JPS6037954B2 - Feature extraction processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a feature extraction processing method, and particularly to a feature extraction processing method that effectively extracts features useful for identification from patterns such as Chinese characters that have multiple structures from a topological perspective. be.

The topological features found when looking in a predetermined direction from each point in a pattern area containing one black-and-white binary pattern are stored and accumulated at each point, and the resulting high-order features are used. Tokubuta No. 51-85708 (Japanese Unexamined Patent Publication No. 53-10931) is a method for character recognition.

The principle of feature extraction processing, which is the basic part of this method, is shown in FIG. In the same figure a, scanning lines are drawn in four directions (rightward direction 1, upward direction 2, leftward direction 3, downward direction 4) from an arbitrary white point 5 in the pattern area 7 including the character part (black) 6. ,
Check whether a character part exists in each direction. A 4-bit feature vector 10 is given corresponding to each white point, and 1 is set in a predetermined position of the feature vector 10 if the character part exists, and 0 is set if it does not exist, depending on the existence status of the character part. When the white point 5 is at the position of this example, the bit position 11 corresponding to the feature vector 1 of this point or the rightward scanning line 1 is 1.
, bit position 12 corresponding to upper scan line 2 is 0, bit position 13 corresponding to left scan line 3 is 1, and bit position 14 corresponding to lower scan line 4 is 1. The above processing is called primary phase feature extraction. By this process, the white point area of the pattern area 7 is divided into several parts according to what kind of feature vector each white point has. In this example, feature vectors having features 21, 22, 23, 24, 25, 26, and 27 are extracted according to the closed state of the character portion 6, as graphically shown in FIG. For example, the feature vector 21 graphically represents that there is a character part only in the right direction 1 and the bottom direction 4, the feature vector 22 graphically represents that there is space only in the top direction 2, and the feature vector 24 graphically represents that there is no space in the direction 4. . Next, in primary phase feature extraction, high-order features are formed based on the topological features formed at each white point of the pattern area 7, that is, the feature vectors 10.

Therefore, for each white point, its own feature vector 10
In addition, when a scanning line is drawn in each direction, feature vectors of white points that are encountered for the first time beyond the character portion 6 are added while preserving direction information. When the white point 5 is located at the position of the main body, its higher-order feature becomes the illustrated higher-order feature 30 if expressed graphically. A small circle 31 in the feature 3 indicates that there is no character part in the upper direction 2 of the white point 5. Also, each black V point'
The trick is to save the direction information and add the feature vector of the white point that is encountered for the first time when scanning lines are drawn in each direction from that point. When the black point 8 is located at the position of this example, its higher-order feature is represented graphically as the illustrated higher-order feature 40. In addition, in Tokugan No. 51-85708 (Japanese Patent Application Laid-Open No. 53-10931), an integration process is performed to convert those that are unstable against character deformation into those that are stable among the feature vectors 10 extracted from the primary phase features. . For example, feature vector 24 of the feature vectors is indistinguishable from a loop, so
It is converted into a white point feature vector 26 that is in contact with 4. In addition, a feature extraction method that gives local familiarity information to a black V point by looking at the black and white state of the points adjacent to that point, and uses this as the primary phase feature of the black V point, Mochibuta Sho 51-1121792 (Special Showa 53
-47239). These above-mentioned formulas are suitable for extracting features of patterns with relatively simple topological structures such as handwritten or printed alphanumeric characters and katakana characters. The input character is identified by comparing the presence status of the higher-order features with a list of higher-order features prepared in advance for each category. In the conventional method described above, the topological feature given to a point in the white background is a primary phase that describes the closed state depending on the presence or absence of a character line when the scanning line is extended vertically and horizontally from the white point to the character frame. It is structured based on characteristics.

However, with the above characteristics, the vertical, horizontal, and
Even if the diagonal direction is relatively clear in the pattern,
It is difficult to distinguish at the feature extraction stage between features caused by graphical deformation/noise of the target character and features that describe stable stroke structure, and The disadvantages caused by the lack of topological information of the characteristics, such as the insufficient information representing the separation state of character lines, which is necessary to distinguish and characterize the existence of partial patterns from the entire character, It is included.
In order to solve these drawbacks, the present invention expresses the primary feature, which is the basic topological feature of the white point on the pattern surface, as the frame of the character line existing in a predetermined direction. The purpose of this invention is to obtain the slanted code by accumulating it on the white point, and also to express the connectivity of the character lines around the white point.The drawings will be described in detail below.

FIG. 2 shows an embodiment of the present invention, which shows a 5-way scanning section, a 6-way pattern memory section, 61 a pattern memory, a 7-way control section, 80 an identification section, and a 9-way feature extraction section. .

The main components of the feature extraction unit 90 include a river slope code extraction unit 910, a feature accumulation unit 9200, a feature code conversion unit 9300, a feature point extraction unit 9400, a background dividing line extraction unit 9500, and a surrounding area 9600. This is a closed state code extraction part. To operate this, the scanning unit 50 scans the paper surface according to a command from the control unit 70, and binary signals representing black and white of an area including one character on the paper are sequentially stored at a predetermined address in the pattern memory 61. be done.

On the pattern memory 61, a memory address consisting of a predetermined number of words is assigned to one mesh within the pattern area. When the signals in the pattern area have been stored in the pattern memory 61, receiving a command from the control unit 70,
The feature extraction unit 9 extracts the pattern from the pattern memory 6 according to a predetermined procedure.
1 is read out, a feature code representing the topological feature of each point on the stored character pattern is obtained, and is stored in the corresponding address on the pattern memory. Feature extraction section 9
When a series of feature code creation processes using 0 are completed, the contents of the pattern memory 61 are sequentially transferred to the identification unit 8.
The identification unit 80 reads out the feature codes stored in the addresses on the pattern memory 61 corresponding to each mesh in the pattern area, determines the type and number of each, and determines what the input character was based on this information. Decide and output the result. In the feature extraction unit 90, first, a forehead slant code extraction unit 9100 forms a forehead slant code representing the harshness of the character line for the black V point on the character pattern stored on the pattern memory 61.

When the operation by the cervical oblique code extraction section 9100 is completed, the pattern memory section 60 reads out the contents while scanning the pattern memory 61 vertically and horizontally according to a predetermined procedure according to a command from the control section 70, and stores the contents in the feature accumulation section 9200. Forward. The feature accumulation unit 9200 selects and extracts a slanting code from the black dots on the character lines that exist in the vertical and horizontal directions of the white dot in the pattern area, and based on this selects and extracts the oblique code from the black dots when viewed from the white dot. An integrated code is created by obtaining a diagonal code that represents the irregularity of character lines and allocating the code according to a predetermined bit configuration, and transfers this code to the address corresponding to the white dot on the pattern memory 61. do. After the operation of the feature accumulation unit 9200 is completed, the feature code conversion unit 9
300, the pattern memory 61 is
The characteristic code including the forehead oblique code read out is converted by scanning the screen vertically, horizontally, and horizontally. That is, by referring to the feature codes read from adjacent points on the pattern memory 61 on the scanning line, the forehead diagonal code is converted according to a predetermined rule, if necessary, to make the pattern more stable. A feature code is created to describe the structure and is stored at a corresponding address on the pattern memory 61. Feature accumulation unit 9200 and feature code conversion unit 93 shown above
Independently of the operation of 00, the feature point extraction unit 9400 reads the contents of the pattern memory 61 in a predetermined procedure and detects topological feature points such as end points and bending points on the stored character patterns. After the operation of the feature point extraction unit 9400 is completed, the background dividing line extraction unit 9500 extracts the topological feature points extracted by the feature point extraction unit 94001 in the process of scanning the pattern memory 61 according to a predetermined procedure. is detected, and when the scanning direction matches a predetermined direction according to the type of the feature point, a line segment that divides the white part of the background is extended starting from the feature point, and the line segment on the pattern memory is A code representing the background dividing line is stored at the address of the point corresponding to the minute. In the peripheral closed state code extraction section 9600, after the operations of the feature accumulation section 9200 and the background dividing line extraction section 9500' are completed, the control section 70
The pattern memory 61 is scanned vertically and horizontally according to the command, and when a white point on the scanning line included in a character line part or an arbitrary white area partitioned by a background dividing line is detected, the pattern memory 61 is scanned vertically and horizontally. The number of times the background dividing line intersects and its direction from passing through to reaching the incandescent point, as well as the type of character line slope code by direction accumulated on the feature code read from the white point during that time Depending on the number of occurrences etc.
A feature code expressing a character line closed state around the white area is created and stored in a predetermined bit configuration at an address corresponding to the white point. By the operation of each part shown above,
The feature extraction unit 90 stores a feature code representing the topological feature of each point on the character pattern stored in the pattern memory 61. Next, each processing unit making up the feature extraction unit 90 will be explained in detail. The slope code extraction unit 9100
The directional component of the character line is determined by counting the number of consecutive black V points in each predetermined direction from an arbitrary black V point on the character line to obtain the distance in each direction, and comparing the distances in each direction. By the method described above, a slant code representing the slant of strokes forming a character pattern is formed.

Several stroke amount oblique extraction methods of this type are already known, and known techniques can be used. FIG. 3 shows an example of a bit configuration expressing a prefix code obtained as an output of the slant code extracting section 9100 at an address corresponding to a black V point on a character pattern, and the bit structure described by the slant code. This shows the definition of the directional component of the stroke.
In this embodiment, the stroke directions are rightward 51, leftward 52, upward 53, downward 54, upper right 45o direction 55, lower left 45o direction 56, upper left 45o direction 57, lower right 45o direction.
It is assumed that the image is transformed into a doji in a total of 8 directions (4 pairs of directions 58). 61
0 is a feature code consisting of one or more words prepared at the address corresponding to the mesh on the pattern memory 61, and when the mesh is located on the character line, the black and white information section 611 becomes "1". , the forehead diagonal code section 620' is a bit area for storing the forehead diagonal code detected on the black V point.

Here bit positions 621, 622, 623, 624, 6
25, 626, 627, 628 are each stroke direction 5
1, 52, 53, 54, 55, 56, 57, and 58. For example, if a stroke is detected in 51 directions centering on the black point V, bit 621 is set to "1".
On the other hand, when no stroke is detected in the 51 direction, the bit 621 is set to "0", thereby encoding the rotational oblique code 620 at the sunspot with an 8-bit area representing the presence or absence of a stroke in the 8 directions. FIG. 4 shows the binary pattern 500 on the pattern memory 61 and the complex code extraction section 91.
001 An example of the slant code of the black V point on the character line of the pattern thus formed is shown as several points 501, 502, 503.
It is illustrated by.

The bit string 506 is connected to the point 501 as shown in the graphical representation 511.
is a diagonal code indicating that the point 502 is on a vertical stroke extending in the 53 and 54 directions, and the bit string 507 is on a stroke in the upward right direction where the point 502 is extending only in the 55 direction, as shown in the graphical representation 512. This is a complicated code indicating that. Further, the bit string 508 is a slanting code indicating that, as shown in the graphical representation 513, at the point 503, a series of character lines not only in the 54th, 55th, and 56th directions but also in the 52nd and 53rd directions was detected. 500 character patterns
This code should be considered as noise that is inconsistent with the structure of the skeleton, but since the character lines have varying thicknesses, it is difficult to suppress the occurrence of such noise. This type of noise can be dealt with in the feature accumulation unit 9200 (described later). FIG. 5 shows a feature accumulation section 9200, a feature code conversion section 9300, a background dividing line extraction section 9500, and a peripheral closed state code extraction section 96.
An example of the bit configuration of a feature code 630 formed at an arbitrary white point on a pattern area by 00 etc., and an accumulation that constitutes the feature code and expresses character line nostalgic information in four directions around the white point. An example of code is shown. bit string 651,
652, 653, and 654 indicate that when a scanning line is drawn in a predetermined direction from an arbitrary white point, there are character lines with horizontal, vertical, upward-right, and upward-left directions. The bit string 655 is an integrated code indicating that there is no character line on the extension of the scanning line, and the bit string 656 is the cervical oblique code extracted from the black dot on the first character line on the scanning line. This is an integrated code indicating that the slope of the character line could not be uniquely coded because it was judged as unstable by the circuit 9211, and the corresponding symbols 661, 662,
663, 664, 665, and 666 are symbols used to graphically represent the integrated code on the white point, an example of which is shown in FIG. The feature code 63 of the white point, the bit 631 for distinguishing between the black V point and the white point, and the integrated code representing the slant of the character line correspond to the direction in which the character line exists starting from the white point. The integrated code section 640 arranges the right direction 51 at 641, the left direction 52 at 642, the upward direction 53 at 643, and the downward direction 54 at bit positions 644, and the feature code conversion section 9300 described later converts the integrated code into the integrated code section 640. The converted accumulated code section 645 stores the results of code conversion, and the pig section 635 encodes other accumulated information formed by the background dividing line extraction section 9500 and surrounding closed state code extraction section 960, which will be described later. Become. FIG. 6 is a block diagram showing one embodiment of the feature accumulation section 9200.

Based on the command from the control unit 70, the characteristic code of the black V point and the white point read out from the pattern memory 61 by scanning the character pattern are input to the input register 921 through the connection 303.
Consolidated into 0. Note that at this point, features corresponding to the white dots have not yet been formed. In response to the characteristic code transferred from the black dot, the oblique code determination circuit 9211 operates, and the oblique code 620 on the characteristic code 610 shown in FIG. The integrated code selection unit 9212 operates only when it is determined that the bit string 65 shown in FIG.
Selecting and outputting an accumulation code among the accumulation codes represented by 1,652, 653, 654, 655, and 656 that corresponds to the case where the oblique information of the black point is accumulated of the white point in the scanning direction, The contents of the integrated code register 9213 consisting of 3 bits are updated. In addition, the input register 921
If the data above 0 is transferred from the white point, the integrated code synthesis unit 9215 operates, and the integrated code register 9213
The contents of the white point feature code 63 on the input register 9210
The value around 0 is stored at a predetermined bit position in the oblique code accumulation portion 640 corresponding to the accumulation direction determined by the scanning direction as described in FIG. The result of this processing means that the character line slanting code in the direction opposite to the scanning direction is accumulated on the characteristic code of the white point. Subsequently, one word created by the integrated code synthesis section 9215 according to a command from the control section 70 is stored at the corresponding address on the pattern memory 61 through the connection 304. An example of the logic for determining the stability of forehead oblique information by the oblique code determination circuit 9211 will be described below.

As shown in FIG. 3, bit pairs 62
1,622, bit pair 623,624, bit pair 625
, 626, bit pairs 627, 628, etc., when one or both bits are set to "1", they represent the presence of horizontal, vertical, upward right, upward left stroke components, respectively. Therefore, of the four bit pairs mentioned above, only one bit is "1", and all others are "0".
We regard the zeni-shaku code as stably expressing the character line's angular information. And "1" for multiple bit pairs
When a bit exists, it is difficult to uniquely determine the mirror slant of a character line based on the code ', so this is regarded as unstable slant information. For example, in the process of scanning in the right horizontal direction 51, bit pairs 621, 6
It is also possible to determine stability in consideration of information in the scanning direction, such as by discarding a slope code in which only bit 622 corresponding to the left horizontal direction 52 is "1" as an unstable information code.

FIG. 7 is an example of this, in which when scanning in the right direction 51, the forehead oblique code indicated by the graphical representation 526 at the point 521 on the character line of the character pattern 520 is determined to be stable in the vertical direction. It is transferred to the integrated code register 9213, but since only bit 622 of the confusing code shown by the graphical representation 527 at point 522 is "1", it is considered unstable for the above reason and is written to the integrated code register. do not have. As a result, at the white point 523 on the extension of the scanning line, the influence of the protrusion near the point 522 is removed, and the information near the point 521, which corresponds to the three branch points of the character line, is saved as character line information in the left direction 52. and can be accumulated. Various methods are possible on a case-by-case basis for determining the stability of the Enoki oblique information as shown in the example above, but the details will be omitted. By the operation of the nostalgic code determination circuit 9211, the integrated code register 9
The contents of 213 are updated one after another in the process of scanning the black dots on the character line, and for white dots, the content of the point closest to the white dot among the black dots with stable forehead slant information located on the scanning line is updated. The oblique information will be encoded and accumulated.

Initial set circuit 9214 writes a predetermined integrated code into integrated code register 9213 based on a command from control unit 70 to initialize the contents of the register.

The content of the initialization is that when scanning starts from the peripheral character frame into the pattern area, an integrated code 655 (5th
However, at the time when the scanning line moves from the white point to the black line, there may be cases where no stable slant code is extracted from the black point on the character line, so the accumulated code 656 (Fig. 5) is used. are transferred to the integrated code register 9213, respectively. Through the above-described processing, feature codes 630 are accumulated at the addresses corresponding to all the white points in the pattern area, encoding the slope of the character line that first crosses when the scanning line is extended in the vertical and horizontal directions from the white point. It is formed.

FIG. 8 shows a diagrammatic representation thereof as an example. White point 50 in the diagram
4, as shown in the graphical representation 531, there are character lines upward to the right, vertical character lines to the right, and further character lines to the left and down to the character frame. A feature code is created indicating its absence. On the other hand, in the white point 505, as shown in the graphical representation 532, there are no character lines in the upper and left directions, there are character lines in the upper right direction in the lower direction, and there are characters in the right direction. Although the line exists, from the black V dot series that first appears on the scanning line extending rightward from the point 505, it becomes unstable, as typified by the disturbing code 508 (FIG. 4) formed at the black dot 503. ”
When it is determined that , a characteristic code indicating that only oblique codes have been extracted is formed. An uncertain accumulation code (* in the graphical representation 532
It is not impossible to take measures such as leaving an uncertain feature code with a mark (mark) as it is and setting the special feature code as dontcare at the stage of identification processing, but as character patterns become more complex like many kanji, This type of indeterminate feature appears more frequently, resulting in many problems such as difficulty in extracting features that reflect the detailed structure of the pattern. This problem,
This problem can be solved by an integrated code determination method that combines a plurality of character line slope information extraction means. This is a block diagram showing an embodiment of the contour shape determining section 922 shown in FIG. 6. The judgment circuit 9221 is an input register 9210
By checking the contents of the character line, the point in time when scanning shifts from the black point to the white point on the character line is detected, and the contents of the integrated code register 9213 at this point are read out. When it is determined that the integrated code stored in the integrated code register is the uncertain integrated code, the character line contour tracing section 9222 outputs a signal to the control section 70 via the connection 701. In response to this, the control unit 70 saves and saves the current pattern memory register in the pattern memory unit 6, and stores the pattern outline in a certain range starting from the previous address (corresponding to the black dot with the unstable accumulation code). is traced and the address of each point is sequentially transferred to the character line contour tracing unit 9222 via the connection line 305. At this time, the forehead slant determination unit 9223 determines the slant of the contour line based on, for example, changes in the vertical and vertical components of the address information of the point on the argument line issued by the character line contour tracing unit 9222, and The determination result is given to the integrated code selection section 9212 and set in the integrated code register 9213 of a predetermined complicated code. When this series of processing is completed, the control unit 70 issues a command to the pattern memory unit 60 to continue scanning from the pattern memory register 60 that has been saved. In the case of the point 505 having the lower definite feature shown in the graphical representation 532 in FIG.
An example of the effect of the operation is shown in FIG.

When scanning the scanning line 534 from the right to the left, the black V point 503
The features extracted from the black V point sequence 535 of the character line portion containing
6, the initial set circuit 921 is placed on the integrated code register 9213 that provides the integrated code from the scanning direction as shown in the bit string 537.
An uncertain accumulation code 656 preset by 4 is stored. Therefore, the contour shape determination unit 9220 operates, for example, traces the black dot rows 538 and 539 on the character line contour by a certain number of meshes, and at this time, the distance changes 540 and 541 on the horizontal pattern memory and the vertical pattern memo I
The inclinations of the character line contours 538 and 539 are detected by using distance changes 542 and 543 on J, and the contents of the integrated code register 9213 are converted to a code 653 (fifth Update to the form shown in Figure). As a result, the code is accumulated at the white point 505 located on the scanning line 534, and finally the feature code shown by the graphical representation 533 is formed. FIG. 10 shows a block diagram of an embodiment of the feature code converter 9300.

The characteristic code of the white point read from the pattern memory 61 by scanning the pattern memory 61 in the vertical and horizontal directions according to the command from the control unit 70 is stored in the current register 93 through the connection 306.
It is stored in 01. The table search circuit 9303 searches the contents of the past register 9302 storing the feature code of the white point finally written in the pattern memory during scanning up to the point immediately before reaching the white point on the scanning line, and the contents of the current register 9302.
1, mirror skew codes relating to specific directions predetermined, such as two directions perpendicular to the scanning direction given by the command from the control unit 70, are extracted. Then, the corresponding address on the code conversion table 9304 in which the contents of the forehead oblique code of the white spot after the conversion process determined by the double oblique code N is stored is specified. The converted code synthesis circuit 9305 converts the converted complicated code read from the code conversion table 9304 into a direction among the bit positions 646, 647, 648, 649 on the feature code 630 (FIG. 5) of the white point. The contents of the corresponding bit positions in each direction of the integrated code section 640 on the feature code are stored as they are in the remaining three bit positions.
The contents of the current register 9301 partially modified by the operation of the conversion code synthesis circuit 9305 are stored in the past register 93.
02 and written to the corresponding address on the pattern memory through the connection 307. The effect of the operation of the feature code conversion unit 9300 depends on the definition of the code conversion rule stored in advance in the code conversion table 9304, and enables noise removal, global feature extraction, and the like.

FIG. 11 shows an embodiment of the code conversion process in the vicinity of a white point 509 on the character line pattern 500 of FIG.

On the scanning line 550 including the white point 552 of interest (corresponding to the white point 509 shown in FIG. 4), first in the 53 direction, then in the 53 direction.
Regarding the case of scanning in four directions, for the sake of simplicity, we will focus only on the transition of the diagonal code in the right direction 51 at each of the white points 551, 552, and 553. Graphical representation 554,
555 and 556 represent the accumulated codes in each direction at the white points 551, 552, and 553 before the feature code converter 9300 operates. 557 is an example of a conversion table of accumulated codes in the right direction 51 of an arbitrary white point when scanning in the upward direction 53 or downward direction 54.

The column of symbols shown at 558 shows the contents of the corresponding accumulated code on the current register, and the row of symbols shown at 559 shows the contents of the corresponding accumulated code on the past register, and the symbols on the table are changed according to the code conversion table. Represents the output of the rewritten integration code.

The graphical representations 563, 562, and 561 indicate that when processing is performed on the scanning line 550 in the 53 direction with the white points 553, 552, and 551, the integrated code of each point given to the graphical representations 556, 555, and 554 is Shows the result of code conversion. Here, 564, 565, 566 are current registers, 567, 56
8,569 are the contents of the accumulated code storage portions in the right direction 51 of the past registers. For example, at point 552, the table 557 Search for new right direction 51
By applying the accumulated code of 2 to the past register 568, the accumulated codes of each direction shown in the graphical representation 562 are newly obtained. Similarly, the graphical representations 571, 572, 573 are scan line 5
50 and downward 54 to white points 551, 552, and 553, the accumulated codes of each point given to graphical representations 561, 562, and 563 show the results of further code conversion. Here, 574, 575, and 576 are the contents of the current register, and 577, 578, and 579 are the contents of the past register, respectively, in the integrated code storage portion in the right direction 51. As a result, when a code conversion table as shown in 557 is provided, the white point 551
.
, 572, 573, an integrated code exhibiting a stable global graphical structure is formed. In the feature point extraction unit 94001, several methods are already known for extracting topological feature points such as end points and bending points on a character pattern, and known techniques can be used.

In particular, as shown in the above embodiment, the slope code extractor 9
100, the number of consecutive black dots in each predetermined direction from an arbitrary black point on the character line is counted to obtain the distance in each direction, and by comparing the distances in each direction, the directional component of the character line is determined as shown in FIG. 6201 When the encoding method shown above is used, the oblique code can be used as it is as an expression of topological feature points, and in that case, the feature point extracting unit 9400 uses the oblique code extracting unit 9100. can be included. As an example, in FIG. 4, a black dot 502 with a feature code 507, which is a graphical expression 512, represents that character lines are connected for a certain length or more only in the right direction, and such a feature code is Let us define the black point with the end point as the end point. Background dividing line extraction unit 950
FIG. 12 shows a block diagram of one embodiment of 0.

The black dot characteristic code 610 and the white dot characteristic code 630 read out from the pattern memory 61 by scanning vertically and horizontally according to a command from the control unit 70 are stored in the input register 9501 through the connection 310. In the case of a sunspot accumulation code, the topological feature point extraction unit 9502 operates, and the information is given by the content of the slope code shown at the bit position 620 on the feature code 610 (FIG. 3) and the command from the control unit 70. The presence or absence of a background dividing line is determined based on the scanning direction, and a background dividing line code determined by the scanning direction is set in the background dividing line code register 9503. FIG. 13 is an example of the 67 river background dividing line code shown in FIG. bit 6
71, 672, 673, 674 are respectively 51,
"1" when there are background dividing lines in the 52, 53, and 54 directions
Therefore, when no background dividing line in any direction exists on the white point, bits 671, 672, 673, and 674 are all "0". In determining the presence or absence of a background dividing line, it is determined that the slanted code of the black V point shown in FIG.
When the condition that only one bit is "1" is satisfied, "
It is assumed that a background dividing line is generated in a scanning direction other than the direction corresponding to the bit that is "1", and for this purpose, 1 is set in the bit position corresponding to the scanning direction on the background dividing line code 670. Ru. Bits 671, 672, and 67 are set when a Japanese oblique code with a black V point that does not satisfy the above conditions is detected.
3,674 are all reset to 01. FIG. 13 shows black dots 584, 585 on the character pattern 500 as an example.
, 586 is a graphical representation of the 684, 685.68
6, and the background dividing lines generated starting from each black point are 851, 852, 852, and 852 starting from the black V point 584, respectively.
853, 854 and 855, 8 starting from black point 585
56 Furthermore, 857, 858, 8 starting from sunspot 586
59. When the content of the input register 9501 is a white dot feature code, the background dividing line code synthesis unit 9504 operates and combines the background dividing line code into the characteristic code 630 (FIG. 5) transferred from the input register 9501. The contents of the represented bit position 680 are updated by performing a bitwise OR with the contents of the background dividing line code register 9503, and this data is stored in the corresponding address on the pattern memory 61 via the connection 311.

Bit representations 681, 682, 683 in Figure 13 are white dots 5
This is the content of the background dividing line code 680 on the feature code 630 in 81, 582, and 583. For example, white point 581
Then, the background dividing line 85 from the black points 584, 685, 586
It is shown that 4,855,857 intersect. A block diagram of an embodiment of the peripheral closed state code extraction unit 9600 is shown in FIG.

The characteristic codes of the black dots and white dots read out from the pattern memory 61 by scanning vertically and horizontally according to a command from the control unit 70 are stored in the input register 961 through the connection line 312. In the case of a feature code with a black V point, the initial set circuit 96
03 is activated, and the peripheral closed status code register section 960
"IJ" is set in the bit position corresponding to the direction opposite to the scanning direction on 5. Reference numeral 695 shown in FIG. 697, 698, and 699 represent peripheral closed states in the directions 51, 52, 53, and 54, respectively, and the characteristics of the white spot outside the frame (Fig. 5) 630
This corresponds to the peripheral closed state code description part 6901 above. For example, if the characteristic code of a black point is stored in the input register 9601 during the process of scanning in the right direction 51, the left direction 52 will be described later for each white point read out from the scanning line following the black point. In order to express that the character line is closed in meaning, the bit position 697 corresponding to the left direction 52 is set to "1". When the content of the input register 9601 is a feature code of a white point, the background dividing line intersection detection unit 9604 analyzes the background dividing line code 680 obtained by the background dividing line extracting unit 95001, and extracts the scanning line It is determined whether or not there is an intersection with the background dividing line in the direction orthogonal to the background dividing line. When the background dividing line is not detected, two of the bit portions 640 representing the integrated code on the feature code 630 are orthogonal to the scan line.
When the accumulation code corresponding to a direction is a definite accumulation code other than the uncertain accumulation code 656 shown in FIG. do.
For example, when scanning in the right direction 51, the number of confirmed accumulated codes in the upward direction 53 is counted by a counter 9611, and the number of confirmed accumulated codes in the downward direction 54 is counted by a counter 9612. On the other hand, if a background dividing line is detected,
The peripheral closed state code reversing unit 9607 operates, and the code counter 961 corresponding to the direction from which the background dividing line starts
When the content of 1 or 9612 exceeds a predetermined threshold value, and the bit corresponding to the direction opposite to the scanning direction among the contents of register 695 of peripheral closed state code register section 9605 becomes "1". If the bit contents are r
peripheral closed status code register section 960
5 internal register 695. code counter 96
The contents of 11 and 9612 are reset to 0 when the background dividing lines having their starting points in the respective corresponding directions are detected and the above processing is completed. The peripheral closed state code synthesis unit 9602 extracts the characteristics of the white dot on the input register 9601 and the inner part of the register 695 of the peripheral closed state code register unit 9605 after bit inversion in the peripheral closed state code inversion unit 9607 has been completed if necessary. By performing an OR with the contents of the corresponding bit position 690 in the code 630, the contents of the surrounding closed state code description portion 690 (FIG. 5) of the white point are updated. Subsequently, the control section 70 stores this content at the corresponding address on the pattern memory 61 through the connection 313. FIG. 16 shows an example of the operation results of the peripheral closed state code extraction unit 9600 described above.

White dot 8 on the same scanning line 810 in the character pattern 500
Bit position 6 of feature code 630 at 11,812
401, the forehead oblique codes shown in graphical representations 814 and 815 are accumulated through processing in the feature accumulation unit 9200'. By the operation of the peripheral closed state code extraction unit 9600, peripheral closed state codes shown at 816 and 817 are extracted at bit positions 690 on the feature code 630 of the white points 811 and 812, respectively. The peripheral closed state code is written in the right direction before and after the white point 813 on the background dividing line 854 when scanning the scanning line 810. 51, the bit position 69 represents the peripheral closed state in the left direction 52.
7 (FIG. 15), while when scanning in the left direction 52, bit position 696 (representing the peripheral closed state in the right direction 51)
(Fig. 15) are extracted by changing from "1" to "0", respectively. The oblique code represented by the graphical representation 814 obtained by the feature accumulation section 920 and the surrounding closed state code 816 obtained by the surrounding closed state code extraction section 9600 are moved in directions 51, 52, 53,
By matching and combining 54, graphical representation 8
A characteristic code shown in 25 is obtained. In the graphical representation, "V'" means that bit 821 is "0" in code 816.
Corresponding to this, the difference from "V" is that there is a vertical stroke in the right direction 51 from the white point 811 as well as in the left direction 52, but there is a vertical stroke in the figure between the white point and the stroke. 84
Background dividing line 85 due to the influence of the character line segment machine part shown in 0
It is shown that the character line state is different in either an upward direction 53 or a downward direction 54 that is orthogonal to the right direction 51 with 4 as a boundary. Similarly, by combining the graphical representation 815 and the code 817 regarding the white point 812, the graphical representation 8
A characteristic code shown in 35 is obtained. As explained above, according to the present invention, features that describe the pattern structure can be extracted, and the identification process is consistent with that of Tokubuta No. 51-85708. Compared to the case where only the existence of directional character lines could be expressed,
Features coded by accumulating slant information of character lines existing in each direction can be extracted from each white point.

For this reason, it has the advantage of being able to extract features that are stable against noise for patterns that have many straight strokes and have clear directions, such as kanji characters. In addition, the feature that combines the feature that accumulates the character line slope information and the surrounding closed state code that is extracted using the feature can be used to extract features that focus on only a part of the character pattern, It is used as an effective feature for identifying addresses that consist of several partial patterns, such as.

[Brief explanation of the drawing]

Figure 1 is an explanatory diagram conceptually explaining the conventional feature extraction shown in Samurai Sho 51-85708, Figure 2 is a block diagram of an embodiment of the present invention, and Figure 3 is the second figure. FIG. 4 is an explanatory diagram showing the bit structure of the feature of the black V point obtained by the diagonal code extraction section; FIG. The figure is an explanatory diagram showing the bit configuration of the feature code of a white point obtained by the feature accumulation section, feature code conversion section, background dividing line extraction section, and peripheral closed state code extraction section in FIG.
FIG. 7 is a block diagram of an embodiment of the feature accumulation unit shown in the figure; FIG.
9 is an explanatory diagram diagrammatically showing the concept of features obtained by the feature accumulating unit on a pattern area, FIG. 9 is an explanatory diagram illustrating the processing of the contour shape determining unit shown in FIG. The figure is a block diagram of an embodiment of the feature code conversion section shown in FIG. 2, FIG. 13th example of a block diagram of a line extraction section
The figure is an explanatory diagram explaining the processing of the background dividing line extraction section for a pattern region, FIG. 14 is a block diagram of an embodiment of the surrounding closed state code extraction section, and FIG. 15 is a diagram showing the processing of the surrounding closed state code extraction section. FIG. 16 is an explanatory diagram showing the bit configuration of a surrounding closed state code of a white point. FIG. 16 is an explanatory diagram that graphically shows the concept of the surrounding closed state code on a pattern area. In the figure, there are 5 scanning lines, 60 a pattern memory section, 61 a pattern memory, 70 a control section, 80 an identification section, 90 a feature extraction section, 9100 a complex code extraction section, and 920 a feature accumulation section. , 930 represents a feature code conversion unit, 9400 represents a feature point extraction unit, 9500 represents a background dividing line extraction unit, and 9600 represents a peripheral closed state code extraction unit. Figure 3

Claims (1)

[Scope of Claims] 1. Topological features found when looking in a predetermined direction from each point of a pattern area including one black-and-white binary pattern are accumulated at each point while storing direction information, and the result is In a character reading method that identifies characters using the obtained higher-order features, a slope code is formed by detecting the slope of a character line as the topological feature and encoding the slope as a slope code on a corresponding black point on a pattern surface. A feature extraction unit comprising an extraction means and a feature accumulation means for forming a higher-order feature by accumulating the slope code on the character line existing in a predetermined direction with respect to the white point in association with the direction. A feature extraction processing method characterized in that the feature extraction unit extracts a feature code representing the presence and slope information of a character line around a white point. 2 Scan the characters on the document and convert the scanning signal into black and white2
a scanning unit that converts and outputs a value; a pattern memory unit that stores a pattern area including one binary pattern as part of an M×N mesh with a plurality of bits per mesh; a feature extraction unit that processes a pattern to form a feature code representing a topological feature on the pattern memory corresponding to each point on the binary pattern; In the character reading device, the character reading device is composed of an identification section that determines whether the black dot belongs to a black dot, and a control section that generates a timing signal necessary for each circuit. , scan the pattern surface in a predetermined direction, select the slope code on the character line, associate the slope code with the direction, accumulate white points on the pattern memory, and form a new higher-order feature. Feature extraction according to claim 1, characterized in that the identification section also uses the higher-order features on the white points to make a comprehensive judgment and identify the characters on the document. Processing method. 3 Extraction of the slope code that expresses the slope of the character line is as follows:
Means for detecting the inclination of a character line by determining the length of a series of black dots on a character line in several predetermined directions and comparing the sizes in each direction; and a means for detecting the inclination of the contour line with respect to the white background part by tracing the contour line, and the inclination code is determined by combining and determining the inclination information of the character line detected by the two means. Claim 1
Feature extraction processing method described in section. 4. Comparing the higher-order features on the white point with higher-order features on neighboring white points in a predetermined direction from the white point, and if necessary, applying a slope code integrated to a part of the higher-order feature. A feature extraction processing method according to claim 1, characterized in that a means for code conversion is provided according to a predetermined rule, and a new feature resulting from the code conversion is extracted as a higher-order feature. . 5. Means for extracting feature points such as end points, branch points, bending points, etc. as topological features on a character line and encoding them into corresponding black points, and a method for determining the type of feature points using the feature points on the black points as a starting point. means for extending a background dividing line that divides a white part of the background in a predetermined direction according to the background dividing line and forming a code representing the direction of the background dividing line at a white point located on the background dividing line; and a character line or the background dividing line. The area of the white part of the background separated by a line is regarded as one unit area, and the type of higher-order feature of the white point within the unit area that exists in a predetermined direction from the area, and the type of higher-order feature that exists in the direction A means for forming a code at a white point in the area, which is determined from the direction and number of the background dividing lines and representing a closed state around the unit area, is provided, and the discontinuity of the surrounding character lines is detected on the white point. 2. The feature extraction processing method according to claim 1, wherein a feature code representing a topological feature reflecting the above is formed.