JPH0432431B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an online handwritten character recognition device that inputs and recognizes handwritten characters written on a tablet online.

Conventionally, when recognizing handwritten characters input online using a pattern matching method, one stroke is taken as a unit from when the pen is lowered to when it is raised, that is, from when the pen is turned on until it is turned off. That is, the patterns for each stroke are compared to determine the degree of similarity, and the degree of similarity between relative position patterns between strokes is also determined, and recognition is performed based on these degrees of similarity. When such a recognition method is used, there is a drawback that if the number of strokes of an input character does not match a predetermined standard pattern, it cannot be recognized. One way to overcome this drawback was to prepare standard patterns for all possible numbers of strokes.

Normally, in handwritten characters, if the person writing is the same, the order of strokes rarely changes. However, the speed of writing often makes the difference between writing letters in succession or not. As a result, there is a possibility that many types of stroke numbers may occur for one character. Therefore, the method of preparing all the patterns in advance requires a large amount of standard pattern memory and calculation, making it difficult to take into account all changes in the number of strokes.

An object of the present invention is to stably recognize handwritten characters input online without being affected by changes in the number of strokes.

In order to achieve this object, the online handwritten character recognition device according to the present invention includes a tablet unit into which handwriting of handwritten characters is input as positional information, and a writing start and a writing end of one character based on the positional information input from the tablet unit. A start/end end detection section that detects, a conversion section that converts the position information into a character pattern at a predetermined interval regardless of whether the pen is on or off, and a standard pattern that stores a predetermined character pattern as a standard pattern. It is characterized by comprising a storage unit, and a pattern matching unit that identifies the type of input character by time-axis normalized matching between the character pattern of the input handwritten character and the standard pattern. shall be.

Next, before describing an embodiment according to the present invention, its operating principle will be explained. Suppose you want to write the word ``mouth'' in handwriting. Figure 1 and 2 depending on writing speed etc.
As shown in FIG. 3, there can be three strokes, three strokes, two strokes, and one stroke.

It is necessary to match these three patterns as being the same. The data inputted from the tablet is usually position data obtained by sampling only the part where the pen is on at constant time intervals, such as the sample point 41 indicated by a black circle in FIG. The sample points 41 are located at irregular intervals. Apply this data along the handwriting as shown in Figure 5, turn on the pen,
Regardless of whether it is off or not, it is converted into sample points that are positioned at equal intervals, and data on the movement direction between these sample points is used as a direction pattern 51 (indicated by 'ⓓ' in the figure). In this case, the online handwritten character recognition device according to the present invention treats the pen-off portion in the same way as the pen-on portion.

In other words, the position of the pen is detected even when the pen is off and is used as position data. However, in this case, there is no guarantee as to what kind of trajectory the pen will take when the pen is turned off, especially when writing slowly.
A good idea is to perform a linear interpolation from the point where the pen turns off to the next point where it turns on. This linear interpolation method will be used below. Direction pattern 51 in Figure 5, which is a pattern of the character "mouth" written in three strokes,
A direction pattern 61 in FIG. 6, which is a pattern drawn in one stroke, is matched. In Figure 7,
The time series pattern 71 of the arrow '〓' written above is one directional pattern when drawn with three strokes shown in Figure 5.
It is drawn on a straight line, and the time series pattern 72 of the arrow 'ⓓ' also drawn below is the directional pattern shown in FIG. By matching patterns connected vertically by straight lines 73 in FIG. 7 through time axis normalization matching, these two patterns can be matched as substantially the same pattern. As a result, if the pattern written using three strokes is used as a standard pattern, matching can be performed even if the number of strokes changes.

An embodiment according to the present invention will be described below. FIG. 8 is a block diagram of an embodiment according to the present invention. The position information input through the tablet 81 is sent to the start/end detection unit 8 as coordinate value data xy on the tablet.
2, data for one character is extracted.

The data from the beginning to the end is processed as one character of data.

To detect the start and end of a character, for example, as shown in Figure 9, a square 91 for writing a character is determined in advance, and a point 92 in the square 91 where the pen is turned on for the first time is determined as the starting point. , when the pen is turned on outside the square, the end point is the point 93 where the pen was last turned on inside the square. It is assumed that the pen position data when the pen is off is not captured, and that the data is connected by a straight line between when the pen is turned off and when it is turned on again.

If data of n sample points are obtained from the beginning to the end of one character, they are x ₀ , x ₁ , ..., x _i , ..., x _j , ..., x _o y ₀ , y ₁ , ..., It can be written as y _i ,…, y _j ,…, y _o . Using these coordinate value data, the movement amount Δl _i between the adjacent data is calculated in the movement amount calculation section 83 in FIG. This value can be obtained, for example, as follows: Δl _i =(x _i+1 −x _i ) ² +(y _i+1 −y _i ) ² (1). This value is integrated by the movement amount integration section 84. If the sum of the movement distance from the i-th point to the j-th point from the starting end is larger than the predetermined movement amount L, that is, when L> _j 〓 ^t=i △l ₁ (2), from the i-th point Between the j+1th point, i
Point to interpolate the point whose movement distance from the th point is L
Find x _j ′ and y _j ′. For example, for linear interpolation these points are becomes. In this way, the coordinate data of the point for each movement distance L can be obtained.

The coordinate storage unit 86 stores data of the previous point, ie, x _i , y _i . Through the processing up to now, points x _{' j and} y' _j , which are moved by the amount of movement L along the handwriting from these points x _i and y _i , have been found. Conversion section 85
Then, for example, the direction angle Θ _k between the two points is determined as data representing the relative positional relationship between the two points.

Θ _k is expressed as Θ _k =ton ^-1 y′ _j −y _i /x′ _j −x _i (0≦Θ _k <3
60°)(4). The coordinate storage unit 86 newly stores x′ _j , y′ _j
The value of is stored. When i=0, x ₀ in advance,
Remember the value of y ₀ .

In this way, the value of Θ _k is determined up to j+1=n.

As a result, it is possible to obtain a time-series pattern of movement direction angles Θ ₀ , Θ ₁ , . . . , Θ _{k ,} .
A pattern matching unit 87 performs time axis normalization matching on this pattern with a standard pattern previously stored in a standard pattern storage unit 88. As standard patterns, time series patterns of movement direction angles RΘ ₀ , RΘ ₁ , ..., RΘ _l , ..., which are determined in advance for characters to be recognized, are used.
Use _RΘN . As time axis normalization matching methods, there are the following methods. First, the inter-pattern distance d k, _l between the k-th data Θ _k of the input character pattern and the l-th data RΘ _l of the standard pattern is calculated as d _k,l = |Θ _k −RΘ _l | (|Θ _k −RΘ _l ｜≦
180°) d _k,l = |Θ _k −RΘ _l | (|Θ _k −RΘ _l |≦
180°) 360°−｜Θ _k −RΘ _l | (｜Θ _k −RΘ _l | > 180°)
(5). Then, the pattern is patterned using dynamic programming as described in detail in "Comparison of various DP matching methods in speech recognition", Material S73-22 of the Acoustical Society of Japan's Speech Research Group (issued December 1970). Find the distance D, D=min l=l(K) [ _M 〓 ^k=1 d _k,l ] (6).

In equation (6), min j=j(i) is for function j(i) [ ]
means to minimize the value within. Function j(i) represents a function that corrects time axis distortion between both patterns.

As a specific method for calculating equation (6), the DP method can be used as follows.

Under the initial condition g _0,0 = d _0,0 (7), the recurrence formula g _k,l = d _k,l + ming _k-1 , l g _k-1 , l- ₁ g _k-1 , l- ₂ (8) in the range of 1≦k≦M and 1≦l≦N, D=g _M,N (9) is found. The character of the pattern with the minimum D becomes the recognition result, and is output as a signal R from the pattern matching section 87.

Although one embodiment of the present invention has been described above, the present invention is not limited to this embodiment. For example, as the time-axis normalized matching method, a time-axis normalized matching method for information-compressed patterns as described in Japanese Patent Application No. 56-032060 entitled "Pattern Matching Apparatus" can also be used. In this case, the time series pattern △l ₀ , △l ₁ , ..., △l _i , ..., △l _o-1 of the amount of movement △ _l i between the input position data,
Using the time series pattern Θ ₀ , Θ ₁ , ..., Θ _i , ..., Θ _o-1 of the moving direction Θ _i between adjacent data (8)
_{Instead of the recurrence formula of Eq .} Just solve the recurrence formula.

As described above, according to the present invention, it is possible to stably recognize online handwritten characters without being affected by changes in the number of strokes, and the effects are significant.

[Brief explanation of drawings]

Figures 1, 2, and 3 are diagrams showing examples of the handwriting patterns of the character ``mouth'' written with 3 strokes, 2 strokes, and 1 stroke, respectively, and the parts indicated by solid lines in the figures are The area where the pen is on is the area where the pen is on, and the area indicated by the broken line is where the pen is off. Figure 4 is a diagram showing an example of the position data when the handwriting of the character "mouth" is input from a tablet, Figure 5 is a diagram showing an example of the direction pattern obtained from the character shown in Figure 1, and Figure 6 is a diagram showing an example of the directional pattern obtained from the character shown in Figure 1. The figure shows an example of a directional pattern obtained from the characters shown in Fig. 3, Fig. 7 shows an example of time-series matching of directional patterns, and Fig. 8 is a configuration diagram of an embodiment according to the present invention. FIG. 9 is a diagram showing an example of detecting the start and end of one character. In the figure, 41 is a sample point, 51 and 61 are directional patterns, 71 and 72 are time series of directional patterns, and 73 is a correspondence relationship in the case of matching. 81 is a tablet, 82 is a start/end detection section, 83 is a movement amount calculation section, 84 is a movement amount integration section, 85 is a conversion section, 8
6 is a coordinate storage unit, 87 is a pattern matching unit, 8
Reference numeral 8 indicates a standard pattern storage section, 91 indicates a square, 92 indicates a start end, and 93 indicates an end.

Claims (1)

[Scope of Claims] 1. A tablet section for inputting handwriting of handwritten characters as positional information, a start/end detection section for detecting the beginning and end of one character from the positional information input from the tablet section, A conversion unit that converts position information into a character pattern consisting of a data string sampled at a predetermined interval regardless of whether the pen is on or off, and a standard pattern memory that stores predetermined character patterns as standard patterns. and a pattern matching unit that identifies the type of input character by time-axis normalized matching between the character pattern of the input handwritten character and the standard pattern. Device. 2. The converting unit converts the handwriting of the part where the pen is off into a character pattern by linear interpolation between the point where the pen is turned off and the point where the pen is next turned on. The online handwritten character recognition device according to scope 1.