JPS6046753B2 - Pattern feature extraction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a feature extraction method for an optical character reading device (hereinafter referred to as OCR), and particularly to a method for extracting features that distinguish between concavities and straightness of a character line connected to a branch point.

When recognizing characters written on paper, it is generally optically scanned and quantized into a binary pattern such as ``1'' for the characters and ``o'' for the white part of the paper. After separation, the signals undergo preprocessing such as noise removal and thinning, and then pass through a feature extraction section and a determination section.

The feature extraction section usually extracts line end points, branch points, bending points, depressions, convexities, etc. as features, and the judgment section judges input characters based on the presence or absence or quantity of these features. It's on. Regarding this feature extraction, various feature extraction methods have been proposed and put into practical use. However, when the number of characters to be read is expanded to include alphabets and katakana in addition to numbers, similar patterns in different categories increase. In order to distinguish patterns that are similar to each other, it becomes necessary to extract special features for distinguishing only these similar patterns. To do this, we first extract features that are relatively effective for identifying the entire character to be read (hereinafter referred to as basic features), determine candidate characters based on the extracted basic features, and then A list of multiple characteristics described in (hereinafter referred to as the basic list)
Accordingly, a method of extracting features from basic features or input patterns and narrowing down to one candidate character that satisfies the feature list has been adopted. An example of the distinction that is difficult to make unless determined by the above means is the distinction 8.B.

In order to read the number 8 as shown in Figure 1, it is prerequisite that you have a nose that can distinguish between Figures 1 and 2 in Figure 1. FIG. 12 shows the alphabetic letter B. With OCR that only reads numbers, there is no problem even if both 1 and 2 are recognized as the number 8, but with OCR that reads alphanumeric characters
We need to make a precise distinction. The basic feature of the pattern shown in Figure 1 is that there are no endpoints and there are two branching points, so it is assumed that either pattern 1 or 2 is the candidate character 8,
After B is determined and narrowed down to two candidate characters, the large indentation on the left side of the pattern indicates that it is 1, and the small indentation indicates that it is 2. However, conventionally it has been difficult to detect the extent of the indentation in a character pattern, and there is no easy method for determining this. SUMMARY OF THE INVENTION An object of the present invention is to provide a method for extracting features that are effective for distinguishing character patterns in which the difference is the size of a depression near a branch point. The present invention provides a pattern memory that thins and stores a binary pattern for one character, and a stroke end point, branch point, etc. from the pattern of this pattern memory.
A basic feature extraction means that extracts basic features such as the direction of the stroke, a dictionary for determining candidates that stores a large number of standard patterns with similar formats to the basic features, and a matching method that matches the standard patterns with the basic features of input characters. candidate determining means for extracting character names of a standard pattern as candidate character names; a judgment dictionary that stores one or more feature lists necessary for recognizing each character; and a judgment dictionary that stores the feature list of candidate character names. The last component of 0CR consists of a determination means that extracts the features of the extracted feature list from the input pattern or basic features, and when all the features in the feature list are extracted, determines the candidate character name as the character name of the input pattern. The determination means is selected from the feature list extracted from the determination dictionary.
One feature is a means for specifying one of the upper, lower, left, and right directions in the input pattern and extracting a branch point in the specified direction and one side of the parallelogram including the input pattern; means for determining, on the side, positions spaced from the branch point by a specified distance on both sides in a direction perpendicular to the direction; means for detecting two distances before intersecting the linear pattern for the first time;
The feature extraction means includes means for verifying the length relationship between the two distances obtained by the means and the distance from the side to the branch point.

1 to 4 in FIG. 2 indicate the directions for extracting branch points and determining the edge that is the starting point of scanning. For example, if direction 4 is specified, the leftmost branch point is selected as shown in FIG. The starting point of scanning is on the left side 100 of the quadrilateral surrounding the thinning pattern.

The same applies to directions 1 to 3. As a method of feature extraction,
As shown in FIG. 3, distance d in direction 4 and two directions orthogonal to direction 4, that is, both directions 1 and 3, from the branch point that takes the maximum coordinate value in direction 4 among the branch points.
By specifying, the branch point B1 on the left side is selected from among the branch points Bl and B2, and the starting points a1 and A2 of the scan are determined on the left side by the coordinate values of the branch point B1 and the distance d, and the starting point of the scan is Scan horizontally from points a1 and A2, that is, parallel to the specified direction, and find the distances 11 and 12 until they intersect the thinning pattern for the first time, and calculate the distances 10 and 12 from the top of the left side to the branch point B1. and the distances 11 and 12, a feature value 1 representing the degree of the dent is determined. -Q') 10\&1I12No Figure 8 is a diagram explaining the relationship between the surrounding state of a branch point and the feature value. , c are when the area around the branch point B is concave when viewed from the reference line position 10 to the right.

Distance 1. , 11, and 12 have the following value relationship. In Figure A, in Figures B and C, the above-mentioned characteristic value 1 is a value close to O if the area around the branch point is close to a straight line, and is a positive value if it is in a concave condition. becomes.

Also, as you can easily see, the more depressed the
In the distance relationship described above, the difference between the value on the left side and the value on the right side becomes large, so the feature value becomes a large positive value. Using the feature value 1, in FIG. 1 1 is assigned when l is greater than or equal to the threshold t1 (for example, 6), and in FIG. 1 2 is assigned when l is less than or equal to the value T2 (for example, 4).

Here, t1>T2. FIG. 4 is a block diagram showing an embodiment of an OCR device using the pattern feature extraction method of the present invention.

Reference numeral 10 denotes a pattern memory for storing one character's worth of patterns, in which patterns that have undergone processing such as OCR photoelectric conversion and cutting are thinned and stored.

The basic feature extraction means 11 extracts basic features such as full marks, branching points, inflection points, and stroke directions from the character pattern stored in the pattern memory 10, and
Store in 2. The basic features obtained here are features that can be commonly used for all characters, and there are few that will be unnecessary in later processing, so the basic feature extraction means 11
In this method, uniform features are extracted for all input characters. When the basic features are extracted, the control is started by the candidate character determining means 1.
4, the basic features stored in the basic feature register 12 are compared with the standard patterns stored in the candidate determination dictionary 13. Here, one or more standard patterns are prepared in the candidate determination dictionary 13 for each character to be recognized. The candidate determining means 14 performs a comparison with the standard pattern of the candidate determining dictionary 13, and the register 12
The character names of standard patterns that include the basic features in the candidate character name register 15 are stored in the candidate character name register 15. In this register 15, one or more candidate character names may be set,
Sometimes it doesn't work out. Once the candidate character name is determined based on the basic characteristics, control is transferred to the determining means 17 in response to a control transfer signal from the candidate character determining means 14. In the determination means 17, if there is no candidate character stored in the candidate character name register 15, a signal indicating that it is unreadable is output to the output register 19, but if there is one or more candidate characters, the character name is output to the keypad. The contents of the judgment dictionary 18 are indexed as , the necessary features are extracted from the pattern memory 10 and the basic feature register 12 according to the feature list taken out from there, and character names that satisfy the feature list are output as input character names. register 1
Output to 9. Each item in the judgment dictionary 18 expresses the type and parameter of a feature, and when all of a plurality of features need to be satisfied, they are listed together.

When the features to be used cannot be extracted from the basic features, a feature extraction method that extracts them from the pattern is also required.
One of them is the present invention. FIG. 5 is a block diagram showing an embodiment in which the pattern feature extraction method of the present invention is implemented in the determination means 17. The determination means 17 includes several feature extraction means 40 under the control of the determination control circuit 20.
, 50, . Note that some parts common to FIG. 4 are omitted from illustration. The feature list corresponding to the candidate character set in the candidate register 15 is decoded by the determination control circuit 20, and the feature A, that is, A9m9d9t9 is selected as one of the feature lists.
r However, when m = 19゛゜゜゜4 (T and r will be explained later), a branch point from the basic feature register 12 is selected by the parameter m corresponding to the direction in Fig. 2, and its coordinate value is (corresponding to the parameter m, x coordinate value when m = 1, 3, y coordinate value when m = 2, 4) is set in the register 21, and d is set in the register 22 as a parameter representing the distance from the branch point. Set.

Therefore, specify one of the top, bottom, left, and right directions in the input pattern mentioned above, and extract the coordinate value of the branch point from one side of the parallelogram that includes the input pattern determined by the specified direction. The means selects a branch point from the basic feature register 12 under the control of the determination control circuit 20 under the condition of the parameter m, that is, the direction shown in FIG. This corresponds to the part that calculates the X or Y coordinate value and sets it in the register 21. Further, a parameter m representing the direction of scanning from the designated side (vertical scanning when m=1, 3, horizontal scanning when m=2, 4) is set in the register 23. The coordinate value of the branch point in the register 21 and the value of the parameter d representing the distance from the branch point set in the register 22 are determined by the addition circuit 26 and the subtraction circuit 27.
They are added and subtracted, respectively, and set in the register 28.
The register 21 is a means for finding a position on the side that is separated from the branch point by a specified distance in a specified direction and perpendicular direction.
This corresponds to the part in which addition and subtraction between the value of and the distance value stored in the register 22 is obtained by the addition circuit 26 and the subtraction circuit 27, and the results are set in the register 28 as Al and a2, respectively. Also, two distances 11, 1 from each position found on the side to the point where the pattern first intersects J by scanning parallel to the specified direction. The means for detecting this corresponds to the following intersection detection means 29. The intersection detection detecting means 29 detects the scanning starting point set in the register 28 and the register 23.
Corresponding to the circuits 26 and 27 that set a value in the register 28, the distance from when the pattern memory 10 is scanned in the scanning direction set to the point where the pattern memory 10 first intersects the pattern,
Set in registers 31 and 30. Next, the values of registers 30 and 31 are added in an adder circuit 32, and a feature calculation circuit 33
Now, based on the above addition value and the coordinate values of the branch point, the length relationship L
Calculate and output. Here, as shown in FIG. 6, the pattern memory area is in column I and column J, and the coordinate values of branch point B are (1,
j), if the above-mentioned addition value is K, the length relationship L is as follows depending on the parameter m.

The verification circuit 34 outputs "1" if the output value L satisfies the convex or concave threshold t set in the register 24 and the convex or concave size relationship r set in the register 25, and "1" otherwise. 0'' is output to the determination control circuit 20.

The two distances 11 and 12 obtained and the distance 1 to the branch point.
The means for verifying the length relationship with
2), the addition circuit 32 and the feature calculation circuit 33 calculate the feature value [1=2×10−(11+12)], and the verification circuit 3
4 corresponds to a portion that determines the presence or absence of unevenness based on the output value of the feature calculation circuit 33 and the parameter values (threshold values) reset in the registers 24 and 25 by the determination control circuit 20. If the parameter r representing the magnitude relationship of feature A is less than or equal to r = 0 and greater than or equal to r = 1, the list for determining FIG. 1 is, for example, 8 from the determination dictionary 18,
In order for the left-hand dent to be 6 or more within ±5 meshes above and below the branching point, that is, to be B, the left-hand dent must be 4 or less within ±5 meshes above and below the branching point.

In the above explanation, the features were determined by fixing the distance d to a constant value, but the influence of the font size was removed. It is effective to replace the distance d with a ratio to the height of the character.

FIG. 7 shows this embodiment.

Here, the ratio to the height of the character is set in the register 35, and the circuit 70 detects the height or width based on the length of the pattern on the pattern memory 10 from the top end to the bottom end or from the left end to the right end. The distance calculation circuit 36 calculates a designated distance from the output value of the register 35 in which the height or width of the pattern from the circuit 70 and a preset ratio thereof are set, and sets it in the register 22.

Other functions are the same as those described in the embodiment of FIG.

[Brief explanation of drawings]

FIG. 1 is a diagram showing examples of patterns that can be distinguished by applying the present invention, FIG. 2 is a diagram related to branch point extraction and scanning direction, FIG. 3 is a diagram related to feature extraction, and FIGS. 5 and 7 are block diagrams showing embodiments of the present invention, FIG. 6 is a diagram regarding the area of the pattern memory and the coordinates of the branch point, and FIG. 8 is a diagram explaining the relationship between the surrounding state of the branch point and the characteristic value. There are plans to do so. In the figure, 10 is a pattern memory, 11 is a basic feature extraction means, 12 is a basic feature register, 13 is a dictionary for candidate determination, 14 is a candidate determination means, 15 is a candidate character name register,
17 is a judgment means, 18 is a judgment dictionary, 19 is an output register, 20 is a judgment control circuit, 29 is an intersection detection means, 33
is a feature calculation circuit, 34 is a verification circuit, and 36 is a distance calculation circuit.

Claims (1)

[Claims] 1. Means for specifying one direction among the top, bottom, left, and right in an input pattern and extracting the coordinate value of a branch point from one side of a parallelogram including the input pattern determined by the specified direction; means for determining a position on the side that is spaced apart from the branch point by a specified distance on both sides in a direction perpendicular to the direction;
means for detecting two distances from each position found on the side until it intersects the pattern for the first time by scanning parallel to the specified direction; and two distances obtained by the means. and a means for verifying a length relationship between the length and the distance from the side to the branch point, and detects a feature value representing the degree of dent.