JPH0150954B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to an online handwritten character recognition device, and more particularly to an online handwritten character recognition device that reduces the burden on a scribe and prevents a decrease in character entry speed. The prior art and its problems will be explained with reference to FIGS. 1 and 2. In Figure 1, 1 is a doublet (coordinate input device), 2 is a character entry sheet, 3 is an input pen, 4 is a preprocessing section, 5 is a feature extraction circuit, 6 is a matching circuit, and 7 is a minimum value selection circuit. , 8 is a code conversion circuit, 9 is an output end, 10 is an UP/DOWN detection circuit, 11 is a stroke number detection circuit, 12 is a standard pattern memory, 13 is a position comparison circuit, 14 is an instruction frame position memory, 15 is a capital letter / Lowercase flag memory. The scribe uses the input pen 3 to
Characters are written in each character entry frame 2-1 on the character entry sheet 2 placed on the tablet 1. At this time, the tablet 1 outputs the positional information of the XY coordinates of the tip of the input pen 3 from the output lines 1-3 at regular intervals (sampling period). In addition, the input pen 3 has a built-in switch that detects whether or not the input pen 3 is pressed against the character entry sheet 2, and the output of this switch also provides the position information of the XY coordinates as Z-axis information. It is also output from output lines 1-3 every sampling period. These X, Y, and Z axis information are supplied to the preprocessing section 4 as handwriting information. The preprocessing unit 4 first looks at the Z-axis information, selects and imports only the data that the input pen 3 presses onto the character entry sheet 2 (hereinafter referred to as handwriting points), and performs the following normalization process. . The series of handwriting points includes redundant points. This is because the pen tip does not move at a constant speed when writing characters, and the spatial distance between adjacent handwriting points is very close. For this purpose, redundant handwriting points are removed. The removal method is to resample handwriting points that are a certain distance away from the starting point of a stroke (one line segment drawn from when the input pen 3 presses on the character entry sheet 2 until it leaves the character writing sheet 2, a series of handwriting points). Then, the process of again setting handwriting points that are a certain distance away from this re-marking point as re-sampling points is performed until the end point of the stroke. That is, the handwriting point series of each stroke is converted from a time-space series to a distance-space series (hereinafter, this process will be referred to as resampling process). Next, the preprocessing unit 4 normalizes the position and size of the data for one character that has been resampled. Which character entry frame 2 on character entry sheet 2
The XY coordinate values of the handwriting point differ depending on where the character is written in 2-1.
Since the XY coordinate values are different, position normalization involves converting the coordinates of each character so that the center of gravity of the character remains constant. In addition, since the size of the characters written in the character entry frame 2-1 differs depending on the scribe, it is necessary to perform coordinate transformation of each resampling point so that the size of the characters written is constant. This is the normalization of This is done by ensuring that the average distance of each resampling point to the center of gravity of the character is constant. The input character data preprocessed in this manner is expressed by the feature extraction circuit 5 in a form with a reduced amount of information so that subsequent processing can be easily performed.
For example, an input character I〓 consisting of M strokes is expressed in the writing order of the strokes, such as I〓=(I ₁ , I ₂ , . . . , I _M ), where the mth stroke written is Im.
In addition, each stroke I ₁ to I _M is made up of N+1 broken line approximation points that divide the line segment of one stroke from the start point (starting handwriting point) to the end point (finishing handwriting point) into N equal parts. Express as a series.
In other words, the mth stroke I _n is the approximate point of the broken line
Using the sequence of P _n1 , P _n2 , ..., P _nN+1, it is expressed as I _n = (P _n1 , P _n2 , ..., P _nN+1 ). Here, the broken line approximate point P _no is the XY coordinate value shown by P _no = (x _no , y _no ). The input characters described by the feature extraction circuit 5 in this manner are supplied to one input terminal of the matching circuit 6. At the other input terminal of the matching circuit 6, a standard pattern memory stores an average pattern of characters input by a large number of scribes, which has undergone the same preprocessing and feature extraction as for the input characters for each character to be recognized. 12. Here, let the standard pattern S〓〓 for the character θ be S〓〓=(S〓 ₁ , S〓 ₂ ,..., S〓 _M ). However, M is the number of strokes of the character θ and S〓
_n is the m-th stroke expressed as S〓 _n = (P〓 _n1 , P〓 _n2 , . . . , P〓 _nN+1 ). P〓 _no is the line segment of the m-th stroke expressed as P〓 _no = (x〓 _no , y〓 _no )
This is the nth XY coordinate value of the approximate point of the polygonal line that divides into equal parts. The matching circuit 6 calculates the distance D (θ) between the input character I〓 and the standard pattern S〓 whose stroke number is equal to the stroke number M of this I〓 as follows. D (θ) = _M 〓 ^m=1 dS (S〓 _n , I _n ) (7) Here, I _n is the mth stroke of the input character I〓, and S〓 _n is the mth stroke of the standard pattern S〓〓. The m-th stroke, dS (S〓 _n , I _n ) indicates the distance between the m-th stroke S〓 _n , I _n of both patterns, and dS (S〓 _n , I _n )= _N+1 〓 ^{n= 1} dP (P〓 _no , P _no ) (8). N+1 is the number of stroke line approximation points, P〓
_no , P _no are the n-th broken line approximate points of the m-th strokes of both patterns, and dP (P〓 _no , P _no ) are the points between the n-th broken line approximate points of the m-th strokes of both patterns. Indicates the distance dP (P〓 _no , P _no ) = √ (〓 _no − _no ) ²
+(〓 _no − _no ) ² (9). Combining the above equations (7) to (9), matching circuit 6
_is ^the _{distance between} ^both _patterns ^D _{( θ} ) ^.
Calculate. However, x〓 _no , x _no is the X coordinate value of the nth broken line approximate point of the mth stroke, y〓
_no and y _no are the Y coordinate values as well. Here, if there are L standard patterns equal to the number of strokes M of the input character I〓, the matching circuit 6 sequentially matches the input character I〓 to these L standard patterns.
Calculate the inter-pattern distance D(θ) with
It is supplied to the minimum value selection circuit 7. The number of strokes of the input character I〓 is determined by detecting UP and DOWN of the input pen 3 by the UP/DOWN detection circuit 10 based on the Z-axis information from the tablet 1, and then from UP to DOWN. The stroke number detection circuit 11 counts the change in the number of strokes over one character. The output of the stroke number detection circuit 11 is stored in the standard pattern memory 1.
2, selects a standard pattern S〓 equal to the number of strokes M of the input character I〓, and supplies it to the matching circuit 6. The minimum value selection circuit 7 detects the minimum value of L inter-pattern distances D(θ ₁ ) to D(θ _L ) sequentially supplied. If the detected minimum value is D(θ ¹ ), the input character I〓 is recognized as a character whose standard pattern is S〓〓 ¹ , and the character code corresponding to the standard pattern S〓〓 ¹ is set as the standard. It is fetched from the pattern memory 12 and outputted to the code conversion circuit 8. The problem with the conventional online handwritten character recognition method described above is that, for example, the uppercase letter ``ki'' and the lowercase letters ``ya'', ``yu'', and ``yo'' in the kana characters ``kiya'', ``kiyu'', and ``kiyo'' are difficult to recognize. It is an input method and recognition method for uppercase and lowercase letters.
These lowercase letters have exactly the same shape as the uppercase letters, and differ only in size. In the conventional technology, the difference in the size of input characters is normalized in the preprocessing unit, and is not recognized in the recognition unit. This is because the characters will be exactly the same, making it impossible to determine whether they are uppercase or lowercase. For example, when the kana ``tsu'' is written in the entire text box 2-1 as shown in a in Figure 2, and when it is written small in the corner of the text box 2-1 as shown in c. Then, the preprocessing results are as shown in b and d.
With W ₀ as the center of gravity, the pattern becomes exactly the same. In this way, uppercase and lowercase letter patterns with the same shape are Japanese katakana,
It exists not only in hiragana but also in English. In order to cope with this problem, a method has been adopted in the past in which a scribe supplies information on whether a letter is an uppercase or a lowercase letter via the tablet 1 when writing. For example, as shown in FIG. 1, the tablet 1 is provided with an uppercase letter instruction frame 1-1 and a lowercase letter instruction frame 1-2, and by pressing either of the instruction frames with the input pen 3, subsequent input can be performed. Declares a character to be uppercase or lowercase. Then, in the recognition section, the position comparison circuit 13 selects the instruction frame 1--
Using the coordinate values corresponding to 1 and 1-2 (stored in the instruction frame position memory 14) as a reference for comparison, it is detected which instruction frame is pressed by the input pen 3, or whether it is not pressed at all. Then, the uppercase/lowercase flag memory 15 is set in accordance with the detection result. Here, the output F of the uppercase/lowercase flag memory 15 is set to F=1 for uppercase letters and F=1 for lowercase letters.
Set to 0. Also, the character code for the capital letter “tsu” is
Assume that A5C4, the character code for the lowercase "tsu" is A5C3, and the uppercase character code A5C4 is given to the standard pattern for "tsu". When the character "tsu" is written with the input pen 3, the minimum value selection circuit 7 outputs the character code A5C4.
This character code is supplied to the code conversion circuit 8. When the instruction from the uppercase/lowercase flag memory 15 is F=1 (uppercase instruction), the code conversion circuit 8 directly outputs the character code A5C4, and F=
When 0 (lowercase letter instruction), the character code A5C3 corresponding to lowercase letters is output. That is, the code conversion circuit 8 internally has a character code correspondence table of lowercase letters corresponding to uppercase letters, and uses this to convert the uppercase character code to a lowercase character code when F=0 and outputs the converted character code. However, in the conventional method described above, in addition to writing the letters, it is also necessary to indicate whether the letters are uppercase or lowercase, which not only burdens the scribe but also reduces the input speed. There is. The purpose of the present invention is to provide online handwritten character recognition that can solve the above-mentioned problems in the prior art, reduce the burden on scribes, prevent a decrease in character entry speed, and improve the recognition efficiency for English characters. The goal is to provide equipment. In order to achieve the above object, the features of the present invention are as follows:
a font size detection means for detecting the size of a written character; a font size determination means for comparing the detected value with a set value and determining that it is an uppercase letter when it is larger than the set value; and a lowercase letter when it is smaller than the set value; receives the character code corresponding to the result of pattern recognition of the written character and the above size judgment result signal as input, and when the size judgment result is an uppercase letter, outputs the character code corresponding to the written character as is, and the size judgment result When the character code is a lowercase letter, the code conversion means performs a predetermined code conversion based on the input character code and outputs the converted code. An embodiment of the present invention will be described below with reference to FIG. In Fig. 3, 4-1 is a resampling circuit;
-2 is a position normalization circuit, 4-3 is a centroid point extraction circuit, 4-4 is a size normalization circuit, 4-5 is an average radius extraction circuit, 16 is a comparison circuit, and the others are as in Figure 1. is the same as The handwriting information of the input characters is input to the resampling circuit 4-1 from the output line 1-3 of the tablet, and as described above, redundant points among the handwriting points of each stroke are removed, and each stroke is Convert the handwriting point series from a spatiotemporal series to a metric space series. In other words, the line segments from the start point to the end point of the stroke are resampled at regular distance intervals. The mth stroke _{I n} of this resampled input character I is resampled at points Q _n1 , Q _n2 ..., Q _nE (m)
Expressed as a series of I _n = (Q _n1 , Q _n2 , ..., Q _nE (m)). However, Q _ne is the e-th resampling point of the m-th stroke, and E(m) is the number of resampling points of the m-th stroke. Further, the resampling point Q _ne indicates the XY coordinate value and is Q _ne =(X _ne , Y _ne ). The input character data resampled in this way is supplied to the centroid point extraction circuit 4-3, and the centroid position W ₀ of the input character is extracted. This centroid position W ₀ is the total resampling point Q _ne for one character (m=1~M, e=1~
The average value X ₀ of the X coordinate value and the average value Y ₀ of the Y coordinate value of E(m)) are expressed as W ₀ =(X ₀ , Y ₀ ). Next, in the position normalization circuit 4-2, the coordinate values of each re-sampling point _Qne are transformed so that this center of gravity position _W0 becomes the origin of the new XY coordinate axes. In other words, as Q _ne = (X _ne −X ₀ , Y _ne −Y ₀ ), the XY coordinate value X _ne of each resampling point Q _ne ,
Subtract X ₀ and Y ₀ from Y _ne . Here, if x _ne =X _ne -X ₀ and y _ne =Y _ne -Y ₀ , then the center of gravity will be x=0 and y=0. Next, for the input character data whose position has been normalized as described above, the average radius extraction circuit 4-5 calculates the size of the input character, in this case the average radius R of the input character, by R=1. /U _M 〓 ^m=1 _E(n) 〓 ^e=1 {|xme|+|y _ne |}. Here, U= _M 〓 ^m=1 E(m) is the number of resampling points of the input character, M is the number of strokes of the input character, |x _ne |, |y _ne | is the center of gravity position of the input character
These are the absolute values of the X-axis value and Y-axis value of the e-th resampling point of the m-th stroke with W ₀ as the origin. That is, the average radius R is determined as the average value of the distance of each resampling point Q _ne from the center of gravity position W ₀ . This average radius R becomes a larger value as the characters are written larger, and is a parameter corresponding to the size of the input characters. The size normalization circuit 4-4 converts the coordinate values of each resampling point _Qne so that this average radius R becomes the set value _R0 . The resampling point Q _ne after this transformation process
Let _{the values of the X and Y} axes _of This is to normalize (divide) each XY coordinate value by the average radius R. This average radius R of the input characters is also input to one input terminal of the comparator circuit 16 and compared with a set value Rth input to the other input terminal. In other words, the comparison circuit 16 determines whether the input character is larger than the set value. The determination result is supplied to the code conversion circuit 8, and the operation of the code conversion circuit 8 is controlled. The output signal of the magnitude normalization circuit 4-4 is supplied to the feature extraction circuit 5, and as explained in the prior art,
The input character is expressed in a form with the amount of information compressed, and a matching circuit 6 performs a matching calculation (calculation of distance between patterns) with a standard pattern. In the standard pattern memory 12, capital letters such as A, B, C, . The average pattern of the processed and feature-extracted patterns is stored together with the character code corresponding to that character. Also, at this time, the characters are classified and stored according to the number of strokes. Here, as shown in Figures 1 and 2 of Figure 4, write the uppercase alphabetic letter "A" in the character entry frame 2-1 with the first stroke.
Assuming that the handwriting information on output lines 1-3 is written with two strokes: I ₁ "∧" and the second stroke I ₂ "-", the handwriting information on output lines 1-3 is pre-processed and feature extracted, and then stored in the standard pattern memory in the matching circuit 6. A matching calculation (calculation of inter-pattern distance) with a standard pattern consisting of two strokes is performed, and the results are sequentially supplied to the minimum value selection circuit 7. Then, the minimum value selection circuit 7 detects the minimum value of the inter-pattern distance D(θ) and supplies the character code of the standard pattern corresponding to the minimum value to the code conversion circuit 8. Here, set the standard pattern and the character code corresponding to the character (uppercase) as shown in the left part of Table 1,
It is also assumed that the character codes of lowercase letters corresponding to the uppercase letters are set as shown on the right side of Table 1.

[Table] In the minimum value selection circuit 7, when the distance between patterns for the capital letter "A" becomes the minimum, the minimum value selection circuit 7 supplies the character code A3C1 given to the capital letter "A" to the code conversion circuit 8. . and,
Are the average radii R (1) and R _{( 2) extracted by the average radius extraction circuit 4-5 larger than the set value Rth for the writing of the capital letter "A" in Figures 1 and 2} in Figure 4? Comparison circuit 16 compares whether or not it is true. Here, if it is determined that RthR ₍₁₎ ...in the case of Figure 1 Rth>R ₍₂₎ ...in the case of Figure 2, the code conversion circuit 8 converts the input character code A3C1 (character for "A") in the case of Figure 1 In the case of Figure 2, assuming that a small "A" is written, the input character code A3C1 (based on the character code for "A") is output as is.
The character code A3E1 of the corresponding lowercase alphabet "a" is
It outputs by referring to the uppercase-lowercase character code correspondence table (built in the code conversion circuit 8) shown in Table 1. The same is true when the katakana ``tsu'' is written as shown in FIGS. 3 and 4 of FIG. These conditions are summarized in Table 2.

[Table] (Note) Larger… Fill in the text box in larger size.
Small...means to write small in the character entry frame.
In the above example, the position is normalized so that the center of gravity of the input character is the origin, and the size of the input character is set to the average distance (average radius) between the center of gravity of the input character and the resampling point. , a configuration has been described in which the size is normalized based on this average radius, but the present invention is not limited to this, and instead of the above, position normalization is performed by using the center position of the circumscribed rectangle of the input character. Alternatively, the input character size may be set to the diagonal length of the circumscribed rectangle, and the size may be normalized based on this diagonal length so that the diagonal length is a constant. can. As explained above, according to the present invention, the size of an input character is detected, and when the detection result is larger than a certain value, it is judged as an uppercase letter, and when it is smaller than a certain value, it is judged as a lowercase letter, and the size of the input character is determined. By having a configuration that outputs the character code to be input, it becomes unnecessary to specify the uppercase and lowercase letters of input characters by pressing the input pen into the uppercase and lowercase letter indication frame area, thereby eliminating the burden on the scribe. At the same time, it has become possible to prevent a decrease in character input speed, and by excluding lowercase letters that are mainly composed of curved lines from being recognized, lowercase letters that were previously composed of mainly curved lines are no longer recognized. This has the effect of preventing a reduction in recognition rate due to various deformations caused by the scribe.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a conventional example, Fig. 2 is a diagram showing input characters and preprocessing results, Fig. 3 is a block diagram of an embodiment of the present invention, and Fig. 4 is a diagram showing an example of character writing according to the present invention. It is. Explanation of symbols, 1...Tablet, 2...Character entry sheet, 2-1...Character entry frame, 3...Input pen, 4
...Preprocessing section, 4-4...Size normalization circuit, 4-5
...Average radius extraction circuit, 5. Feature extraction circuit, 7. Minimum value selection circuit, 8. Code conversion circuit, 12. Standard pattern memory, 16. Comparison circuit.

Claims (1)

[Scope of Claims] 1. In an online handwritten character recognition device that writes characters on a tablet with an input pen and recognizes the written characters based on handwriting information of the written characters output from the tablet, font size detection means for detecting the font size; font size determination means for comparing the detected value with a set value and determining it as an uppercase letter when it is larger than the set value; and a lowercase letter when it is smaller than the set value; It receives the character code corresponding to the result of character pattern recognition and the above size judgment result signal as input, and when the size judgment result is an uppercase letter, the character code corresponding to the written character is output as is, and when the size judgment result is a lowercase letter, the character code corresponding to the written character is output as is. What is claimed is: 1. An online handwritten character recognition device comprising code conversion means for converting a predetermined code based on a character code received as input and outputting the converted code. 2. In the device according to claim 1,
The character size detecting means calculates the average distance between the center of gravity of the written character and each of the remarking points, which is obtained based on the resampling points extracted from the line segments of each stroke of the written character at regular distance intervals. An online handwritten character recognition device characterized in that it is a character size detection means whose value is the size of a handwritten character. 3. In the device according to claim 1,
The online handwritten character recognition device is characterized in that the character size detection means is a character size detection means that determines the size of the handwritten character to be the diagonal length of a rectangle circumscribing the handwritten character.