into ×N rectangular partial regions, counts the number of sunspots in each partial region, creates M×N-dimensional vectors, and performs classification in vector space.Experimentally,
The best classification performance is obtained when dividing into 8×8 partial regions and performing classification using 64-dimensional vectors. Another feature quantity is to divide the horizontal side of the circumscribed rectangle of the same binarized character pattern into M divisions, and the vertical side into N divisions, and scan horizontally to the right from the left vertical side. , the number of white points until reaching the black point for the first time and the number of white points until reaching the black point for the second time are counted. By summing these two types of white point numbers for each of the N sections, 2×N elements are obtained. The same process is performed by scanning horizontally to the left from the right vertical side, vertically downward from the top horizontal side, and vertically upward from the bottom horizontal side. 2×N, 2×M, 2
xM elements are obtained. Therefore the total (M+N)
×2×2 elements are obtained, so (M+N)×2
x Create a two-dimensional vector and perform classification in vector space. In practice, the best performance is achieved by setting M=N=8 and processing using a 64-dimensional vector.

The former feature vector indicates the global distribution of the character parts of the character pattern, and the latter feature vector reflects the shape of the periphery of the character pattern, and both reflect the structure of the character pattern in the absolute position in a global manner. It has become a different characteristic. However, since all of these characteristics capture global characteristics, we expanded the recognition target to approximately 3,000 characters of the JIS first level, and the number of characters with complex structures and shapes became extremely large. As the number of combinations of character pairs increases, there is a problem in that the subsequent identification processing becomes burdensome, making it difficult to maintain classification performance.

On the other hand, there is also a classification method using feature vectors that reflect information on the relative positions of character patterns.

Conventionally, the following method is known as a typical method of this type. That is, first, create a rectangular area whose four sides surround the character to be recognized, divide this area into m pieces vertically and n pieces horizontally, and divide the area into m×n pieces.
A set of microregions. Each minute area is defined as a point within the area containing the character. Such points are divided into two types: background points that do not include lines representing characters, and character points that include lines that represent characters. Now, a certain background point P defined in this way
From that point, scan lines are drawn horizontally to the left and right, and vertically to the top and bottom, and these scan lines are connected to the character line c times to the left and d times to the right, respectively.
Suppose that it has crossed upwards a times and downwards b times.
Considering this as a quantity representing one characteristic associated with the background point P, all background points are associated with a characteristic having certain respective values of c, d, a, and b.

Now, if the values of c, d, a, and b, which represent the number of times the character lines intersect, are 2 or more than 2, then if we limit it to 2, the allowed values are 0, 1, and 2. There are three types. Therefore, the number of types allowed as characteristics of each point is 3 ⁴ = 81,
We aggregate background points for each of the 81 types of characteristics to generate an 81-dimensional vector, and use this vector to perform character recognition using a linear discriminant function.

This is Glucksman's method [H.A. Year reference (HA
Glucksman: Classification of Mixed font
Alphabetics by Characteristic Loci, Digest
of First IEEE computer Conference, P.138,
1967)], but when trying to apply this to the recognition of Japanese characters including Kanji, the following problems arise.

First, the number of times a horizontal or vertical scanning line intersects a character is limited to two times, which is too small for characters with a large number of strokes, such as Kanji characters.

Second, simply counting the number of intersections with the character portion will make the values of the vector elements unstable with respect to the inclination of the character.

Third, if the upper limit of the number of intersections is set to about 7 to match the plurality of kanji, then creating a vector like the one above will process a vector in a space with a huge number of dimensions, 8 ⁴ = 4096 dimensions. The need arises and the process becomes very difficult.

In order to eliminate the above-mentioned drawbacks, a character recognition method that increases the upper limit of the number of intersections, stably determines the count value of the number of intersections with respect to the inclination of the character, and further reduces the amount of processing to a feasible amount is proposed. , has already been published in Japanese Patent Publication No. 1983-
This method is disclosed in Japanese Patent Application No. 70383 (Patent Application No. 169153/1983 Character Recognition System). This recognition method is
Although effective classification is performed using only relative position information, if the application is limited to the classification of printed characters, even more effective classification can be expected by introducing absolute position information specific to printed characters.

One such attempt was already made in 1983.
There is a publication No. 16503 (character identification processing method). Here, first, as in Glucksman's method, the number of intersections a, when scanning from each background point in the vertical and horizontal directions,
Although b, c, and d are determined, the number of intersections is not particularly limited to two times. After the number of intersections is assigned, the m×n character pattern is
Divide, divide into N in the vertical direction to define M×N local areas. This local area consists of p pieces in the horizontal direction and q pieces in the vertical direction, and the relationship is m=M×p, n=N×q. In each local area, calculate the total number of intersections in each of the vertical and horizontal directions assigned to each point in the local area, and calculate the total number of intersections in the upward direction as A, the total number of intersections in the downward direction as B, When the total number of intersections in the left direction is C and the total number of intersections in the right direction is D, the four-dimensional vector represented by (A, B, C, D) is the feature vector representing the local area. . For each local area, a representative four-dimensional feature vector is obtained, and classification is performed by vector calculation using all feature vectors or selected feature vectors. Although this method does reflect position information, by calculating the total number of intersections in each direction, up, down, left, and right within each local area and making it into a four-dimensional vector, when the local area is relatively small, the Relative position information, which is a feature of Glucksmann's method or the method of JP-A-58-70383, is almost preserved in the four-dimensional vector, but as the local area becomes larger, the relative position information deteriorates. There is a problem with doing so. Therefore, if priority is given to classification ability, the local area will be made smaller. In other words, the values of p and q are reduced. However, from the relationship of m, n, M, N, p, q, the p,
When the value of q is reduced, the values of M and N become larger, and the number of local regions M×N becomes larger. Each of the local regions is represented by a four-dimensional vector, so if all the local regions are used to represent a character pattern, it becomes a 4×M×N-dimensional vector, and the number M×N of the local regions increases. This means that the dimension of the vector representing the character pattern increases, which leads to an increase in the processing required for vector calculations for classification. For example, the character pattern is 64
When the pattern is ×64, that is, m=n=64, and the size of the local area is 4×4, that is, p=q=4, the m, n, M, N, p,
From the relationship of q, M=N=16, and the number of local regions is
256 pieces, that is, M×N=256. Therefore, if the character pattern is represented using all the local regions, it will become a vector with an enormous number of dimensions of 4×M×N=1024 dimensions. In addition, even if we choose local regions to reduce the number of dimensions, if 1/8th of 32
Even if this number is selected, the number of dimensions of the vector becomes 128 dimensions (=4 dimensions x 32), but the problem remains that too much information can be lost as a character pattern. Also, by increasing the size of the local region and assuming that p
= q = 8, M = N = 8, and the number of dimensions of the vector representing the character pattern is 256 dimensions (= 4
dimension × 8 × 8), but as already mentioned, 8
A problem arises in the classification ability because the ×8 area is represented by only a four-dimensional vector.

(Objective of the Invention) The object of the present invention is to maintain sufficient classification performance for printed characters even when the number of characters with complex structures increases and the number of combinations of similar characters increases with the increase in character types. In order to do this, we have developed a character classification method that can prevent an increase in the amount of processing by realizing stable features that reflect relative position information and absolute position information using a 64-dimensional vector, which is a very low-dimensional vector. It provides:

(Structure of the Invention) According to the present invention, a character pattern formed by converting a character portion and a background portion into a quantized signal consisting of binary values is scanned in multiple directions from each point of the background portion and intersected. In a character classification method that uses features obtained by counting the number of strokes and performing characterization, the circumscribed rectangular area of the character pattern is divided into two halves in the horizontal and vertical directions, and four types of characters are classified. A small area is defined, and in the upper left small area, scanning is performed in two directions upward and to the left from each point in the background area, and in the upper right small area, scanning is performed in the upward direction from each point in the background area. In the lower left small area, scanning is performed in two directions, downward and to the left from each point in the background part, and in the lower right small area, scanning is performed in two directions from each point in the background part to the downward direction. Scanning is performed in two directions, ie, the horizontal direction and the right direction, and the number of times the horizontal and diagonal strokes intersect with each other is counted in the upward and downward scans, and the number of times the strokes intersect with the horizontal and diagonal strokes is counted in the left and right scans. Count the number of times the stroke intersects with the stroke of A feature pattern generation unit that assigns 16 types of feature codes to each point in the background part of the region by combining the number of intersections in two directions, and the appearance frequency of each of the 16 types of feature codes in the four small regions mentioned above. a feature vector extraction unit that takes as an input feature vector a 64-dimensional vector obtained by calculating 16 elements, and normalizing the 64 elements found in the entire region by the area of the circumscribed rectangle; Compare the input feature vector and the standard feature vector for each character type with a standard vector storage unit that stores standard vectors described in 64-dimensional vectors in the same format as the feature vectors and prepared for each character type. and a dissimilarity calculation unit that calculates the degree of dissimilarity between the input feature vector and the standard vector for each character type by vector calculation, and a dissimilarity calculation unit that ranks each character type based on the dissimilarity and performs detailed discrimination. a classification processing unit that determines candidate character types for character recognition, and performs broad classification for character recognition based on feature amounts that reflect absolute position information and character pattern structure information. can be realized.

(Principle of operation of the invention) The principle of the invention will be explained in detail below using the drawings. FIG. 6 is a diagram for explaining the principle of the present invention. In Figure 6a, A, B, C, D
are the points that bisect each side of the circumscribed rectangle of the binarized character pattern, and E, F, G, and H are the points A, B, and H that bisect the circumscribed rectangle, respectively.
These are the background points of characters belonging to the upper left partial area, upper right partial area, lower left partial area, and lower right partial area when divided into four by C and D. From this background point,
Unlike the prior art, which scans in four directions: up, down, left, and right, scanning is performed in only two directions specified by the partial area to which the point belongs. That is, from all points in the background part of the upper left partial area, scanning is performed upward and to the left, from all points in the background part of the upper right partial area, scanning is performed upward and rightward, and from all points in the background part of the upper left partial area, scanning is performed upward and to the left. Scanning is performed downward and to the left from all points in the background area, and scanning is performed downward and to the right from all points in the background area in the lower right partial area, and the outline of the character pattern that intersects with the scanning direction is scanned. If the tangential direction intersects at an angle greater than or equal to a predetermined threshold, each background point is characterized by counting the number of times of intersection. FIG. 6b shows the background point E,
The feature expressions of F, G, and H are shown, where the background point E intersects upward twice and to the left once, the background point F intersects upward once and to the right zero times, and the background point G intersects upward once and to the right zero times. is below 2
times, cross once to the left, and the background point H crosses downward once,
It crosses to the right twice. FIGS. 6c and 6d are diagrams showing the effect of counting the number of times a scanning line intersects a character stroke, taking into account the direction of the tangent to the contour of the character pattern. Figure 6c
shows that the same character pattern as in the 6th a was inputted at an angle, and A', B', C', and D' are the points that bisect each side of the circumscribed rectangle, and E' and F. ′、G′、
H′ is a point in the background area. The background point
When scanning leftward from E', it intersects the character part once, but it is not counted because the angle between the scanning direction and the tangential direction of the outline part is small. Similarly, when scanning upward from the background point E', it intersects the character part twice, but 1
Since the angle between the scanning direction and the tangential direction of the contour is small, the number of times of intersection is set to 1 when characterizing. The same applies to the other background points F', G', and H', and the results of characterizing each of the background points E', F', G', and H' are shown in FIG. 6d. As shown in this example, by considering the angle between the tangential direction of the contour of the character pattern and the scanning direction when counting the number of intersections, it is possible to obtain the effect of absorbing the inclination of the character pattern. In printed character recognition, character patterns are frequently input at an angle, so whether or not the inclination of the character pattern can be absorbed has a large effect on recognition ability.

An upper limit of 3 is set for the number of intersections. That is, even if the scanning crosses four or more times, the number of crosses is set to three. This means that it is sufficient to allocate 2 bits when encoding the number of crossings. The reason for this is that when a character pattern is divided into the four partial areas, it is extremely rare for more than four intersections to occur. This is because the feature for classification does not deteriorate its performance. In this way, in each partial area, each background point is characterized by the number of intersections in two directions, vertical and horizontal, and since two bits are required for the feature in one direction, two
The directional feature will require 4 bits, and each background point characterized in two directions will be allocated 4 bits. Therefore, in each of the partial regions,
One of the 16 types of features is assigned to each of the background points, and if each of the 16 types of features is characterized by the four partial areas to which it belongs, 64 types of features are assigned to the background points of the entire area of the character pattern. will be assigned. Here, if we limit the number of intersections to 2 or less, there are 3 types of intersections in one direction: 0, 1, and 2, so there are 9 (= 3 × 3) types of features that can be assigned to background points characterized by two directions. Then,
Any one of 36 (=9×4) types of features is assigned to the background points in the entire area. This is less than half the number of features compared to the case where the number of intersections is set to 3 at most, but when the classification target is a Chinese character with a complex shape, the features are not very coarse. Therefore, when setting four partial regions as in the present method, the best effect can be obtained by setting the number of intersections to a maximum of three in consideration of classification ability and processing amount.

Now, the above processing will be formulated as follows. Let the code indicating the partial area be i, i = 0 when indicating the partial area on the upper left, i = 1 when indicating the partial area on the upper right, and i when indicating the partial area on the lower left.
=2, and when indicating the partial area at the lower right, i=3. Next, let j and k be codes indicating the number of intersections in two directions in each of the partial regions. For example, i=0
When , j is the number of crosses in the upward direction and k is the number of crosses in the left direction.When i=1, j is the number of crosses in the upward direction and k is the number of crosses in the right direction.When i=2, j is the number of crosses in the downward direction. The number of crossings in the direction, k is the number of crossings in the left direction, when i=3, j is the number of crossings in the downward direction, and k is the number of crossings in the right direction. If the feature assigned to each background point using the codes i, j, and k is expressed as fijk, the feature assigned to the background point E in FIG. 6a is f ₀₂₁ and the feature assigned to the background point F is f 021 . is f ₁₁₀ and the feature assigned to the background point G is
f ₂₂₁ , and the feature assigned to the background point H is f ₃₁₂ . In the explanation so far, the codes i, j, and k take on four types of values, 0, 1, 2, and 3, so the characteristics
There are 64 types of f _ijk .

When each background point of one character pattern is characterized, the number of background points to which the feature f _ijk is assigned is F _ijk , and the value of the feature point F _ijk normalized by the area of the circumscribed rectangle. Let be F^ _ijk . A 64-dimensional vector having the normalized value F^ _ijk as an element is a feature vector of the character pattern. The classification process is performed as follows. The 64-dimensional feature vector is prepared for each character type, and this is defined as the standard vector 〓(S) _l .
S is a code representing the character type. If the 64-dimensional feature vector obtained from the input character pattern is the input vector 〓, then the degree of dissimilarity defined by the vector operation between the input vector 〓 and each of the standard vectors 〓(S) _l is expressed as D(〓, 〓 (S) _l ), and the ranking is performed in descending order of the value of the degree of difference, and classification processing is performed. In this case, the degree of difference may be determined by a vector distance such as a Euclidean distance or a City Block distance, and processing that is performed in normal character recognition may be performed.

(Embodiment) FIG. 1 is a block diagram showing the configuration of the present invention.

Reference numeral 1 denotes a feature pattern generation unit, which inputs a binarized character pattern as a signal 90, performs the characterization on each of the background points of the four types of partial areas, and generates a signal containing circumscribed rectangle information of the character pattern. 91, outputting the characterized feature pattern for each partial region as a signal 92; 2 is a feature vector extraction unit which inputs the feature pattern signal 92 and counts the number of background points of the character pattern assigned to each feature f _ijk to obtain the frequency of appearance F^ _ijk ; At the same time, input the circumscribed rectangle information signal 91 to calculate the area of the circumscribed rectangle, normalize the frequency of appearance F _ijk by the area of the circumscribed rectangle, and obtain the normalized value F^ _ijk ;
A signal 93 is output as the 64-dimensional input feature vector. 3 is a dissimilarity calculation unit which inputs the input feature vector as a signal 93 and the standard feature vector as a signal 95, and calculates the input feature vector F and the standard feature vector 〓(S) _l by a predetermined vector calculation. The degree of dissimilarity D (〓,〓(S) _l ) with respect to the calculated value is calculated and outputted as a degree of dissimilarity signal 94. 4 is a classification section which inputs the dissimilarity signal 94, performs processing such as assigning levels, and outputs the classification result as a signal 96. 5
is a standard feature vector storage unit, and the above 64 characters are stored for each character type.
A dimensional standard feature vector is stored, and the standard feature vector is read out from the dissimilarity calculation unit 3 as the signal 95. In the figure, a dissimilarity calculation unit 3, a classification unit 4, and a standard vector storage unit 5 are used in a normal character recognition method, etc., and the dissimilarity calculation unit 3 includes an adder, a multiplier, a comparator, and a register. Similarly, the classification section 4 can be easily realized using a comparator, a register, etc., and the standard feature vector storage section 5 can be easily realized using a normal storage element.

FIG. 2a is a diagram showing an example of the configuration of the feature pattern generation section 1. As shown in FIG.

Reference numeral 11 denotes a binary pattern storage section which inputs and stores the binary character pattern as a signal 90, which can be easily realized with a normal storage element. Reference numeral 12 denotes a feature processing area extraction unit which reads the binarized character pattern from the binary pattern storage unit 11 as a signal 911, determines a circumscribing rectangle of the character pattern, and outputs the coordinate values of the vertices of the circumscribed rectangle as a signal.
Output as 91. The feature processing is performed inside the circumscribed rectangle. The process of determining the main circumscribed rectangle is performed not only in character recognition but also in general image processing, and is carried out by the feature processing area extraction unit 12.
can be easily realized. Reference numeral 13 denotes a contour direction code generation section which reads the binarized character pattern as the signal 911, and at the same time reads the coordinate values of the vertices of the circumscribed rectangle from the feature processing area extraction section 12 as the signal 91; For each pixel inside the character pattern, detect whether it is a point on the background part, a point on the outline part, or a point inside the character pattern, and if it is a point on the outline part, a temporary slope code at the outline point is detected. In the case of a point in the background area, refer to each coordinate of the quadrant partial area obtained from the coordinate values of the circumscribed rectangle, determine a code to identify whether it is the upper left partial area, and detect the pixel. The pixel information is output as a signal 9131, and then the pixel of the characteristic pattern in which the pixel information is written and the adjacent pixel are simultaneously output as a signal 9132. At the contour point of the character area, the pixel and the pixel are The direction code at the contour point is determined from the temporary slope code of the adjacent pixel and output as a signal 9133. Reference numeral 14 denotes a feature pattern storage unit, in which each pixel stores an identifier for a point on the pattern such as the background point, the contour point, the interior point, etc., and for the contour point, the temporary slope code, the direction code, and the The background points have attributes such as the partial area identifier and the number of intersections for characterization, and the identifiers of the points on the pattern, the temporary slope codes for the contour points, and the partial area identifiers are based on the signal 9131. , the tentative slope code for the contour point is read out as the signal 9132, the direction code is written as the signal 9133, the entire pixel is read out as the signal 9141 and written as the signal 9142, After all characterization processing has been performed, each of the pixels is read out as a signal 92. 1
Reference numeral 5 denotes a crossing number counting unit which reads the coordinate values of the circumscribed rectangle using the signal 91 from the feature processing area extraction unit 12, calculates the coordinates of the four-part partial area, and calculates the coordinates of the partial area from the feature pattern storage unit 14. The pixels of the feature pattern are scanned as signals 9141 according to the corresponding scanning procedure.
The number of intersections is written as a signal 9142 in the characteristic pattern storage section 14 at points in the background area according to the direction code assigned to the contour point of the character pattern.

FIG. 2b is a block diagram showing one embodiment of the configuration of the contour direction code generation section 13. 1
Reference numeral 30 denotes a partial area determination unit, which reads the coordinate values of the vertices of the circumscribed rectangle as the signal 91, calculates the coordinates of points that divide the horizontal and vertical sides into two, and determines the four-part partial area. 4 above
The coordinates of the vertices of the four rectangles that define the divided partial areas are output as signals 9134, which can be easily realized. 131 is a local pixel processing unit that reads each reference pixel of the character pattern binarized as the signal 911 and 8 pixels adjacent to the reference pixel, a total of 9 pixels, and when the reference pixel is a contour point of the character pattern; Then, the temporary slope code is determined using the nine pixels, and the code indicating the contour point and the temporary slope code are written as the signal 9131 in the corresponding pixel of the feature pattern storage unit 14, and the reference If the pixel is an internal point of the character pattern, a code indicating the internal point is written as the signal 9131 to the corresponding pixel of the feature pattern storage unit 14, and if the reference pixel is a background point within the circumscribed rectangle of the character pattern. In this case, a code indicating the background point and a code indicating the quadrant partial area to which the background point belongs are written into the corresponding pixels of the feature pattern storage unit 14 as the signal 9131. .
Reference numeral 132 denotes a contour direction determining unit, which receives the 4th line inputted as a signal 9134 from the partial area determining unit 130.
A region to be processed is determined based on the coordinates of the vertices of the divided partial regions, and the pixels in the feature pattern storage section 14 determined by the local pixel processing section 131 are used as signals.
9132, the direction code is determined from the temporary inclination code, and is written as a signal 9133 to the corresponding pixel of the feature pattern storage section. FIG. 3a shows the positional relationship between the reference pixel and the eight adjacent pixels, where x is the reference pixel, a,
b, c, d, e, f, g, and h indicate eight adjacent pixels. FIG. 3b is a diagram showing the temporary slope code, which is expressed as a 3-digit binary number, that is, 3 bits. FIG. 3c is a diagram showing a rule for assigning the tentative slope codes 0, 1, ..., 7 when the reference pixel is a contour point on a character pattern. In the figure, the column for the slope code is + The points marked with ``*'' are inside the character pattern, so the slope code cannot be assigned to them, and the points marked with * are not regarded as the contour points, but are assigned the same code as the interior points. Indicates that the point is excluded from the process of assigning a direction code.

How to apply FIG. 3c is shown below.

If the value of the reference pixel x is 0, that is, it is a background point, it is not subject to the process of assigning the direction code. When the value of the reference pixel x is 1, FIG.
a, b, c, d, e, f, g, h are 0, 0,
In the case of 0, 0, 1, 1, 1, 1, the rows where a, b, c, d in Figure 3 c are 0, 0, 0, 0, and e, f, g, h If you pay attention to the intersecting column of the columns where is 1, 1, 1, 1, it is '3', so set the temporary slope code to 3. Similarly, if the value of the reference pixel x is 1, a, b, c,
d, e, f, g, h are 0, 1, 1, 0, 1, 0,
In the case of 1, 1, the rows where a, b, c, d in Figure 3 c are 0, 1, 1, 0, and e, f, g,
If we pay attention to the intersecting column of the columns where h is 1, 0, 1, 1, it is '0', so we set the temporary slope code to 0. In this way, the rule in FIG. 3c is applied from the values of the eight pixels adjacent to the reference pixel. The local pixel processing section 131 performs writing to each pixel of the characteristic pattern storage section 14, such as the process of assigning a temporary slope code, as described above, and therefore can be easily realized using a normal logic circuit, sequential circuit, etc.

FIG. 3d is a diagram showing the direction code, and 4
It is expressed as a binary number of digits, that is, 4 bits. Figures 3e to 1Temporary slope code shown in Figure 3b above
₃ is a diagram showing the rules for generating the direction code Y ₁ Y _{2 Y 3} Y ₄ from X ₁ X ₂ X _3. For example, when the tentative slope code X ₁ X 2 Check the tentative slope code of the pixel corresponding to b and g in FIG. The code is 111, respectively.
When the value is 000, the direction code is determined to be 1111, and when the combinations of pixels b and g shown in FIG.

Similarly, when _the tentative slope _code X ₁
When the temporary slope code X ₁ X ₂ X ₃ is 010, examine the combination of pixels c and f of the adjacent pixels;
When X ₁ X ₂ X _{3 and} 011, the combination of c and d or e and f of _the adjacent pixels is checked, and when X ₁ When the X ₁ X ₂ X ₃ is 101, the combination of the adjacent pixels e and a or the combination of d and _h is examined; when the X ₁ X ₂ The above X ₁ X ₂ X ₃ is 111
When , the combination of adjacent pixels b and h or the combination of g and a is checked, and if there is a combination of adjacent pixels that corresponds to the direction code determination rules shown in FIG. 3 e to l, the direction is determined from the direction code determination rule. The code Y ₁ Y ₂ Y ₃ Y ₄ is determined, and when there is no corresponding combination, the direction code Y _{1 Y 2 Y 3 Y 4 is determined as Y 1 Y 2 Y 3 Y 4} = X ₁ X ₂ X _{3 0} . do. These processes can be easily realized using ordinary logic circuits and sequential circuits, and therefore the contour direction determining section 132 can be easily realized.

FIG. 4 is a diagram showing an example of the pixel configuration of the feature pattern storage section 14. In FIG.

Each pixel has 8 bits of information, and bit 0 in the figure indicates the outline of the character pattern. When bit 0 is "1", it is the outline pixel, and when bit 0 is "0", it is the outline pixel. is a pixel that is not a contour part. In the case of a pixel in the contour part, bits 1, 2,
The three bits 3 correspond to the temporary slope code X ₁ X ₂ X ₃ , and the four bits 4, 5, 6, and 7 correspond to the direction code Y ₁ Y ₂ Y ₃ Y ₄ .

In the case of a pixel that is not the outline part, bit 1 is an identifier for whether it is a point in the character part or a point in the background part, and if bit 1 is "1", it is an identifier for the point inside the character, and if bit 1 is "0", it is an identifier for the point inside the character. is the point in the background. At points within the character, bits 2, 3, 4, 5, 6, and 7 are Don't Care.

In the case of the point in the background part, bits 2 and 3 are the identifier of the quadrant partial area, and bits 4 and 5 are the identifier of the quadrant partial area.
2 bits and 2 bits 6 and 7 are count information of the number of crossings in two directions, respectively. The characteristic pattern storage section 14 has an attribute for each bit as described above, and may be a normal storage element.

FIGS. 5A and 5B are diagrams effectively showing the scanning direction of the characterized pixels for counting the number of intersections in each of the quadrant regions by the intersection number counting unit 15.

As a process for counting the number of intersections, there is also a method of scanning in each of two directions using all background points as starting points, as shown in FIGS. 6a and 6c. However, although this method is suitable for explaining the characterization of pixels of background points in an easy-to-understand manner, it is not efficient in actual processing because scanning is performed many times in the same direction. However, if scanning is performed as shown in FIGS. 5a and 5b, all background points can be efficiently characterized.

The upper left and lower left partial areas of the quadrant partial areas are scanned horizontally to the right, the upper right and lower right partial areas are scanned horizontally to the left, and the upper left and upper right partial areas are scanned vertically downward. The lower left and lower right partial areas are scanned vertically upward, and intersection values initialized to 0 are sequentially written to the pixels of the background points, which intersect with the outline of the character pattern and intersect with the stroke. If the condition is satisfied, the process of adding 1 to the intersection value to create a new intersection value and writing the new intersection value to the next appearing background pixel is repeated. According to the above processing order, scanning rightward once for every row in the upper left and lower right subregions results in the same number of intersections as scanning leftward from all background points in the subregion. is found. The same holds true for rightward, downward, and upward scanning, and by performing scanning as shown in FIGS. 5a and 5b, all background points can be efficiently characterized without duplicating the number of scans.

The above processing consists of processing such as read/write processing, comparison, and addition to/from the storage section, and can be easily realized using ordinary logic circuits, sequential circuits, and the like.

(Effects of the Invention) As described above, according to the present invention, absolute position information can be reflected by determining characteristics for each partial area obtained by dividing the area defined by the circumscribed rectangle of a character pattern into four, and By determining the scanning direction for each character and extracting features, we were able to create vectors that reflect stable relative position information in as low as 16 dimensions. It can be expressed as a dimensional vector, and with this 64-dimensional vector, it is possible to provide a character classification method that can classify with sufficient performance a large variety of printed characters with complex structures and similar glyph shapes, such as kanji. .

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2a is a block diagram showing an embodiment of the configuration of the feature pattern generation section, and FIG. 2b is a block diagram of the configuration of the contour direction code generation section. FIG. 2 is a block diagram showing one embodiment. 3A to 3L are diagrams explaining the principle of the contour direction code generation section, FIG. 3A is a diagram showing adjacent pixels, FIG. 3B is a diagram showing a temporary tilt direction, and FIG. 3C 3 is a diagram showing rules for determining a temporary tilt direction, FIG. 3 d is a diagram showing a direction code, and FIGS. 3 e to 1 are diagrams showing rules for determining a direction code from a temporary tilt code. FIG. 4 is a diagram showing an example of the pixel configuration of the feature pattern storage section 14, and FIG. 5 is a diagram showing the scanning direction for each partial area when counting the number of intersections. FIG. 6 is a diagram showing the principle of the present invention. In the figure, 1 is a feature pattern generation section, 2 is a feature vector extraction section, 3 is a dissimilarity calculation section, 4 is a classification processing section, 5 is a standard feature vector storage section, 11 is a binary pattern storage section, and 12 is a feature processing section. Area extraction section, 13
14 is a contour direction code generation section, 14 is a feature pattern storage section, 15 is an intersection count section, 130 is a partial area determination section, 131 is a local pixel processing section, and 132 is a contour direction determination section.

Claims (1)

[Claims] 1. On a character pattern formed by converting a character part and a background part into a quantized signal consisting of binary values, scan in multiple directions from each point of the background part, count the number of intersecting strokes, and perform characterization. In a character classification method that uses features obtained by In the small area, scanning is performed in two directions upward and to the left from each point in the background area, and in the upper right small area, scanning is performed in two directions upward and to the right from each point in the background area, and scanning is performed in two directions upward and to the left from each point in the background area. In the small area, scanning is performed in two directions, downward and to the left from each point in the background part, and in the lower right small area, scanning is performed in two directions, downward and to the right from each point in the background part, In the upward and downward scanning, the number of times the stroke intersects with the horizontal direction and the diagonal direction is counted, and in the leftward and rightward scanning, the number of times the stroke intersects with the vertical direction and the diagonal direction is counted,
When the number of intersections is 0, 1, or 2, the number of intersections is kept as is, and when the number of intersections is 3 or more, the number of intersections is set to 3, and 2 A feature pattern generation unit that assigns 16 types of feature codes based on the number of intersections in the directions, and
The frequency of appearance of 16 types of feature codes is counted to find 16 elements, and the 64 elements found in the entire area are normalized by the area of the circumscribed rectangle, and the resulting 64-dimensional vector is used as the input feature vector. A feature vector extraction unit and a feature vector having the same format as the input feature vector.
Vector calculation is performed by comparing the input feature vector with the standard feature vector for each character type with a standard feature vector storage unit that stores standard feature vectors described in 64-dimensional vectors and prepared for each character type. a dissimilarity calculation unit that calculates the dissimilarity between the input feature vector and the standard feature vector for each character type; and a dissimilarity calculation unit that ranks each character type based on the dissimilarity and generates candidates for detailed identification. 1. A character classification method, comprising a classification processing unit that determines character types, and performs broad classification for character recognition based on feature amounts that reflect absolute position information and character pattern structure information.