JPH0246988B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition device that recognizes characters with many strokes, and more particularly to a character recognition device that recognizes Chinese characters.

Conventionally, when recognizing Chinese characters, particularly in devices that recognize printed Chinese characters, a pattern matching method or a similar method has been used. These methods were effective in the case of printed kanji and extremely carefully written block script.

However, as shown in FIG. 1, when the input character pattern 3 is tilted with respect to the reference character pattern 2 in the entry frame 1, the degree of similarity between the two becomes small. Furthermore, as shown in FIG. 2, when the input character pattern 5 has a different line width from the reference character pattern 4, the degree of similarity between the two also decreases. Therefore, even if the pattern matching method is applied to characters such as handwritten kanji where each stroke is tilted with respect to the standard character pattern or the line width is not constant, a high recognition rate cannot be obtained. It was hot.

In order to eliminate these drawbacks, this invention extracts the direction code of the contour line, feature points such as end points, branch points, and bending points, and strokes from the input character pattern,
The present invention is characterized in that recognition is performed by combining these special features, and its purpose is to realize a character recognition device with high recognition accuracy.

This will be explained in detail below using the drawings.

FIG. 3 is a block diagram of an embodiment of the present invention. First, the input characters on the form 6 are scanned by the scanning means 7,
The character pattern is stored in the pattern memory 8. The direction code extracting means 9 extracts the direction code distribution of the contour line, ie, the number of contour points having the same direction code, for each of several divided regions of the character pattern stored in the pattern memory 8. Here, the direction code is a code that encodes the direction from a contour point of interest to an adjacent contour point. Next, thinning means 1
0 thins the character pattern stored in the pattern memory 8 and stores it in the thinning pattern memory 11. The feature point extracting means 12 extracts the number of feature points such as end points, branch points, and inflection points for each of several divided regions of the thinning pattern stored in the thinning pattern memory 11. Moreover, the stroke extraction means 13
extracts a stroke from the thinning pattern stored in the thinning pattern memory 11, and determines its direction,
Find the length and midpoint position. The determining means 14 compares the direction code distribution, the feature point distribution, and the stroke midpoint distribution with the direction code distribution, feature point distribution, and stroke midpoint distribution of the reference character stored in advance in the determining means. The degree of similarity is determined, and the character name of the input character is determined using these degrees of similarity.

FIG. 4 is a diagram showing the direction code of the outline obtained by the direction code extracting means 9 from the input character pattern of the Chinese character "kin". Here, the symbol "H",
"V", "R", "L" are horizontal, vertical, upper right, respectively.
Represents the upper left direction code. Next, divide the character pattern into areas of n ₁ × n ₂ (n ₁ and n ₂ are integers),
For each divided area, the frequencies of the direction codes of the four reward directions (horizontal, vertical, upward right, upward left) are determined and used as a direction code distribution. That is, the direction code distribution 〓 _d is m
If =4n ₁ ×n ₂ , then it is expressed as 〓 _d = (d ₁ , d ₂ , ..., d _n ). Here, d _i is the frequency of the direction code in a specific direction in each region.

FIG. 5 shows an example in which the input character pattern of the Chinese character "kin" is thinned by the thinning means 10, and the character portion 15 of the thinning pattern is represented by "1".

FIG. 6 is a diagram showing feature points such as end points, branch points, and bending points obtained from the thinning pattern by the feature point extraction means 12, and strokes obtained by the stroke extraction means 13. Here, each feature point is marked with "*", and each stroke is marked with "A"~
Indicated by "H". From this, the feature point distribution and stroke midpoint distribution are determined. The feature point distribution is obtained by calculating the number of feature points for each region divided into n ₃ ×n ₄ (n ₃ and n ₄ are integers) of the thinning pattern. That is, the feature point distribution _f is expressed as _f = (f ₁ , f ₂ , . . . f _l ) where l=n ₃ ×n ₄ . Here f _i is the feature point of each area.

The midpoint distribution of strokes is the position of midpoints of strokes determined for each stroke direction.
The direction of the stroke is horizontal, vertical, upward to the right,
Use the four types shown on the top left. FIG. 7 shows a midpoint pattern of horizontal strokes.
The midpoints of the strokes are marked with symbols corresponding to the stroke symbols "C", "E", and "H" in FIG.
The midpoint patterns of vertical, upward right, and upward left strokes are found in the same way.

The determining means 14 determines the direction code distribution, feature point distribution, and stroke midpoint distribution of the input character, and the direction code distribution, feature point distribution, and stroke midpoint distribution of the reference character C _i stored in the determining means 14. The similarity of each is calculated. The similarity of these three
Let Sd(i), S _f (i), and S _c (i).

Various definitions of S _d (i), S _f (i), and S _c (i) can be considered, but for example, they can be determined as follows.

S _d (i) is the direction code distribution of the input character and the reference character, respectively 〓 _d = (d ₁ , d ₂ , ..., d _n ), 〓 _d (i) = (d ₁ ′,
d ₂ ′,…,d _n ′), it is defined as S _d (i)=(〓 _d , 〓 _d (i))/‖〓 _d ‖・‖〓 _d (i)‖.

_{S f} (i) _{is the} feature _point distribution _{of the} input character and _the reference character, respectively:
...,
f _l ′), it is defined as S _f (i)=(〓 _f , 〓 _f (i))/‖〓 _f ‖・‖〓 _f (i)‖.

S _c (i) is obtained by associating strokes using the direction-specific midpoint pattern of strokes of the input character and the direction-specific midpoint pattern of strokes of the reference character. Specifically, first, the closest point among the midpoints of the stroke of the reference character is associated with the midpoint of the stroke of the input character. Next, the distances between the two points correlated above are added, and the reciprocal thereof is set as S _c (i).

Finally, the overall similarity S(i) is calculated from S _d (i), S _f (i), and S _c (i).
, and determine the character name of the input character. That is, W _d ,
With W _f and W _c as weights for each feature, the overall similarity expressed as S(i) = W _d S _d (i) + W _f S _f (i) + W _c S _c (i) has the largest value. Determine the reference character as the recognized character.

In this way, recognition is performed using the direction code distribution, feature point distribution, and stroke midpoint distribution, so
Higher recognition accuracy can be obtained compared to conventional methods.

Although the above description has been made regarding the case of recognizing handwritten kanji characters, the present invention may also be used to recognize characters with many straight lines, such as handwritten katakana characters.

In addition, as a method for determining the determining means, we have explained the case where the characteristics of the direction code distribution, feature point distribution, and stroke midpoint distribution are used simultaneously, but this is not limited to this. It's okay. For example, after performing major classification using direction code distribution, recognition may be performed using feature point distribution and stroke midpoint distribution. By doing so, processing time can be shortened.

As described above, according to the present invention, the direction code,
Since recognition is performed using the distribution of feature points and midpoints of strokes, it is stable even with small changes in the slope of line segments and changes in line width, and it is possible to use a combination of three types of features with different properties. The advantage of this method is that it can recognize handwritten characters with high accuracy by taking advantage of the strengths of each feature. Furthermore, among the above three types of features, the feature of stroke midpoint distribution has the highest discrimination ability and is essential. However, the number of characters to be recognized is small,
If there are no sets of characters that are similar to each other, sufficient discrimination ability can be obtained even if the use of either the feature point distribution or the direction code distribution is excluded. Moreover, the processing time is shortened, and the cost can be reduced by reducing the capacity of the recognition dictionary. Therefore, under the above conditions, it is not possible to configure a character recognition device that uses only the stroke midpoint distribution and direction code distribution, or a character recognition device that uses only the stroke midpoint distribution and feature point distribution. Can be done.

[Brief explanation of drawings]

Figures 1 and 2 are diagrams showing examples of variations in input character patterns, Figure 3 is a block diagram of a device according to the present invention, Figure 4 is an example of an input character pattern from which a direction code has been extracted, and Figure 5 is a thin line FIG. 6 is an example of an input character pattern from which feature points and strokes have been extracted, and FIG. 7 is an example of a stroke midpoint pattern of input characters in the horizontal direction. In the figure, 6 is a form, 7 is a scanning means, 8 is a pattern memory, 9 is a direction code extraction means, 10 is a thinning means, 11 is a thinning pattern memory, 12 is a feature point extraction means, 13 is a stroke extraction means, 14 is a determining means. In the drawings, the same or corresponding parts are denoted by the same reference numerals.

Claims (1)

[Claims] 1 A character recognition device that recognizes characters written or printed on a recording medium such as a form includes a scanning means that scans and photoelectrically converts characters, and extracts a direction code of an outline from an input character pattern obtained by the scanning means. a direction code extracting means for thinning an input character pattern to obtain a thinning pattern; a feature point extracting means for extracting feature points such as end points, branch points, and bending points from the thinning pattern; a stroke extracting means for extracting a stroke that is a straight line segment from a pattern, and a determining means for determining a character name of an input character using the direction code, feature points, and stroke; , Extract the direction code distribution of the contour line for each region divided into multiple parts of the input character pattern,
The feature point extraction means extracts a feature point distribution for each region divided into a plurality of areas of the thinning pattern, the stroke extraction means extracts a midpoint distribution of the stroke, and the determination means extracts a distribution of midpoints in the direction Calculating the degree of similarity between the code distribution, the feature point distribution, the midpoint distribution of the stroke, and the direction code distribution, feature point distribution, and stroke midpoint distribution of a reference character stored in advance in the determining means, respectively; Characters are recognized using the similarity between the above direction code distribution and stroke midpoint distribution, or the above feature point distribution and stroke midpoint distribution, or the above direction code distribution, feature point distribution, and stroke midpoint distribution. Characteristic character recognition device.