JPH0533434B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an online continuous character recognition device that recognizes one or more consecutively written characters from the handwriting of characters input online.

(Prior Art) Conventionally, a device that recognizes a string of consecutively written characters without the user instructing the end of one character includes, for example, the Japanese Patent Application No. 173421.
There was an online continuous character recognition device described in "No. 1" (hereinafter referred to as Reference (1)). Hereinafter, a conventional online continuous character recognition device will be explained using this device as an example.

This device uses continuously written characters, including characters written without breaks, as an input pattern, and recognizes the characters by pattern matching this pattern with a standard pattern stored in the recognition device in advance. .

The input character string is exchanged into a time-series pattern of direction angles a _i of line segments constituting the characters A={a _i |i=1, 2, . . . , } (1). Let this time-series pattern be input pattern A. Similarly, the standard pattern for each character is stored in the device as a time series pattern of direction angle b _j as BC = {b ⁿ _j | j = 1, 2, ..., S ⁿ , ..., J ⁿ } (2) Retained. Here, n represents a character name. Let this time series pattern be a continuous character pattern BC. Continuous character pattern BC is an isolated character pattern BI that represents the pattern from the beginning of writing to the end of one character BI = {b ⁿ _j | j = 1, 2, ..., S ⁿ } (3) BS representing the inter-character strokes from the beginning consists of BS = {b ⁿ _j | j = S ⁿ +1, S ⁿ +2, ..., J ⁿ } (4).

An end point table Ej(n)=S ⁿ (5) storing the end point S ⁿ of the isolated character pattern BI included in the continuous character pattern BC is held in the device in the same way as the continuous character pattern BC.

Connect one or more of these continuous character patterns BC,
Synthesizes a standard string pattern for matching with the input pattern. However, only one pattern to be connected at the end is an isolated character pattern BI up to the end point listed in the end point table Ej.

Here, the i-th data of input pattern A,
The distance between the j-th data of standard pattern BC is
It is expressed as the angle formed between the direction angles a _i and b _j . Let this distance be the point-to-point distance d _o (i, j)=|a _i −b ⁿ _j | (6). The inter-pattern distance D between the input pattern A and the character string standard pattern is the point-to-point distance d _o (i,
The distance between points d _o (i,
Let it be the cumulative value of j). The combination of consecutive character patterns BC that minimizes the inter-pattern distance D between the character string standard pattern synthesized in this way and the input pattern A is set as a recognition result.

This matching on the time axis can be performed by the DP matching method described in document (1). DP matching of input pattern A with continuous character pattern BC at time i is the recurrence formula go ₍ i, j) = d _o (i, j) + ming _o (i-1, j) _go (i-1, By calculating j-1) (7) for j=1, 2, . . . , the minimum cumulative distance up to time i can be obtained. However, as initial values, _go (0, 0) = 0, _go (0, j) =
Give ∽(j=1, 2,...j ⁿ ).

At the same time as this recurrence formula operation, for the path value h _o (i, j) required to trace the combination of standard patterns that gives the minimum inter-pattern distance, h _o (i, j) = | # | h _o (i-1, j): When g _o (i-1, j) is the minimum _ho (i-1, j-1): When g _o (i-1, j-1) is the minimum (8 ) performs the assignment operation. However, as the initial value h _o
Let (0, 0)=0. For the route value h _o (i, j),
When performing the recurrence formula calculation of equation (7) at times i and j,
The boundary value indicating where the boundary distance was given, that is, where the starting edge of the standard pattern corresponds to the input pattern, is stored.

Character name n, minimum cumulative distance to time i g _o (i,
The minimum value for n among j ⁿ ) (T= min ⁿ g _o (i, j ⁿ ) (9) becomes the boundary distance of the recurrence formula operation at the next time i+1. That is, _go (i, 0)=T (10). Moreover, the boundary value of the route at that time is given by h _o (i, 0)=i (11).

Also, substitute the character name nt that satisfies equation (9) and the path value h _ot (i, j ^nt ) at that time into the boundary character name N(i) and boundary path L(i), respectively, as follows. .

n(i)=nt (12) L(i)=h _ot (i, j ^nt ) (13) The boundary character name n(i) and the boundary path L( ⁱ ) include its The character name nt of the standard pattern with the minimum cumulative distance and the position of the starting edge of the standard pattern are stored. In other words, it can be seen that the standard pattern of the character name n(i) matches the portion from time L(i) to i.

On the other hand, for the last part of the input pattern A, the minimum inter-pattern distance for the isolated character pattern BI must be found. Therefore, at the end point i=I of input pattern A, find the end point S ⁿ =Ej(n) (14) of the isolated character pattern BI from the end point table Ej(n), and calculate the recurrence formula of equation (7) and (8). and the assignment operation j=
1, 2, . . . Perform only on the isolated character pattern BI portion of S ⁿ . As a result, the inter-pattern distance D
is found as D= min ⁿ _go (I, S ⁿ ) (15).

The above recognition results are obtained as follows. The above-mentioned boundary character name N(i) stores the character name n of the standard pattern whose cumulative distance is the minimum at time i. Further, the boundary value is stored in the boundary route L(i). When i=I, the value nr=N(i) (16) of the boundary character name N(i) becomes the character name nr of the standard pattern that matches the portion from time L(i) to I. Next, the value of N(i) is found in the same way by replacing the value of i with L(i). By tracing L(i) until L(i)=0, recognition results are obtained in reverse order of time.

(Problems to be Solved by the Invention) Conventionally, the point-to-point distance d _o (i, j) between the input pattern and the standard pattern has been determined using a direction angle, etc., without using position information. This is because the boundaries between characters in the input pattern are not known, and the position information of the input string and
This is because it was not possible to compare the positional information of standard patterns separated for each character. However, when recognizing the katakana letters ``ko'' and ``yu'', the only difference is the lateral position of the end point of the final stroke, and it is not possible to distinguish based on the direction angle.

As described above, conventionally, there were characters that could not be recognized because character position information could not be used. An object of the present invention is to provide an online character recognition device that solves this problem.

(Means for solving the problem) The online continuous character recognition device according to the present invention has the following features:
an input section that reads the handwriting of written characters as a time-series input pattern; normalizes the positional information of the input pattern using boundaries between characters; a normalization section that reads the handwriting of characters as a time-series input pattern; , a standard pattern storage unit for storing a standard pattern in which inter-character strokes from the end point of a character to the start point of the next character are connected; an end point table for storing the end point of the isolated character pattern; and 0 standard patterns. a recognition unit that recognizes the input pattern using positional information based on a character string standard pattern in which an isolated character pattern whose end point is the end point stored in the end point table is connected to the end of the connected pattern; Consisted of.

(Operation) In the online character recognition device of the present invention,
When matching an input character string with a standard pattern, the boundary value when taking the optimal route at time i-1, that is, the boundary between characters in the input pattern (hereinafter referred to as character boundary) is saved. Then, by setting the boundary value as the origin, the position of one character in the input character string is normalized, and the position information is used as the distance between points during matching.

In this device, the input character string is a time-series pattern of position coordinates (ax _i , ay _i ) of the input points that make up the character A = {(ax _i , ay _i ) | i = 1, 2, ..., I } Converted to (17). Let this time-series pattern be input pattern A. Similarly, the standard pattern for each character is the position coordinates (bx _i , by _i ) BC = {(bx ⁿ _j , by ⁿ _j ) | j = 1, 2, ..., S ⁿ , ..., J ⁿ } (18) is held within the device. Let this time series pattern be a continuous character pattern BC. The origin of the position coordinates of these patterns (hereinafter referred to as the position origin) is the character string or the starting point of writing the character. The continuous character pattern BC is an isolated character pattern BI that represents the pattern from the beginning to the end of one character. BI = {(bx ⁿ _j , by ⁿ _j ) | j = 1, 2, ..., S ⁿ } (19) BS representing the stroke between characters from the end of writing to the beginning of writing BS = {(bx ⁿ _j , by ⁿ _j ) | j = S ⁿ +1, S ⁿ +2, ..., J ⁿ } (20) Become.

An end point table Ej(n)=S ⁿ (21) storing the end point S ⁿ of the isolated character pattern BI included in the continuous character pattern BC is held in the device in the same way as the continuous character pattern BC.

Connect one or more of these continuous character patterns BC,
Synthesizes a standard string pattern for matching with the input pattern. However, only one pattern to be connected at the end is an isolated character pattern BI up to the end point listed in the end point table Ej.

Here, the position coordinates (ax _i , ay _i ) of input pattern A at time i are normalized as follows.
When the position coordinates of input pattern A at time K and the position coordinates of standard pattern BC at time L are optimally correlated, that is, when the cumulative distance to time (K, L) is the minimum, standard pattern BC The time H of the input pattern associated with the character boundary that is the starting edge is determined. Normalization is performed by subtracting the position coordinates of input pattern A at time h from the position coordinates at time i. In other words, if the character boundary that gives the minimum cumulative distance matched up to time (K, L) is h = h _o (K, L) (22), then the position coordinate (hereinafter referred to as (referred to as normalized input pattern) becomes (ax _i , ax _h , ay _i −ay _h ).

The point-to-point distance between the normalized input pattern at time i and the standard pattern of character name N (bx ⁿ _j , by ⁿ _j ) at time j is d _o (i, j, K, L) = | (ax _i , ax h _o(K , _L) )−bx ⁿ _j |+|(ay _i −ay h _o(K , _L) )−by ⁿ _j | (23).

The inter-pattern distance D between the input pattern A and the character string standard pattern is the point-to-point distance d _o (i, j,
The distance between points d _o when input pattern A and character string standard pattern are aligned on the time axis so that the cumulative value of K, L) in the time direction is the minimum
Let it be the cumulative value of (i, j, K, L). The combination of consecutive character patterns BC that minimizes the inter-pattern distance D between the character string standard pattern synthesized in this way and the input pattern A is set as a recognition result.

This matching on the time axis is performed using the DP matching method. DP matching of input pattern A with continuous character pattern BC at time i is a recurrence formula g _o (i, j) = min | # g _o (i-1, j) + d _o (i, j, i-1, j) g _o (i-1, j-1) + d _o (i, j, i-1, j-1
)(24) However, _go (0, 0) = 0, _go (0, j) = ∽ (j = 1,
2, . . . , J ⁿ ) for j=1, 2 ^, .

At the same time as this recurrence formula operation, the path value h _o (i, j) required to trace the combination of standard patterns that gives the minimum inter-pattern distance is substituted as in equation (8) in the same way as in the prior art. An operation is performed. Further, h _o (i, j) is also used as a boundary value h for normalizing the input pattern A at the next time i+1.

The subsequent recognition processing is performed in exactly the same manner as in equations (9) to (16) of the prior art, and finally, the recognition result is obtained and the process ends.

As described above, by using the boundary value as the positional origin when matching an input character string and a standard pattern, it is possible to use positional information, and it is now possible to recognize characters that could not be recognized using direction angles in the past. I can do it.

Furthermore, the distance between points used in the present invention is also d _o (i, j, K, L) = {[(ax _i , ax ho _(K , _L) ) − bx ⁿ _j ] ² + [(ay _i , ay h _o(K , _L) )−bx ⁿ _j ] ² } ^1/2 (25) Various position information can be used.

In addition, as a standard pattern “Patent Application 1986-110743”
The present invention can be used in exactly the same way when using the continuous character pattern and isolated character pattern described in "Japanese Patent Application No. 1982-53397" and the isolated character pattern and intercharacter pattern described in "Japanese Patent Application No. 1983-53397."

In addition to equation (24), the recurrence formula used for DP matching is also included in the Nikkei Electronics magazine, November 7, 1988 issue, pages 171 to 208, "A two-stage method for efficiently recognizing continuously uttered word sounds. Various recurrence formulas published in the literature entitled "DP Matching" (Oie) can be used.

(Example) An example of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention.

The standard pattern storage unit 130 contains a continuous character pattern BC corresponding to equation (18), and an end point table Ej.
In 200, end points S ⁿ corresponding to equation (21) are stored for each character name n (n=1, 2, . . . , N).

A G _memory 170 and an H memory 180 are prepared in which the recurrence formula value go (i, j) and the path value _ho (i, j) are stored and whose addresses are specified by time j and character name n. Similarly, the T memory 210, N memory 230, and L memory 210, which store the boundary distance, boundary character name, and boundary route corresponding to equations (9), (12), and (13), and are addressed by time i.
Each memory 220 is prepared.

Character string data input from the tablet 100 is converted into time series data of position coordinates in a preprocessing section 110. This time series data is sent to the input pattern buffer 120, and the input pattern (ax _i ,
ay _i ). At the same time, j = 1, 2, ..., J and n = 1, 2, ... are stored in the G memory 170.
..., ∽ (the maximum value allowed by the memory) is set for N. Similarly, T(0) is stored in the T memory 210.
=0, T(i)=∽(i=1, 2, . . . , I) are set, and the H memory 180 is cleared with all 0s. Furthermore, the end point I of the input pattern is determined by the preprocessing unit 11.
0 is sent to the end point detection unit 260 and held.

When the above initialization is completed, the control unit 250 outputs the time signal i as 1, 2, . . . , I, and processing is performed in synchronization with this. At this time i, the character designation signal n from the control unit 250 is 1,
It changes like 2,...,N. Input patterns (ax _i , ay _i ) are sent from the input pattern buffer 120 to the normalization unit 270 according to the time signal i. The normalization unit 270 includes a DP matching unit 1, which will be described later.
From the boundary value signals h1 and h2 from 60, the input pattern is normalized as shown in equation (23). Then, the normalized input pattern a1 _i corresponding to h1 and h2,
a2 _i is output to the distance calculation section 140. at the same time,
The standard pattern b n _{j read out from the standard pattern storage section 130 using the character designation signal n and the standard pattern time signal j} ^is also sent to the distance calculation section 140 .
The distance calculation unit 140 uses the normalized input patterns a1 _i ,
The inter-point distances d1 and d2 between a2 _i and the standard pattern b ⁿ _j are calculated as shown in equation (24). These point-to-point distances d1 and d2 are sent to the DP matching section 160, and the minimum cumulative distance is calculated for the character name n.

Next, the operation within one cycle for n will be explained. A detailed configuration example of the DP matching unit 160 is shown in the second example.
As shown in the figure. When the character designation signal n is specified, T(i-1) is read out from the T memory 210 onto the signal line T2, written into the g1 register 1601, and written into the h1 register.
A numerical value i-1 is written into register 1606.

Subsequently, when the standard pattern time signal j is set to 1, the contents of the g1 register 1601 become g2.
The contents of the h1 register 1606 are transferred to the h2 register 1607. At the same time, G(n,
1) is read out and written into the g1 register 1601, and H(n, 1) is read out from the H memory 180 and written into the h1 register 1606. At this point, the values held in g2 register 1602 and h2 register 1607 are the boundary conditions g(i-1, 0) and h(i-
1, 0). Also, the values held in g1 register 1601 and h1 register 1606 are _go (i-1, 1) and _ho (i-1,
1).

Thereafter, the calculations of equations (24) and (8) are repeatedly executed while the standard pattern time signal j is sequentially increased. The operation when the standard pattern time signal generally takes the value j will be described below. At this point g
1・Register 1601, g2・Register 1602
are respectively g1=g _o (i-1, j) and g2=g _o
(i-1, j-1) is also h1/register 16
06, h2 register 1607 has boundary values h1=h _o (i-1, j) and h2=hn (i-1, j-
1) is maintained. The values of h1 and h2 are sent to the normalization unit 270, which calculates the position coordinates (ax _h1 , ay _h2 ) of the input pattern at time i=h1 from the input pattern buffer 120, and calculates a1 as shown in equation (23). The input pattern is normalized as _i (ax _i −ax _h1 , ay _i −ay _h1 ). Similarly for h2, a
2 (ax _i −ax _h2 , ay _i −ay _h2 ) and perform normalization.
These normalized input patterns a1, a2 _i are sent to the distance calculating section 140 as described above, and the point-to-point distances d1, d2 from each standard pattern b ⁿ _j are calculated.

d1 and d2 sent to the DP matching section 160
are calculated by adders 1603 and 1604, respectively,
g1 register 1601, g2 register 1602
The sum of g1 and g2 held in g1+d1,
g2+d2 is calculated and sent to minimum value detection section 1608. In the minimum value detection section 1604, the signal g1+
The minimum value of d1 and g2+d2 is determined and the minimum value signal
Output as gm. At the same time, this minimum value is g1
A switch signal c having a value of 1 or 2 is output depending on whether the value is +d1 or g2+d2.

The minimum value signal gm thus obtained becomes the signal sg2, which is the calculation result _go (i, j) of recurrence formula (24).
Further, the signal sg2 is written into the G memory 170 as a new G(n,j). The multiplexer 1605 selects signals h1 and h2 from the h1 register 1606 and h2 register 1607 under the control of the switch signal c. That is, h1 is selected for c=1, and h2 is selected for c=2.
As a result, the output signal sh2 is the calculation result of equation (8)
It becomes h _o (i, j). This signal sh2 is written into the H memory 180 as a new value of H(n,j).

When the above process is repeated until j=J ⁿ , g _o
(i, J ⁿ ) and _ho (j, J ⁿ ) are obtained. These signals are output as signals T1 and h, respectively.
Thus, the operation of the DP matching section 160 is completed once.

Signal T1 from DP matching section 160 and signal T2 from T memory 210 are sent to comparison circuit 190 and compared. When T1<T2, comparison circuit 1
A write pulse Wp is output from 90. As a result, the signal T1= _go (i,
J ⁿ ) is stored in the L memory 220, and h= _ho (i, J ⁿ ) is stored in the L memory 220.
Then, character designation signals n are written in the N memory 230 as T(i), L(i), and n(i), respectively. On the other hand, when T1>T2, these entry processes are not performed.

Processing similar to the above is repeated while the character designation signal n is changed to 1, 2, . . . , N, and the processing at time i ends. As a result, T memory 21
0 holds the minimum value of g _o (i, J ⁿ ) with respect to n, as in equation (9). Furthermore, this process is performed at time i
Regarding 1, 2, ..., I is repeated,
The processing at end point I is different from before.

The end point detection unit 260 compares the time signal i with the end point I, and when i=I, the end point table Ej20 is
Send end point signal e to 0. When the end point table Ej200 receives the end point signal e, the standard pattern storage section 1
30, the end point S ⁿ for the character designation signal n is sent. Using this sent value, the standard pattern storage section 130 sets the end point of the standard pattern to be output to the distance calculation section 140 as S ⁿ . In other words, the data for j of the standard pattern is b ⁿ _j (j=1, 2, ...,
S ⁿ ). Subsequent distance calculations and recurrence formula calculations are done using j
=1, 2, ..., S. The subsequent processing is the same as before, and by changing the character designation signal n to 1, 2, ..., N, the time I
The process at ends. In the above manner, the inter-pattern distance D has been found as the value of T(I).

Subsequently, a determination process for obtaining a recognition result is started. The determination unit 240 refers to L(i) and N(i) from the L memory and the NM memory and outputs the recognition result nr as in equation (16). Give the value of i=I to N memory,
The value of N(i) at that time is output as a result, and then i
The value of is output to the L memory to obtain the value of L(i). The value of L(i) is newly given as i to the N memory, and the value of N(i) is similarly output as a result. By repeating this process until L(i)=0, the recognition results nr are output in the reverse order of time i.

(Effects of the Invention) According to the present invention, it is possible to discriminate characters based on differences in positional coordinates, which was impossible in conventional online continuous character recognition devices, and an online character recognition device with a higher recognition rate can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention;
FIG. 2 is a detailed configuration diagram of the DP matching section in this embodiment. 100...tablet, 110...pretreatment section,
120...Input pattern buffer, 130...Standard pattern storage unit, 140...Distance calculation unit, 15
0... Distance storage section, 160... DP Matsushig section,
170...G memory, 180...H memory, 19
0... Comparison section, 200... End point table, 210
...T memory, 220...L memory, 230...
n memory, 240...determination unit, 250...control unit, 260...end point detection unit, 270...normalization unit.

Claims (1)

[Claims] 1. An input section that reads the handwriting of written characters as a time-series input pattern; A normalization section that normalizes the positional information of the input pattern using boundaries between characters; ,
a standard pattern storage unit that stores a standard pattern that connects inter-character strokes from the end point of a character to the start point of the next character; an end point table that stores the end point of the isolated character pattern;
Based on a character string standard pattern in which an isolated character pattern whose end point is the end point stored in the end point table is concatenated at the end of a pattern in which two or more are connected,
and a recognition unit that recognizes the input pattern using positional information.