JPS6186881A - Recording system for on-line handwritten character - Google Patents

Abstract

PURPOSE:To shorten the calculating time of the resemblance between patterns with correct correspondence of strokes secured between an input character and a standard pattern even in case the input character has an error of the stroke making order, by defining the sum total of inter-stroke resemblance degrees showing each maximum value as an inter-pattern resemblance degree. CONSTITUTION:The character information equivalent to a character written on a tablet by an input pen is fetched every fixed time point and separated into the writing point coordinate series for each stroke to recognize it after the pre-processing like the normalization, the decision of a size, etc. A feature extraction processor 43 extracts the features based on the pre-processed character information. The extracted input character information is expressed in a multi-dimensional vector space containing each stroke as an element. In a discrimination processing mode, the inter-pattern resemblance degree rhoC is obtained to a standard pattern to be collated based on the input character information. Then a character code PC of the standard pattern which gives the maximum inter-pattern resemblance degree is delivered. The inter-pattern resemblance degree is obtained based on the inter-stroke resemblance degree between each stroke of an input character and that of the standard pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to an online handwritten character recognition system in which the ability to identify characters does not deteriorate even if the order of strokes is incorrect.

[Background of the invention]

Conventionally, in online handwritten character recognition methods, nine handwritten characters, which are different from ordinary printed characters, are classified in stroke order (order of strokes).
Taking advantage of the fact that information is included, the similarity between each stroke that corresponds in stroke order between the input character to be recognized and the standard pattern to be matched is calculated, and the sum of these is taken as the similarity between patterns. , a method is adopted in which the standard pattern that provides the maximum similarity among all the standard patterns to be matched is output as the recognition result of the input character.

This method assumes that each stroke of the input character is written according to the standard stroke order! If the stroke order of the input characters is incorrect, the degree of similarity between strokes corresponding to each other in stroke order will be extremely small9, causing erroneous identification.

Tokuko Sho 58 was developed as a method to correctly match each stroke between the input character and the standard pattern and calculate the similarity between patterns even when there is an error in the stroke order of the input character.
A method using an inter-stroke similarity matrix described in Japanese Patent No. 27551 is known.

This method calculates, for each stroke of the input character, the degree of similarity with all strokes of the standard pattern, and considers the stroke of the standard pattern that gives the maximum degree of similarity to correspond to the stroke of the input character. , the sum of the maximum similarities obtained for all strokes of the input character is used as the inter-pattern similarity between the input character and the standard pattern.

According to this method, no matter what kind of stroke order error p occurs in the input character, it is possible to correctly associate each stroke between the input character and the standard pattern and obtain the inter-pattern similarity.

However, according to this method, the number of calculations of the mutual similarity between each stroke of the input character and the standard pattern (the number of elements of the inter-stroke similarity matrix) is equal to 2 times the number of strokes of the input character.
Since the number of multiplications is large, especially when recognizing large kanji with a large number of strokes and the number of categories (the number of standard patterns to be matched), the number of times the similarity between strokes is calculated becomes excessive and it takes a lot of time. . This impairs the immediacy and real-time nature of online handwritten character recognition, which outputs recognition results immediately after writing input characters.

[Purpose of the invention]

An object of the present invention is to provide an online handwritten character recognition method that can correctly match each stroke between an input character and a standard pattern even when there is a stroke order error in the input character, and can significantly reduce the time required to calculate similarity between patterns. We are committed to providing this.

[Summary of the invention]

Kanji forms one character by combining so-called partial patterns, such as r kai, j, ``″'', thread, word, ... j.For example, the character ``ken'' has three parts: ``word, blade, and heart.'' It can be said that it is composed of partial patterns. When writing these kanji, brush 1'El is ``word''.

The partial patterns of ``blade'' and ``heart'' are written in this order, and the writing order of these partial patterns is almost never incorrect. However, the stroke-based Fudenowan is within each partial pattern.
Mistakes often occur.Here, when the correct stroke order is the stroke that should be drawn second, the stroke order variation is i-)
Then, when writing a character consisting of partial patterns with the number of strokes m, , m, , .・・・・・・
9m・)-11.

For example, if we consider the 14-stroke character ``recognition'' as consisting of three partial patterns: ``word, blade, and heart,'' the number of strokes in each partial pattern is 7.3.4 strokes. The maximum value of stroke order variation is ±1max(7,3,1i)-I
I-±6.However, if you write the 7-stroke partial pattern "Miya" in completely reverse stroke order, this value will change the number of strokes that should be written in the first (seventh) position. 7th (1st
), which corresponds to when written in practice, rarely occurs. In the characters actually written, the amount of variation in stroke order is
It is quite small.

Figure 1 shows 8 stroke characters of educational kanji, 25 characters in total of 27 famous sentences.
Out of 38 characters, 2608 of 626 characters have stroke order errors.
(= 326 x 8 ) This shows the distribution of the stroke order variation amount of strokes.

As shown in the same section, the stroke order variation is not uniformly distributed from -7 to 7, but is distributed in a narrow range centered around 0. For example, the stroke order variation is distributed from -2 to 2. The ratio is 5'7.5%.

Therefore, in this example, the th input character is the th st)o-
In other words, it is possible to correctly associate strokes with any one of the i-2 to i+2 standard pattern strokes at a rate of 97.5%.

Although Fig. 1 shows a character with 8 strokes, as mentioned above, considering that stroke order errors occur within a partial pattern, the distribution of the amount of stroke order change kh for characters with other stroke numbers can also be considered. This is considered to indicate that

The online handwritten character recognition method according to the present invention is based on the distribution of the stroke order variation amount, and the number stroke of the input character and the i-Ath to &+
Calculate the inter-stroke similarity with only 2A+1 strokes up to the A-th stroke, make the standard pattern stroke that gives the maximum similarity correspond to the first stroke of the input character, and calculate for all strokes of the input character. The sum of the maximum angles obtained is the inter-pattern similarity between the input character and the standard pattern.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to FIG.

FIG. 2 is a block diagram of a device used in an embodiment of the present invention, in which 1 is a character input device, for example a tablet, and the tablet 1 is a writing instrument used to write characters, so-called input pen 2. Information indicating whether or not the pen is crimped to the set arrangement (hereinafter referred to as pen up-down information) and coordinate value information of handwriting are output at regular intervals. 3 is a tablet interface unit that converts an output signal from the tablet 1 into a signal that can be input to the microprocessor 4; The microprocessor 4 performs recognition processing based on the input character information inputted through the tablet interface unit 3 and outputs a standard pattern character code, which is the recognition result, to the output terminal 8 via the output interface unit 5. It performs the following actions. 6 is used, for example, as a work area in a random access memory (RAM)-T:l>D recognition process. Reference numeral 7 denotes a read-only memory (ROM), for example, which records information on the standard pattern to be verified, a program defining the operation of the microprocessor 4, and the like.

An overview of the recognition processing performed by the microprocessor 4 in FIG. 2 will be described below with reference to FIG. 3.

The recognition process includes input character data import processing 41. Preprocessing 42. Feature extraction process 43. This is carried out by the steps of identification processing 44 and identification result output 45.

First, the input character data import process 41 is performed on the tablet 1.
The character information (pen up-down information and handwriting coordinate value information) for one character written on the top with the input pen 2 is taken in at regular intervals via the tablet interface section 3, and the R
AM Store it in a predetermined area of B. And pre-processing 42
In this step, the number of strokes of an input character is detected based on the pen up-down information, and the input character information is separated into a series of pen point coordinate values for each stroke. Further, known noise removal, normalization of the writing point coordinate value series, and preprocessing of position, size, etc. are performed.

In a feature extraction process 43, known feature extraction is performed based on the preprocessed input character information. The input character information sentence from which features have been extracted has the number of strokes of the input character M,
If each of the extracted strokes is written in stroke order (stroke order) as X, , X, ,..., xM, then large = (X+, X2,...,
It is expressed in a multidimensional vector space called ・・・-・・・・・・・・・・・・・・・(1).

Based on the extracted input characters/jh information Xi, in the identification process 44, the inter-pattern similarity ρ with the standard pattern to be matched is determined, and the character code PC of the standard pattern that gives the maximum inter-pattern similarity is identified. Result. Then, this identification result is outputted to the output terminal 8 via the output interface section 5 as an identification result output 45.

In the identification process 44, first, in 441, the maximum inter-pattern similarity ρ11111J is initialized to Co (constant),
Inter-pattern similarity ρ with the standard pattern S to be matched (where t=1 to L, L is the number of standard patterns to be matched)
0 is calculated at 443, and when ρ' is greater than ρ2, 1 at 4441, ρ21 is replaced with this ρ0 at 4442, and the standard pattern y at this time is calculated at 4443.
The character code PC' is stored in the character code PC. This procedure is repeated at step 442 for each standard pattern to be matched. At this time, the group of standard patterns to be matched are standard patterns equal to the number of strokes M of the input characters, which are stored in the ROM y in a predetermined order, and are sequentially imported into the microprocessor 4 to calculate the similarity between the patterns. do.

The method of calculating the inter-pattern similarity ρ6 in step 443, which is a feature of the present invention, will be described below with reference to an example shown in FIG.

The similarity between patterns ρ゛ is the stroke X of the input sentence,
and the stroke S' of the standard pattern y! Step 2) Calculate based on the similarity between the two points.

Here, if each stroke stroke order of the standard pattern S is S, , S, , ..., S5, the standard pattern selection is +t S = (S,, St, ..., 5.)・・・・・・・・・
・・・・・・・・・Tap・ It is expressed in a multidimensional vector space of +21.

Because the input characters are not necessarily written in the correct stroke order,
Even when input characters are matched with standard patterns in the same category, there is no guarantee that they will match in stroke order. That is, X,: S”;,
There is no guarantee that X2; S: , , , , XM: Do. Therefore, each stroke Xi of the input character is made to correspond correctly to each stroke St of the standard pattern Tobi 6, and the degree of similarity between the strokes that corresponds correctly is ρ8. It is necessary to use the sum total as the inter-pattern distance.

The correspondence between the stroke Xi of the input character and the standard pattern stroke S is based on the distribution of stroke order variation amount (first
) based on the stroke order K S, -j to Si
Among strokes up to +B, maximum inter-stroke similarity ρ: 1. Match the strokes of a standard pattern that gives . Then, the sum of the inter-stroke similarities ρ with the strokes S' of the standard pattern corresponding to each stroke Xi is defined as the inter-pattern similarity ρC.

The procedure for calculating this inter-pattern similarity ρC will be explained in detail below.

First, in 4431, the value of inter-pattern similarity ρC is initialized to 0. Then, at 4432, for each stroke order of the input character, 4433 is described below.

The processes of 4434, 4435, and 4436 are repeated 9 times from i=1 to M. In 4433, the range in which the strokes of the standard pattern should correspond to the strokes Xi of the input character is set. That is, for Xt, 'S' and jE are set which should correspond to any one of the standard pattern strokes S up to S in order.

Here, ノS is K as shown in 44331, and ノS=muA
No.E is 44332K. As shown, No.E=i+B.

Also, when the judgment of 443 is 5<1, 1.11
If S=1 at 4334 and E>N as determined at 44335, then E=N at 44336. However, N is the number of strokes of the standard pattern to be matched,
Usually, it is the same value as the number of strokes M of the input character.

Further, the numerical values A and B are values that should be determined based on the distribution of the stroke order variation amount, and will be explained below assuming that both are 2 here.

Figure 5 shows each Xi when the number of strokes of the input character is 8 strokes.
The target stroke range (other than the shaded area) of the standard pattern is shown.

Next, in step 4434, the value of the maximum degree of gradation ρ2,5 is initialized to a constant C8. And at 4435,
For the stroke S of the standard pattern consisting of ノS and ノE set previously, () = ノS ~ ノE), each inter-stroke similarity ρ with the stroke Xi of the input character is calculated, and among them The degree of image between strokes that gives the maximum value ρ2, 1
Detect. ρ8 which gives the maximum value in FIG. It is expressed by surrounding it with a circle. This detection is performed in 714353 when the inter-stroke similarity ρ, , determined in 44352 becomes ρ, 〉ρ,
This is done by replacing =ρ, . And 44
At step 36, this ρ;,1 is added to the inter-pattern similarity ρ', and the following formula is performed: ρC=ρ+ρ2,1.

Through this series of processing steps, the stroke of the input character
Lc, standard pattern stroke 54 (no i-2~L
+2), which gives the maximum inter-stroke similarity, and the sum of the inter-stroke similarities (the value of ρ1. circled in Fig. 5) with the corresponding S for each of these Xi. Thus, the inter-pattern similarity ρC can be obtained.

In the above embodiment, the stroke Xi of the input character is a standard pattern stroker (where: )=i-A-i+B)
Stroke SL that gives the maximum similarity between strokes
However, instead of this, the stroke S' of the standard pattern is given the maximum similarity between strokes of the input character strokes Xi (L-No A to No 10 B). It goes without saying that the same effect can be obtained even if the stroke is made to correspond to the stroke Xi.

Furthermore, in the above embodiment, the standard pattern to be matched is
Although it is assumed that the stroke IIM is equal to the stroke IIM of the input character, this is not limited to this. For example, the present processing method can also be applied to the recognition of input characters in which the number of strokes is reduced compared to the original character due to the occurrence of writing in the input character. can be made recognizable by For example, if it is allowed to write the continuation of the input character, the standard pattern to be matched is, when the number of strokes of the input character is M, M, M+1, . ...
A standard pattern group of 1M+le may be used.

Also, at this time, the number of strokes of the standard pattern A' =
For example, the constant B that sets the target stroke range for M, Jf + i, . . . 1M+le is 2.
3. ..., just set it like 2+7L.

Further, the constants A and B can also be changed depending on the accuracy of recognition.

[Effects of the invention-]

As explained above, according to the present invention, even if there is a flower order error in the input character, each stroke between the input character and the standard pattern can be correctly matched, and the time required to calculate the similarity between patterns can be significantly reduced. can be reduced.

For example, if the calculation range of the stroke similarity ρ, , between the stroke Xi of the input character and the stroke S of the standard pattern is from No2 to No+2, if the input character has 8 strokes, then ρ, , The number of calculations is 34, which can be reduced to about 53 compared to 64 in the conventional method. This value is more effective as the number of strokes increases; for 12 strokes, it is approximately 35%
, it is about 26% for 18 strokes.

This shows that for kanji characters with a large number of strokes and a large number of categories, it is possible to significantly reduce the calculation of similarity between patterns according to the present invention. In addition, the present invention not only has the effect of reducing the amount of processing, but also
(7) by limiting the range of correspondence of standard pattern strokes to the stroke order based on the distribution of stroke order variations.
Roke Xi and standard pattern stroke SL Nor-no 1 can avoid mismatching of extremely large strokes. This is effective in a recognition method in which the inter-stroke similarity ρ, J is defined only by stroke shape information. This is because kanji are a combination of letters 1-o-ri with the same or similar shape, so I i
-)' This is because it often happens that correspondences with extremely large I give the maximum similarity.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the distribution of stroke order variation, FIG. 2 is a diagram showing an example of the configuration for performing the online handwritten character recognition method according to the present invention, FIG. 3 is a diagram showing the general procedure of recognition processing, and FIG.
FIG. 5 is a diagram showing a procedure for calculating similarity between patterns according to the present invention, and FIG. 5 is a conceptual diagram for explaining the operation of calculating similarity between patterns. 1...Tablet 2...Input 4...Microprocessor 6...
...... RAM 7...ROM 443... Inter-pattern similarity calculation process 4433
... Setting processing of target stroke range 1 Figure

Claims (1)

[Claims] 1. Inter-stroke similarity between each stroke of an input character to be recognized and each stroke of a standard pattern to be matched is calculated, and the inter-stroke similarity is used as a matrix element. In an online handwritten character recognition method that recognizes the input character based on the degree matrix, the inter-stroke similarity ρ_i_j between the i-th stroke Xi of the input character in stroke order and the j-th stroke of the standard pattern θ So, j=jS to jE, jS=i-A, jE=i+B However, when jS<1, jS=1 When jE>N, jE=N N is the number of strokes of the standard pattern 3 A and B are the positive Find the integer of , detect the inter-stroke similarity that shows the maximum value, and calculate the sum of the inter-stroke similarities that show the detected maximum value for all strokes of the input character, and calculate the integer between the patterns with the standard pattern θ. An online handwritten character recognition method characterized in that a code of a standard pattern that gives a maximum similarity between patterns among standard patterns to be matched is used as a recognition result of the input character to be recognized. 2. The online handwritten character recognition method according to claim 1, wherein the positive integers A and B are constant values regardless of the number of strokes of the input character to be recognized. 3. Claims characterized in that the positive integers A and B are varied according to the difference between the number of strokes M of the input character to be recognized and the number of strokes N of the standard pattern to be compared. The online handwritten character recognition method described in Section 1.