JPS58163080A - Feature extracting system of character - Google Patents

Abstract

PURPOSE:To recognize the characters having many curves like HIRAGANA (Japanses syllabary), etc. with high accuracy and no error, by using a means which extracts the rectilinearity of strokes and then decides the number of key points in response at least to said rectilinearity. CONSTITUTION:The character information supplied from a tablet 1 is fed to a rectilinearity measuring circuit 11 in the form of a coordinate point series. The circuit 11 calculates the rectilinearity Cs, i.e., the ratio between the total distance of stroke and the rectilinear distance. This rectilinearity is supplied to a rectilinearity/representative point number converting table 12. Then the number of representative points corresponding to the rectilinearity and increasing toward a curve is fed to feature extracting circuits 13 and 14. The circuits 13 and 14 extract the feature series an input data and a data read out of a dictionary 4 in response to each number of representative points respectively. A matching circuit 3 compares the feature series sent from the circuits 13 and 14 with each other to obtain a distance.

Description

[Detailed Description of the Invention] (Technical Field of the Invention) The present invention relates to a real-time handwritten character recognition device, and in particular to a feature extraction method that can approximate characters such as hiragana, which have many curved parts, to a high degree of #1. be.

(2) Conventional technology and problems For example, Odaka is a conventional feature extraction method.

Arayo, Danta 2 "Online recognition of handwritten characters using stroke approximation points", Proceedings of the Institute of Electronics and Communication Engineers 80/2 V
As was discovered in OL J65D 應2, the number of representative points is fixedly determined by the number of strokes, regardless of the length of the character's (1) stroke or the length of the song. For this reason, the configuration shown in FIG. 2 has been greatly simplified, and the input data from the tablet 1 is sent to the % fine extraction circuit 2 in which the number of representative points is fixed by the number of strokes, and the features are sent to the matching circuit B, which reads them from the dictionary 4. The candidate character is compared with the candidate character, the distance is taken, and a determination is made. However, with this method, as shown in Figure 1, when the character "wa" ■ with a curved stroke is input, the candidate characters read from the dictionary are "wa" ■ and "re".
■, and the part ■ where the distance between ■ and ■ appears is α of ■ and A of ■
The distance between ■ and ■ is the distance between ■ and ■.

In this way, the distance calculated from a straight line part and the distance calculated from a curved part are treated equally, so gd&
rate becomes worse.

(3) Purpose of the Invention The purpose of the present invention is to provide a feature extraction method that can approximate characters with many curved parts with high accuracy.

(4) Structure of the Invention (2) In order to achieve the above object, the character feature extraction method of the present invention is based on the real-time handwritten character g! *In the device, when extracting features of characters with many curved parts, a means for measuring the straightness of the stroke, and a means for determining the number of representative points to be extracted according to at least the straightness measured by the means. It is characterized by having the following.

(5) Examples of the invention Handwritten characters, especially hiragana, have six uras.

Identification information is mainly contained in curved parts, such as 1.Warene'. On the other hand, in the past, straight line portions and curved portions were approximated using exactly the same number of representative points. but,
In this case, the % mark changes due to a change in the 11" position of 6" which is not essential [or the 61st position of 1"], and the curve part that is originally effective for identification changes. It is difficult to see the difference.

Therefore, the principle of the present invention is to calculate the linearity of the stroke before extracting representative points, and strokes that are close to a straight line will have a small number of points, and strokes that are close to a curve will have a large number of points that are close to 1.
By doing this, weight is given to the curved stroke (6) loaf, and characters with many curved parts, such as hiragana, are approximated with a high degree of curvature, thereby increasing the 4-line rate.

45 (α) and (6) are explanatory diagrams of an embodiment according to the above-mentioned principle, FIG. 45 (α) is a configuration diagram, and FIG. 45 (6) is a detailed explanatory diagram of the main part.

In the same figure (α), the character 11v input from the tablet 1 is input to the straightness measuring circuit 11 as a coordinate point series. Here, linearity is calculated using the following method.

That is, as shown in FIG. 3B, if vectors PI, P2, . . . , Pn are representative points on the stroke S of the curve, then the straightness C is defined as follows. s)l=
- If the total distance of the strip S is , then the straightness C1 can be calculated by (1) and (2). C1 has a real number of 0≦C, ≦1 and is close to 1 (4) is close to a straight line.

An example of the linearity C1 and the number of representative points is shown in the following table.

Returning to table (8) in the same figure, 9, the linearity measurement circuit 11 calculates the linearity C, using the method described above, and the value is sent to the If linearity/representative point conversion table 12 that stores the contents of the above table. becomes the input. Then, both feature extraction circuits 1 calculate the number of representative points according to the linearity.
Send to 6 and 14. The feature extraction circuit 13 extracts a feature series from the input data from the tablet 1 according to the number of representative points from the conversion table 12. Also, the feature extraction circuit 14 is
ii A feature series is extracted from the data successively generated from F4 according to the number of representative points from the conversion table 12. The matching circuit 3 compares the feature sequences sent from the feature extraction circuits 15 (5) and 14, and sends out the category with the smallest distance as the answer. To find the distance, use the human data feature series “” (2X1y Z2+...
π1) Dictionary feature series Xj=(z!jr
If W2j+...ff1J, distance dj=Σ1 town-
Town j1 11+1.

FIG. 4 is an explanatory diagram showing the configuration of another embodiment of the present invention.

In the embodiment shown in FIG. 46, only the linearity of the stroke is used as a parameter for the number of representative points, but in this embodiment, the parameter is a combination of the length of the stroke.

In the same figure, a representative point number determining circuit 16 is provided in place of the straightness/representative point number conversion table 12, and a new stroke length measuring circuit 15 is added to input data in parallel with the straightness measuring circuit 11.
The number of representative points is determined by inputting the stroke length together with the linearity to the representative point number determining circuit 16. The effect is similar to the previous embodiment.

(6) As described in detail, according to the present invention, when extracting the features of characters with many curved parts, such as hiragana, the straightness of the stroke is extracted, and at least the number of representative points is determined according to this straightness. By providing a means to determine the characters (with high precision)
It becomes possible to recognize the information without any problems.

[Brief explanation of drawings]

Fig. 1 is an explanatory diagram of the problems of the conventional example, Fig. 2 is an explanatory diagram of the configuration of the conventional example, and Figs. 6 (α) and (b) are an explanatory diagram of the configuration of the embodiment of the present invention and detailed explanation of the main parts 4 are explanatory diagrams of other embodiments of the present invention, in which 1 is a tablet, 3 is a matching circuit, 4 is a dictionary, 11 is a linearity measurement circuit, and 12 is a linearity/representative point conversion table. , 15 and 14 are feature extraction circuits, 15 is a stroke length measurement circuit, and 16 is a representative point number determination circuit. Patent applicant Fujitsu Ltd. sub-agent Patent attorney 1) Yoshishige Saka (7) Figure 1 Figure 2 Figure 3 (a) Figure 4

Claims (1)

[Claims] In a real-time handwritten character recognition device, when extracting features of a character with many curved parts, there is provided a means for measuring the straightness of a stroke, and at least a number of representative points to be extracted according to the straightness measured by the means. 1. A method for extracting a character source, characterized in that it includes a means for determining.