JPH0474755B2 - - Google Patents

Description

TECHNICAL FIELD The present invention relates to character recognition, and more particularly to online character recognition for alphabets, numbers, katakana, and the like.

(Background Art) Conventional online character recognition devices have been constructed using information that is available online. That is, the information includes information about the stroke itself and information indicating the order of the strokes. Using these two pieces of information,
It was a so-called stroke analysis method in which the characteristics of each stroke were stored in a dictionary in detail, and the characteristics of the input strokes were recognized by referring to the dictionary.

In the case of the stroke analysis method, alphabets, numbers, katakana, etc. (hereinafter referred to as ANK, etc.)
Even in the case of around 100 characters, the dictionary becomes enormous because the characteristics of those characters are described in the dictionary for each stroke, including the transformation of each character, and therefore the processing time becomes large.If you try to shorten the processing time, The drawback was that the hardware scale was large.
In addition, in order to compensate for these drawbacks, it is necessary to impose restrictions on the number of strokes and the order of strokes, and these restrictions on the font, number of strokes, and order of strokes are also major drawbacks that have forced scribes to inconvenience.

(Problem to be solved by the invention) The purpose of the present invention is to eliminate these drawbacks.
A major classification of online character recognition that recognizes written characters by considering the entire character as a figure and identifying how stroke line segments are distributed within the written character, without relying on the stroke analysis method. It is intended to provide a legal framework and will be explained in detail below.

(Structure and operation of the invention) FIG. 1 is a block diagram showing an embodiment of the present invention, in which 1 is a tablet, 2 is an analog-to-digital converter (hereinafter referred to as A/D), and 3 is an input register (hereinafter referred to as A/D). 4 is an arithmetic unit containing arithmetic registers (hereinafter referred to as ALU), 5 is a partial value register (hereinafter referred to as PR), and 6 is a position register (hereinafter referred to as PR).
7 is a reference register (hereinafter referred to as AR), 7 is a reference register (hereinafter referred to as
8 is a comparator (hereinafter referred to as COMP),
9 consists of candidate registers (hereinafter referred to as SR), and the characters written by the scribe on tablet 1 are A/
digitized by D2 and this data is IR
3 are sequentially stored. IR3 has tablet 1
The data is stored according to the sampling rate of the tablet 1, and the stored data is data indicating the absolute coordinates of the tablet 1. At the same time as data is stored from A/D2 to IR3, IR3 data is stored in a time-sharing manner.
It is sent to the ALU 4 and converted into a relative coordinate system centered on the stroke start point first written as a character within a character frame, and the same points are deleted. This is to avoid complicating subsequent processing due to input of the same point if the scribe places the writing instrument at the same position on the tablet 1.

Note that the start and end points of the stroke are IR3,
It is assumed that markings are made inside PR5 so that it can be determined which data is the start point or end point of each stroke. In this way, the data stored in PR5 is preprocessed using ALU4. Preprocessing consists of denoising and smoothing, but is not the subject of this invention.

As explained above, the data group input from the tablet 1 is subjected to identical point removal, noise removal, and smoothing, and is stored in the PR5. The number "6" is written on tablet 1, and the data group stored in PR5 is connected with a straight line in the order of occurrence of each data.
It is a diagram.

From the figure in Figure 2, (1) the starting point, (2) the ending point, (3) the point where the x and y stroke direction changes (hereinafter referred to as the extreme point), (4)
When the figure in Figure 2 is rotated 45 degrees, the x and y poles are taken as representative points, and if the distances between these representative points are close, the figure in Figure 3 can be created by ignoring the nearby representative points. obtain. FIG. 3 shows a data group resulting from the above-described processing performed on the data stored in the PR5, which is connected by a straight line in the order of occurrence of each data, and the resulting data group is placed in the PR5.

Now, when characters corresponding to ANK are viewed as figures, the position of the strokes is relatively characteristic. This is a proposal to look at it from the perspective of where it is, and will be explained in detail below. Note that the straight line connecting the representative points in FIG. 3 is hereinafter referred to as a segment. Furthermore, it is assumed that all the explanatory texts describe data that has been normalized in both the x and y directions.

First, the line length of the entire stroke described is used as the first feature. As shown in Figure 3, if Hx and Hy are the maximum coordinate values among the representative points in the x and y directions, then Q ₁ = 1/Hx _o 〓 ⁱ⁼¹ ｜x _i −x _i+1 ｜ Q The value expressed by ₂ = 1/Hy _o 〓 ⁱ⁼¹ |y _i −y _i+1 | is used. That is, Q ₁ is the normalized sum of the x components of the line segments obtained by connecting the representative points, and Q ₂ is the normalized sum of the y components of the same line segments. These Q ₁ and Q ₂
represents the distance that the writing instrument has moved in the x and y directions as a ratio to the size of the character.

The second feature is the position where each segment exists. The term "segment" here refers to a line segment obtained by connecting representative points in a predetermined order along the character. The position where a certain segment exists in the x and y directions is digitized for each x and y component, and the average value thereof is calculated. Regarding the position of the x component on the x-axis, Q ₃ = _o 〓 〓 ⁱ⁼¹ {|x _i −x _i+1 |/Hx×1/Hx・x _i +x _i+1 |/2}=
1/2Hx ² _o 〓 ⁱ⁼¹ ｜x _i ² −x _i+1 ｜ Similarly for the position of the y component on the y axis, Q ₄ = 1/2Hy ² _o 〓 ⁱ⁼¹ ｜y _i ² −y It is represented by _i+1 | Regarding the position of the x component on the y axis, Q ₅ = 1/2Hx・Hy _o 〓 ⁱ⁼¹ {|x _i −x _i+1 |・(y _i +y _i+1 )}, and the position of the y component The position on the x-axis is expressed as Q ₆ =1/2Hx·Hy _o 〓 ⁱ⁼¹ {|y _i −y _i+1 |·(x _i +x _i+1 )}. That is, Q ₃ indicates whether the horizontal strokes in a character are distributed toward the right or left, and Q ₄ indicates whether the vertical strokes are distributed toward the top or bottom. Similarly, Q ₅ indicates whether horizontal strokes are distributed toward the top or bottom, and Q ₆ indicates whether vertical strokes are distributed toward the left or right.

To give a specific example, for the character ``Ri'', the value of _Q1 is large because there are many horizontal strokes, and because the horizontal strokes are distributed towards the right side of the character,
The value of Q ₃ increases. On the other hand, for the character "river", the x component is only a slight curvature of the first stroke, so Q ₁ takes a very small value. Furthermore,
Since the number of strokes itself is small, Q ₂ also takes a relatively small value. In addition, regarding the character "Yu", Q ₃ and Q ₅ take large values because the right side of the character is complex. In particular, _Q3 takes a large value because there are many horizontal strokes on the right side.

All of the above examples are kanji, and only vertical strokes and horizontal strokes are illustrated, but since the above processing is performed by finding the x and y components of the segment from representative coordinate values, it can be used for ``hiragana'' or ``hiragana''. Curved lines and diagonal strokes, which are often found in Katakana, are also subject to processing.

As is clear from FIG. 3, the above Q ₃ to _{Q 6} are calculated using a coordinate system whose origin is the lower left of a rectangle circumscribing the written character. Therefore, the position where the x and y components of each segment are located becomes smaller as the position on the y-axis is lower, and as the position on the x-axis is further left, the feature amount becomes smaller. Since all the weights related to the positions of all segments are added, changes in segments near the origin appear as small differences, and are influenced by changes in segments in the upper right part of the rectangle circumscribing the character. The Rukoto. Therefore, as shown in FIG. 4, calculations regarding the above-mentioned Q ₃ to _{Q 6} are performed in a coordinate system whose origin is the upper right corner of a rectangle circumscribing the character, and the results are set as Q ₇ to _{Q 10} .

Specifically, for the coordinate transformation P _i = x _i −Hx, q _i = y _i −Hy, Q ₇ = 1/2Hx ² _o 〓 ⁱ⁼¹ | p _i ² −p _i+1 ² | Q ₈ =1/2Hy ² _o 〓 ⁱ⁼¹ |q _i ² −q _i+1 ² | Q ₉ =1/2Hx・Hy _o 〓 ⁱ⁼¹ {|p _i −p _i+1 |・(q _i +q _{i +1} )} Q ₁₀ = 1/2Hx・Hy _o 〓 ⁱ⁼¹ {|q _i −q _i+1 |・(p _i +p _i+1 )}. Since these Q ₇ to _{Q 10} are simply coordinate transformed to Q ₃ to _{Q 6} , their graphical meaning is
Same as Q ₃ to Q ₆ . However, in the calculations of Q ₇ to _{Q 10} , the upper right is the origin as described above, so the influence of segment changes in the upper right portion of the character is reduced. Therefore, by comparing Q ₃ to _{Q 6} and Q ₇ to _{Q 10} , we can distinguish between a relatively large change in the segment in the lower left part of the character and a relatively small change in the upper right part of the character. . Therefore, more detailed classification is possible than when _Q7 to _Q10 are not used.
Note that the above calculation is performed on all segments of all strokes for a character with multiple strokes.

Using the formulas Q ₁ to Q ₁₀ explained above, the data of PR5 is calculated by ALU4 and stored in AR6. The maximum and minimum values of Q ₁ to Q ₁₀ for each character, including various transformations, are stored in RR7. Characteristics of input characters stored in AR6 Q ₁ ~ _{Q 10}
is compared with the characteristics of each character in RR7 in COMP8. Since the minimum and maximum values are stored in RR7, COMP8 checks whether the contents of AR6 are within the range from the minimum value to _the maximum value stored in _RR7 . compared with
If all of Q ₁ to _{Q 10} of AR6 are included in the range indicated by RR7, the character giving the value of RR7 at that time is sent to SR9 and stored. As a result, one candidate character is stored in SR9. In this way, when COMP8 finishes verifying whether all characters are within the range for about 100 characters such as ANK, the characters stored in SR9 become a group of roughly classified characters.

After this, each character can be recognized by a known method, such as a stroke analysis method.

As the number of characters increases, some characters cannot be clearly distinguished even in this way. Furthermore, since the characters that are written differ from the standard character shape depending on the scribe and from time to time, the feature values of the coordinates to be compared have a wide range, which increases the number of characters that cannot be distinguished. However, since the number of candidate characters can be significantly reduced by simple calculations as described above, the recognition processing time is significantly reduced compared to when all characters were recognized using the conventional method.
This reduces the dependence on hardware, and eliminates the need to impose restrictions on the number of strokes or the order of strokes.

I think this can be better understood by showing the number 8 as a familiar example. The number 8 can be written in various ways, such as 〓〓〓〓〓, if the arrow is the direction of the stroke order, but when viewed as a figure, it is similar, and the values of Q ₁ to _{Q 10} , which indicate the distribution of stroke line segments, are In order to fall within a certain range, stroke analysis methods must take different views, but according to the method of the present invention,
No matter which letter is written, it can be identified as the number 8 by just one way of looking at it.

As explained in detail above, the present invention provides a general classification method for recognizing written characters by considering characters as figures and identifying how stroke line segments are distributed within the written characters. Therefore, compared to the stroke analysis method, the calculation costs such as hardware and processing time are lower, and even if "7" is written as one stroke or two strokes, the number of strokes will be reduced. Regardless, even in the case of a character with multiple strokes, the values of Q ₁ to _{Q 10} are determined and compared based on the values for all segments of all strokes, so the order of strokes is not restricted. That is, it is possible to provide a general classification method for online character recognition that is not affected by the number of strokes or the order of strokes.

In the above embodiment, it was explained that COMP8 sends characters whose data in AR6 falls within the data range of RR7 to SR9 as candidates _.
~Q If even one of _{the 10} data in AR6 deviates from the range from the minimum value to the maximum value shown in RR7, it is excluded from the candidate characters. In order to prevent this, a separate flip-flop circuit (hereinafter referred to as FF) is provided, and when _Q1 to _Q10 deviate from the range of RR7 for the first time, that FF is set, and when the range is not deviated from then on, the information of that FF is set. By storing candidate character codes in SR9, it is possible to reduce the number of candidates in the major classification. The state that there are no candidates in the major classification is a phenomenon that does not appear unless the characters are written in a very special way, but even in such a case, the character will not be rejected at the time of the major classification.

Next, there is a method to reduce rejects in the same major classification as above, but by using the data of RR7 for the input character data input to AR6.
If there are no candidate characters even after verifying with COMP8, RR
It is also effective to increase the range of the contents of 7 by about 1/8 to 1/4 and re-verify it.

(Effects of the Invention) As is clear from the above explanation, the present invention recognizes characters by recognizing characters as figures and using the distribution state of stroke line segments as a feature. It is possible to broadly classify the following. Moreover, the present invention
Those skilled in the art will easily understand that the present invention is not limited to recognition of about 100 characters such as ANK, but can also be applied to recognition of partial patterns of hiragana and kanji.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention;
Figure 2 is an example of data after smoothing, Figure 3 is an example of the data after smoothing.
FIG. 4, an example of data from which representative points are extracted from the figure, has its origin set at the upper right of a rectangle circumscribing a character. DESCRIPTION OF SYMBOLS 1...Tablet, 2...Analog-digital converter, 3...Input register, 4...Arithmetic unit including arithmetic registers, 5...Partial value register, 6...Arrangement register, 7...Reference register, 8...Comparator , 9... Candidate register.

Claims (1)

[Claims] 1. A tablet that generates data indicating the coordinates of a writing instrument when writing a character, and an online character recognition device that recognizes written characters based on the data generated by the tablet, converting analog data from the tablet into digital data. means for temporarily storing the digital data;
After performing preprocessing to determine representative points, calculate the sum of the lengths of line segments obtained by connecting the representative points in a predetermined order along the characters, and the distance between each point on the line segment and the coordinate axis. means for determining the average value of and normalizing the sum of the lengths of the line segments and the average value of the distances, and means for comparing the normalized value with a predetermined value, A character recognition general classification method characterized by classifying written characters according to the length and distribution state of line segments.