JPH0253193A - Recognizing device for on-line handwriting character - Google Patents

Abstract

PURPOSE:To contrive the miniaturization of a dictionary and to speed up retrieval by obtaining a degree of approximation between an inputted stroke and the basic constituting element of a character, which is defined in advance, and evaluating the character to be inputted from a feature vector, which defines this degree of approximation as a component, with fuzzy. CONSTITUTION:One stroke part of data is supplied through a temporary buffer memory 3 to approximation degree arithmetic circuits 401-426 and the approximation degree with ten plates T1-T26 is calculated. The calculated approximation degree is supplied to a feature vector storing buffer 6. 26 memory areas are obtained in a row direction in correspondence to the ten plate and memory areas are obtained in a column direction in correspondence to the maximum stroke number of the character to be recognized. Consequently, the feature vector of each stroke is stored at every stroke. This feature vector by one character and a character code of a feature dictionary 7 are processed in a detecting circuit 8 and the code number of the inputted character is outputted.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to an online handwritten character recognition device.

[Summary of the invention]

In an online handwritten character recognition device, the present invention calculates the degree of similarity between an input stroke and the basic constituent elements of a character defined in advance, and calculates the degree of similarity between an input stroke and the basic constituent elements of a character defined in advance. By evaluating characters with fuzzy properties, their initial characteristics are improved.

(Prior art) As shown in Fig. 6, as a method for recognizing raw characters online, input handwriting of i, one stroke (A in the same figure) is converted to sample points Pa, Pz, . . . on the handwriting.・, Pn and its time series information are approximated by a polygonal line (B in the figure).

ii. "Stroke recognition" is performed in which the basic shape of the stroke defined in advance, that is, the [basic stroke type] is compared with the polygonal line of the i term.

Based on the results of sections iii and ii, the input stroke is converted to the code number of the nearest basic stroke type.

IV, for all strokes of one character, i to i
Repeat the term.

(2) Refer to the dictionary and determine the character whose stroke type (stroke order) has the code number of item 111 as the input character.

This method is often used.

According to this method, since the input stroke is once replaced with the basic stroke type, most of the information based on the sample point Po = Pn of the input handwriting is discarded except for the data necessary for the subsequent recognition process. is possible,
Therefore, even a device with a small memory capacity can perform character recognition. In addition, by arranging the code numbers of the basic stroke types in the dictionary in stroke classes for each character,
Since the input character can be identified by comparing this with the order of the code numbers of the basic stroke type of the input character, the dictionary can be made smaller and the time required for comparison can be shortened.

Literature: [Nikkei Electronics J December 5, 19B3
[Problem to be solved by the invention] However, when using this method, if the input IE trace changes or is deformed due to input noise, scribe's writing style, etc., errors may occur in the stroke recognition described in Section 11. For example, when writing a "-" or a horizontal bar, if there is a "brush holder" as shown in the O ink pad in Figure 7A, the recognition accuracy will be greatly reduced. The type is recognized as shown in Figure B.

Conventionally, this has been dealt with by writing in the dictionary "stroke type codes that are easy to misrecognize" in parallel, such as "stroke S has code number ct l C2+..." or CnJ.

However, this increases the size of the dictionary and increases the search time.

Furthermore, if the user attempts to add unregistered characters to the dictionary, it is not possible to uniquely determine the stroke type code that needs to be registered from the input strokes, so such additional registration will not be possible. This is extremely difficult to achieve.

This invention attempts to solve the above problems.

[Means to solve the problem]

Therefore, in the present invention, the degree of approximation between an input stroke and the basic constituent elements of a character defined in advance is determined, and the input stroke is expressed in the form of a feature vector characterized by this degree of approximation. The degree of judgment can be changed using modifiers during evaluation.

[Effect]

Character recognition is performed with fuzzy properties, making it possible to reduce the size of the dictionary and speed up searches without reducing the recognition rate.

(Example) First, an outline of the present invention and an example will be explained.

That is, in the present invention, 1. For example, templates (basic stroke types) T1 to T2G as shown in FIG. 2 are prepared.

■For each large input character, approximate the i#th stroke Si and templates 1゛1 to 1゛2ε to each template 'rj (j-1 to 26). Calculate degree Eij.

For example, if the input character is katakana 1me,
The ff11 stroke St is the template TI,
'I' 2, 'l' 3, etc. are more common,
Since the degree of approximation is low with template 1, etc., hloL
-(Capital)%・Etoz”80%・EIQ3-95%
.

..., Etov = 0%r """+ Et2g
-0%. Similarly, for the second stroke s2, Etov-5%, E202-0%11! :203-0%
.

...rk! , 2 o te -95%, ......
1! :22=O% (the numerical value is a tentative size for explanation).

■For each uppercase letter, the result of term ■ is converted into a feature vector Vi Vi = (Etoz + Eioz *
= r Ei2g ).

In the upper case, Vl = (90, 80, 95, =, 0. . . .
0)V2- (5,0,O,...”l 951 ・
..., 0).

IV. If the katakana character ``me'' is written correctly, its first stroke S1 will be approximately 1 in the template 'r1.''
match, the second stroke s2 has the template number ``
Definitely match.

Therefore, in the dictionary, if it is the character ``Me'', it is written as ``JI'' of ``Me''.
Write the character data in the S code, 'rv-approximately,'("7w-definitely.

That is, for each character, write the code number of that character, the number Tj of the template closest to the j-th stroke Si of that character, and its proximity (a modifier indicating the degree of coincidence). Also, this template number Tj
and modifiers are written in stroke order for the number of strokes and strokes.

Further, this character data is divided into sections according to the total number of strokes of the character.

(8) For example, prepare a table of fuzzy functions as shown in FIG.

Among the character data based on the Vl and IV terms, the character data for the first character is extracted from the term of the total number of strokes of the input character.

In the upper case, the total number of strokes for the character "Me" is two strokes, so the first character data is extracted from the term 21ji.

■For the sake of simplicity, let us assume that the character data extracted by vtLJI is character data for the character "Me".For the first stroke s1, i"a = "approximately", so "approximately" in Figure 3. ” and select the function curve m
Among the degrees of approximation of feature vector (1) obtained in section 1, the degree of approximation of 5% (=hto3) to template r3 is extracted.

Then, this degree of approximation of 95% is converted into pass *Ct, for example, G1-%%, using the "approximate" function curve shown in FIG.

Similarly, for the second stroke S2, since Tv - "definitely", the function curve "definitely" in FIG. % (=E 20 ma) is taken out and converted to a passing grade G2, for example, G2-98%.

That is, according to the character data extracted in the Vl term, the function curve shown in FIG. 4 is selected for each stroke S, and the corresponding approximate value Eij of the feature vector Vt is selected.
, by the selected function curve (11 decorated and passing grade G
Convert to i.

Among the pass rates Gi obtained in Section 1.4, the minimum pass rate is
It is assumed that the pass level Gs of the character of the code number indicated by the character data.

In the upper case, G1-96% and G2-98%, so the pass level G of the input character for the character "Me" is 96% ("Gi").

(2) Thereafter, execute Section 1.4 for the character data with the corresponding total number of strokes.

When the X and ■ items are completed, M+16 of the obtained pass score G+m (this is obtained by the number of character data)
The code number of the character giving the pass grade G is output as the first candidate for the input character.

Next, T 1 to T2G (Tembray) are determined in consideration of the following points. The constituent elements of Sunarichi, i, and kanji include ``recon stroke'', ``vertical stroke'', ``harai'', and ``bend'', but the types are limited.

:1. Even strokes that look the same may have different forms of deformation, such as ``splashes'' and ``curves,'' due to differences in brushstrokes.Also, since ``splashes'' and the like do not originally have a specified length or angle, , for strokes that may cause different deformations, separate templates are prepared for each type of deformation, such as templates 1'1 to 'I'3.

iii. Particularly complex basic shapes appear very infrequently in all kanji characters, so they are processed using another recognition method without defining a template.

Note that in the template T 1 ″ T 2 @, the stroke portion indicated by a broken line indicates that the evaluation is small or ignored when determining the degree of approximation Eij.

Next, an example of this invention will be explained.

In FIG. 1, +11 indicates a coordinate input means such as a tablet, and this input means (1) inputs a coordinate string Pa-Pn for one stroke, and this coordinate string Po"Pn is input to the polygonal line pressure type circuit ( 2) and is compressed and converted into a string of polygonal line information and its end point information.

That is, for example, if the input stroke (coordinate string) is preprocessed and consists of polygonal lines #1 to #4 as shown in FIG. 4, it is quantized in 8 directions with an angular range (direction) as shown in A in the same figure, and is converted into the length of the polygonal lines #1 to #4 and the coordinate values of each starting point and ending point, The data is as shown in C of the same figure.

Then, one stroke of this data is sent to the approximation calculation circuit (401) through the -time buffer memo (January 3).
~(426) 2 supplied template 'r1~'!
' The degree of approximation Eij to 26 is calculated (the u term above).

Note that the calculation of this degree of approximation Eij is performed by the rule storage circuit (5
01) to (526) are executed independently and in parallel for each template l''j.

Then, the calculated degree of approximation Eij is supplied to the feature vector storage buffer (6).
Template T1 schematically shows the structure of
~ It has 26 memory areas in the row direction corresponding to T211, and the maximum number of strokes k of characters to be recognized.
It has memory areas in the column direction corresponding to . Therefore, the feature vector Vi of each stroke St of one character is stored in this buffer (6) for each stroke Si (section 2 above).

Then, the feature vector Vi for this one character and the feature dictionary (
The character code from 7) (section IV above) is the character code from the verification circuit (
In 8), the code number of the input character is processed as in items 1 to 1 above and output.

FIG. 5 shows the degree of approximation Ejo of the input stroke Si to the template T1 in the degree of approximation calculating circuit (401).
An example of a rule for calculating t is shown below.

In other words, figure A includes "right", "five", etc.
This is an expanded view of the stroke ``ノ'', but this stroke ``
In contrast, when the template is 1, B in the same figure
As shown in , lengths 1 to L41Lh and LH are measured and Eiot −(aLh −bLw−cLt +dL4+
eL3)/L years However, when E;at>1, Etot-IEto
When t<0, E tot -Oa -e is calculated as a constant (here, the approximate value Eio1 is expressed as a decimal)
.

is calculated.

At this time, in template 1'1, the weight of the stroke portion of the broken line is set to 0 or small, so the value L3. Constants e and d for L4 are other constants a''
- It is made smaller compared to C.

Further, in the arithmetic circuits (402) to (426), arithmetic expressions for the degrees of approximation Eio2 to Ei2ε are defined corresponding to the templates 1'2 to '1' 2G, respectively, and the degrees of approximation are calculated.

(Effects of the Invention) As described above, according to the present invention, online handwritten character recognition is performed. In this case, in particular, according to the present invention, the input stroke Si and a predefined template are Since the degree of approximation Eij between 'r1~+l'l, , , and character recognition is performed based on this degree of approximation Eij and the modifier, there is no reduction in the recognition rate due to natural changes or deformations of handwriting; , template 'r j is template) T L to T3, the recognition rate is improved and the recognition rate is strong against changes or deformations of handwriting.

In addition, the dictionary (7) basically only needs to have one set of representative templates i' j corresponding to the stroke Si and its modifiers, so the dictionary (7) can be made smaller, and Searching the dictionary (7) can be speeded up.

Furthermore, when registering undefined characters, each time the user inputs one stroke, the shape of the template with the highest degree of approximation is displayed on the screen, and R4 confirms whether the template is the correct shape in a conversational manner. This can be achieved in this way.

[Brief explanation of the drawing]

FIG. 1 is a system diagram of an example of the present invention, and FIGS. 2 to 7 are diagrams for explaining the same. (1) is a coordinate input means, (401) to (426) are approximation calculation circuits, (7) is a feature dictionary, and (8) is a verification circuit.

Claims (1)

[Claims] A plurality of templates in which basic constituent elements of characters are defined in advance, a circuit that calculates the degree of similarity between an input stroke and the plurality of templates, and a circuit that calculates the degree of similarity between an input stroke and the plurality of templates, and An online handwritten character recognition device comprising: a feature dictionary having a modifier; and a verification circuit that modifies the degree of approximation with the modifier to change and test the degree of character determination during evaluation.