JPS6047636B2 - Feature extraction processing method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention is effective for identification using a feature extraction processing method, particularly topological feature points such as patterns, end points, and branch points of Chinese characters that have a topologically complex structure. The present invention relates to a feature extraction processing method for efficiently extracting features.

The topological force features found when looking in a predetermined direction from each point in a pattern area containing at least one black and white binary pattern are accumulated at each point while preserving the information of each detected direction. , Japanese Patent Application No. 51-85708 (Japanese Patent Application No. 53-10931) is known as a character reading method for identifying characters using higher-order features obtained as a result.

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 (right 1, top 2, left 3, bottom 4) from an arbitrary white point 5 in the pattern area 7 including the character part (black) 6, and each direction is A code 10 encoded with 4 bits is prepared corresponding to each white point, and if it exists, it is Rlj. If it does not exist, it is Rlj. Set the feature code 10 at a predetermined position.When the white point 5 is at the position of this example, the feature code 10 of this point is that the bit position 11 corresponding to the right scanning line 1 is r1, and the upper scanning line Bit position 12 corresponding to RO
The bit position 13 corresponding to the lower scanning line 4 is r1, and the bit position 14 corresponding to the lower unidirectional scanning line 4 is RL. The above process is called primary phase feature extraction. By this processing, the white dot area of the pattern area 7 is divided into several parts according to what kind of feature code each white dot has. This example is graphically shown in Figure b. 2 depending on the closing status of the character part 6 like so.
The feature codes indicated by 1, 22, 23, 24, 25, 26, and 27 are extracted. For example, 21 graphically indicates that there is a character portion only in the right direction 1 and the downward direction 4, 22 indicates that there is space only in the top direction 2, and 24 indicates that there is no space in the direction. Next, higher-order features are formed based on the topological features formed at each white point in the pattern area 7, that is, the feature code 10, after the first-order topological features have been extracted.

Therefore, for each white point, its own feature code 1
In addition to 0, a characteristic code of a white point that is encountered for the first time beyond the character portion 6 when a scanning line is drawn in each direction is added while preserving the direction information of the scanning line. When the white point 5 is at the position of this example, its higher-order feature is represented graphically as the illustrated feature 30.
become that way. In the higher-order feature 30, a small circle 31 indicates that there is no character portion in the upper direction 2 of the white point 5. Also,
For each black point, the feature code of the white point encountered for the first time when scanning lines are drawn in each direction from that point is added while preserving the direction information of the scanning line. 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. Furthermore, in Japanese Patent Application No. 51-8708 (Japanese Patent Application No. 10931/1982), among the feature codes 10 extracted from the primary phase features, those that are unstable against character deformation are converted into stable ones. I am trying to perform an integration process. For example, the feature code 24 of the feature codes is indistinguishable from a loop, so it is converted into the feature code 26 that is in contact with the feature code 24. Also, the point adjacent to the black point is A feature extraction method that looks at the white and black states, gives local area information, and uses this as the primary phase feature of the sunspot, Japanese Patent Application No. 51-121792 (Japanese Patent Application No. 53-47)
No. 239) is known. These methods 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 this way, the conventional method described above can obtain a feature code that is resistant to noise and character deformation and expresses the structure of the pattern in considerable detail, but on the other hand, it plays an important role in identifying patterns that have a complex structure of strokes such as kanji. In some cases, the ability to describe topological feature points such as end points, branch points, intersections, and bending points that are considered to represent Moreover, on multiple topological feature points,
The drawback was that it was not possible to extract feature codes representing higher-order features that described the positional relationships of bottom, left, and right.

In order to solve these drawbacks, the present invention integrates information codes of topological feature points such as end points, intersections, branch points, bending points, etc. that exist in a predetermined direction on a white point on a pattern surface. The purpose of this invention is to enable extraction of feature codes expressing higher-order features, and the drawings will be described in detail below.

FIG. 2 shows an embodiment of the present invention, in which 50 is a scanning section, 60
is a pattern memory section, 61 is a pattern memory, 70 is a feature extraction section, 80 is an identification section, 90 is a control section, 100 is a feature accumulation section, and 810 is an output terminal. To operate this, the scanning unit 50 scans the paper surface according to a command from the control unit 90, and the binary signals (black and white) of the area containing one character on the paper are sequentially stored at a predetermined address in the pattern memory 61. be done.

On the pattern memory 61, memory addresses consisting of a predetermined number of words are assigned to one mesh within the pattern area. When the signals in the pattern area have been stored in the pattern memory 61, the feature point extraction unit 70 reads out the contents of the pattern memory 61 according to a predetermined procedure in response to a command from the control unit 90, and Detect end points, intersections, branch points, bending points, etc. as topological feature points. When any of the above feature points is detected, a feature point code representing the topological feature point is stored in a predetermined number of bits of the corresponding address on the pattern memory 61. When the operation by the feature point extraction unit 70 is completed,
The pattern memory 60 scans the pattern memory 61 vertically, horizontally, or diagonally according to a predetermined procedure according to a command from the control unit 90, reads out the contents, and then reads the contents of the pattern memory 61.
0, for the content read from the address corresponding to the white point in the pattern area, feature point codes of topological feature points that exist above, below, left, right, or diagonally of the white point are distributed according to a predetermined bit configuration. This rewrites and synthesizes a feature code, and transfers this to the corresponding address on the pattern memory 61. When a series of operations using the pattern memory section 60 and the feature accumulation section 100 are completed, the contents of the pattern memory 61 are sequentially transferred to the identification section 80. Identification part 8
At 0, the feature code recorded at the address on the pattern memory 61 corresponding to each mesh in the pattern area is read out,
The type and number of each character are determined, and based on this information, it is determined what the input character was, and the result is sent to the output terminal 8.
Output to 10. Note that the feature accumulation unit 100 performs other feature accumulation processing in parallel with the above processing of accumulating feature point codes consisting of end points, intersections, branch points, and bending points on the character pattern onto white points. A plurality of types of codes obtained as a result of the integration process can be allocated to a predetermined bit configuration to synthesize a feature code, and the result can be provided to the identification unit 80, but the above-mentioned other integration process and its circuit configuration An example of this is Japanese Patent Application No. 51-85708 (Japanese Patent Application No. 51-85708)
No. 3-10931), and since it is not directly related to the gist of the present invention, detailed explanation will be omitted. The feature point extracting unit 70 thins the character pattern stored in the pattern memory 61, and then overlays a template of each feature having a size of, for example, 3×3 mesh on the black dots on the skeleton of the character line obtained. A method of pattern matching with a combined thinning pattern, or a method of determining based on the number of times the character line intersects the circumference of a small circle with a constant radius centered on an arbitrary black point on the character line and the position of the intersection point. Then, topological feature points such as end points, intersections, branch points, and bending points are detected and stored at predetermined addresses in the pattern memory 61. This type of feature extraction method is already known, and known techniques can be used. Figure 3 shows a thinning pattern 700.
And the end points 701, 702, 703, which will be detected from now on,
704, 705, intersection 706, branching point 707, 708,
709 and bending point 710 are shown as an example. The feature point code recorded at the corresponding address on the pattern memory 61 as the output of the feature point extractor 70 can be arbitrarily determined, and FIG. 4 shows an example bit configuration thereof. 610 is a memory consisting of one or more words prepared at an address corresponding to one mesh on the pattern memory 61 (contents of the address), and 620 is a bit portion that describes the feature point code of the mesh.

Bit 629 indicates RlJ. if the mesh point is detected as the topological feature point. In other cases, RO
Discriminate the feature points as J, bits 621, 622, 6
23, 624, 625, 626, 627, 628 are when the mesh point is detected as a topological feature point,
611, 612, 613, 614, 615, 616 respectively
, 617, and 618, the type of the topological feature point is coded by specifying RlJ only when the character lines are continuous in the directions of , 617, and 618.

FIG. 5 shows that feature point codes representing topological feature points such as end points, intersections, branch points, bending points, etc. existing in a predetermined direction with respect to the white point are accumulated in the feature accumulation unit 100 to obtain higher-order FIG. 2 is a block diagram of an embodiment of a high-order circuit for forming features.

In this part, for each white point on the pattern area 7, the sixth
As shown in the figure, an area 6 sandwiched between scanning lines 631, 632, 633, and 634 passing above the white point 630 from the top, bottom, left, and right directions and a predetermined number of scanning lines on both sides of the center line.
End points that exist inside 41,642,643,644
Among the topological feature points such as intersections, branch points, and bending points, the code of the one closest to the center scanning line 631, 632, 633, and 634 is enhanced with a predetermined bit configuration, and the code is placed on the white point 630. A feature code representing the higher-order features of is formed. For convenience of explaining the operation contents in FIG. 5, the contents 61 of the address corresponding to one mesh on the pattern memory 61
Reference is made to FIG. 7, which depicts one embodiment of a zero bit configuration.

619 is RlJ when the mesh is a black point, ROJ when it is a white point.
620 is a bit that distinguishes between black and white, and 620 is a part that describes the code of the topological feature point on the black point by setting bit 629 to Rlj as shown in Fig. 4. be.

Furthermore, 650 is a part that describes the high-order features of the white point obtained by the feature accumulation unit 100, and 651, 652, 6
Reference numerals 53 and 654 describe feature point codes of topological feature points located above, below, to the left, and to the right of the white point, respectively. Although 620 at the black point and 650 at the white point can share and use a common bit string in the address content 610, they are expressed separately in FIG. 7 for convenience. In Figure 5, connection 6
61 to the center scanning line 631 (
or 632, 633, 634), the contents of the mesh points on the control unit. In accordance with the command 901 from 90, the scanning lines (the number of which is added) included in the area 641 (or 642, 643, 644) on both sides of the central scanning line are read out to the register 101 in synchronization with the current reading. ) The contents of the mesh point address above are added from the connection 622 to the current register 10.
2. The contents 610 of the current registers 101 and 102 are transferred to the logic circuits 103 and 104, respectively.
Based on the determination of bits 619 and 629, the mesh point is
When it is found that it is a topological feature point on a character line such as an end point, intersection, branch point, or bend point, commands 951 and 95 are issued.
2, the contents of the current registers 101 and 102 are transferred to the mask circuits 105 and 106. Mask circuit 105,1
In step 06, only the portion 620 corresponding to the feature point code is extracted from the contents 610 transferred from the current registers 101 and 102 and transferred to the past registers 107 and 108. On the other hand, bits 619 and 6 in logic circuits 103 and 104
29, if it is found that the mesh point is a black point other than the topological feature point, and the previous mesh point on the scanning line is a white point, the end point, intersection point,
The feature point code of the bit string 620 indicating that there is no feature point such as a branch point or bending point is determined by the code generator 109, and is transferred to the past registers 107, 108 by commands 953, 954. Further, as a result of content determination of the mesh point 630 on the center scanning line in the logic circuit 103, if the mesh point is found to be a white point, the command 955
and the past register 107 according to the command from the control unit 90.
, 108 is first stored in the area 641 (or 642, 64
3,644) from the past register 10 corresponding to the scan line far from the center scan line 631 (or 632, 633, 634) to the past register 108 corresponding to the scan line close to the center scan line 631 (or 632, 633, 634) The past register 107 corresponding to the center scanning line 631 (or 632, 633, 634) is finally read out, and the comparison circuit 1
Transferred to 10. In the comparison circuit 110, past register 1
The feature point code sequentially transferred from 07 and 108 is compared with the feature point code obtained from the code generator 109, which means that one feature point does not exist. Transfer the contents to temporary storage register 111. When all the contents of the additional past registers have been processed by the comparison circuit 110, the temporary storage register 111 has an area 641 (or 624, 643,
644) and the most central scan line 631 (or 6
32, 633, 634) are stored, and the contents thereof are transferred to the shift circuit 112 by a command 903 from the control unit 90. The shift circuit 112 converts feature point codes corresponding to 620 parts of the contents stored in the temporary storage register 111 into
Shift processing is performed to correspond to any one of portions 651, 652, 653, and 654 according to the scanning direction given by command 904 from control section 90. The output of the shift circuit 112 and the pattern memory contents of the white point 630 currently stored in the register 101 are transferred to the 0R circuit 113.
By compositing at , a result in which feature point codes are accumulated in any part of 651, 652, 653, and 654 determined by the scanning direction is obtained and is currently written to the register 101. Finally, according to a command 905 from the control unit 90, the contents of the current register in which the feature point code has been newly accumulated at any predetermined position of 651, 652, 653, or 654 are transferred to the pattern memory 61 by a connection 663. written to the corresponding address above. Furthermore, the fifth
In the figure, A is an AND circuit, and 0 is an 0R circuit. An example of a higher-order feature obtained on the pattern memory 61 by the operation of the higher-order circuit of the feature accumulation section 100 described above will be explained with reference to FIG. 8, which corresponds to the thinning pattern 700 of FIG. 3. .

A high-order feature having a meaning expressed by an abstract figure 680 is formed on the pattern memory corresponding to a white point in a small area shown at 670 in the figure. That is, 681,682,684
The types are expressed at the end points 701, 704, and 703, respectively, and the minutiae codes that describe these are shown in Fig. 7, 651, 652.
, 654, and 683 is a symbol indicating that there are no feature points in the left direction, and the feature point code describing this is accumulated in the 653 part in FIG. 7, and the higher-order feature code 65
0 is formed. In this embodiment, by simultaneously scanning a plurality of +1 (m=1 in FIG. 8) scanning lines, as shown in the example shown in FIG. Even if the line 671 and the scanning line 672 passing through the end point 703 are not the same, the lines 681 and 682 are
A stable higher-order feature 680 is formed which is shown to have two endpoints of the type shown in FIG. Similarly, an abstract figure 685 illustrates the meaning of a higher-order feature formed on the pattern memory corresponding to a white point within the small area 674. In the figure, 686 represents the type of bending point 710. In the above embodiment, the scanning lines were arranged in four directions: top, bottom, left and right.
It is not limited to the case where all directions are performed, but the four directions are
It is also possible to accumulate only feature points existing in a specific direction among the directions, and conversely, the scanning direction is not limited to the top, bottom, left, and right.

For example, by performing diagonal scanning as shown in 675, a high-level feature having the meaning shown in abstract figure 690 can be obtained in small area 676. Further, in the above embodiment,
The case where feature accumulation is performed using multiple additions + one scanning line is shown, but although the stability of the features is inferior by setting m = 0, it is possible to accumulate only the central scanning line by using a simpler device configuration. It is also possible to perform feature accumulation processing using .

As explained above, according to the feature extraction processing method according to the present invention, features describing a pattern structure can be extracted with a simple device configuration, and identification processing is also simple.
It is consistent with the method of No. 08 (Japanese Patent Application No. 53-10931), and the topological feature points such as end points, intersections, branch points, bending points, etc. on character lines whose descriptive ability was insufficient in the conventional method. Since it is possible to extract high-order features that accumulate the presence of characters and their positional relationships, it has the advantage of being able to extract features that are effective in identifying patterns with topologically complex structures such as Chinese characters.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram conceptually explaining a conventional feature extraction method, FIG. 2 is a configuration of an embodiment of the present invention, and FIG. 4 is an explanatory diagram showing an example of feature points on a pattern. FIG. 4 is an explanatory diagram showing the bit configuration of the pattern memory in FIG. 2 that describes the feature points in FIG. FIG. 6 is an explanatory diagram for explaining the processing in FIG. 5 for the pattern area; FIG. 7 is an explanatory diagram showing the bit configuration of higher-order features on the pattern memory in FIG. 2. , FIG. 8 is an explanatory diagram that graphically illustrates the concept of higher-order features on a pattern area. In the figure, 50 is a scanning line, 60 is a pattern memory section, 61 is a pattern memory, 70 is a feature point extraction section, 80 is an identification section, 90
1 represents a control unit, and 100 represents a feature accumulation unit.

Claims (1)

[Claims] 1. For each point in a pattern area including a black and white binary pattern, topological features found when looking in a predetermined direction from that point are each detected direction information. Find higher-order features by storing them and accumulating them at each point,
As a result, in a character reading method that identifies characters to be recognized using higher-order features obtained at each point on the pattern area, topological features on the character lines of the pattern include end points, branch points,
A means for detecting at least one or a plurality of intersection points and bending points and coding them into corresponding black points, and accumulating feature point codes thus formed at black points on character lines in a predetermined direction. means for forming a higher-order feature on the white point of the white part of the character background, and a feature code that expresses the existence of a topological feature point on the character line, and a topological feature point seen from the white part of the background. Extract the feature code that expresses the positional relationship of the phases. A feature extraction processing method characterized by: 2 Scan the characters on the document and convert the scanning signal to black and white 2
a scanning unit that converts and outputs a value; an M×N pattern memory unit with multiple bits per mesh that stores a pattern area including a binary pattern as part of the pattern area; Expresses higher-order features obtained by accumulating topological features found when looking at a predetermined direction from an arbitrary point at each point while preserving information on the detected direction. a feature extraction unit that forms a feature code on the pattern memory; an identification unit that processes the feature code representing the higher-order feature extracted on the pattern to determine to which category the input character belongs; In the character reading system, which is composed of a control section that generates timing signals necessary for the circuit, the feature extraction section extracts the phase of at least one or more of end points, branch points, intersection points, and bending points with respect to the black points on the pattern. A feature point extraction unit that extracts geometric features and stores feature point codes expressing them in the pattern memory, and a feature point extraction unit that scans the pattern surface in a predetermined direction to express topological features on the black points. and a feature accumulation unit that forms a feature code that expresses a new higher-order feature obtained by associating the feature point code with the direction and accumulating it on the white point on the pattern memory, and the identification unit includes at least the white point Claim 1, characterized in that characters on a document are identified by making a comprehensive judgment using a feature code including the above-mentioned higher-order features.
Feature extraction processing method described in section.