JPH0620104A - Character pattern generation device and character recognition device using the same - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a character pattern generation device used for character recognition or the like that can generate a natural character pattern close to human handwritten characters. [Configuration] A reference character trajectory generation unit 11 that generates a reference character trajectory indicating a representative pattern of a specified character type, a character type designation unit 12 that specifies a character type for which a reference character trajectory is to be generated, and a motion of a human hand. A dynamic characteristic model having a plurality of parameters, which simulates the characteristic, is included, and a dynamic characteristic simulating unit 13 for a hand that generates a character pattern in which a reference character trajectory is simulated by handwriting and a plurality of parameters. A parameter setting unit 14 for individually variably setting, and a dynamic characteristic simulation unit 13 according to a plurality of dynamic characteristic models having parameters set by the parameter setting unit 14 for each character type designated by the character type designation unit 12 are generated. Character pattern storage unit 16 for storing character patterns as data
Have.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character pattern generating device for generating a character pattern used for learning data of characters in a character recognizing device and a character recognizing device using the same.

[0002]

2. Description of the Related Art A character recognition device is a device for recognizing the content of a character by reading a printed character or a handwritten character, and is used in an optical character reader (OCR) and an automatic mail reading device.

In such a character recognition apparatus, it is important to prepare as many samples of character patterns as possible as the learning data for character recognition, which are considered as possible recognition targets, in order to improve the recognition ability. . In particular,
Since handwritten characters are significantly deformed, a character recognition device that recognizes handwritten characters must collect a large number of character patterns, ranging from thousands to tens of thousands. And costs money. In addition, handwritten characters have regional characteristics, and character patterns collected in one area may not be usable in other areas. In that case, it is necessary to newly collect character patterns. As described above, the collection of character patterns significantly hinders the development efficiency of the current character recognition device.

In order to solve this problem, attempts have been made to artificially create a character pattern. This is to prepare a pattern group of a certain representative character and to transform it to generate a new character pattern group. As a result, in the character recognition device, it is considered that the recognition ability for deformed character patterns other than the typical character pattern is improved to some extent.

As an example of a character pattern generation device based on such an idea, a method of transforming a representative character pattern by a random number is known. This is to obtain a deformed character pattern by adding noise by random numbers to the representative character pattern, as shown in FIG. However, the character pattern deformed by this method has unnatural irregularities as shown in FIG. 7, and is not practical even if it is similar to the human handwritten deformation. In many cases. Therefore, even if a large number of character patterns transformed by this method are created, it is not very useful for improving the recognition ability.

In general, when a character pattern is actually used in a character recognition device or the like, it is necessary to perform a normalization process for adjusting the character pattern to a predetermined size. In the above method of generating a character pattern by deforming a representative character pattern by adding noise due to random numbers, a large deformation far from the deformation of human handwritten characters is added, and the size of the character pattern is equivalently large. In such a character pattern, as a result of normalization, the substantial size of the character pattern is, on the contrary, smaller than the representative character pattern. That is, even if the normalization is performed, a character pattern having the expected size may not be obtained.

On the other hand, a method has also been proposed in which the character pattern is deformed so as to maintain the smoothness of the trajectory of the character without causing unnatural irregularities as shown in FIG. Expression ”IEICE Spring National Convention, D-433, 1989). However, since this method also uses a convenient transformation method unrelated to the characteristics of human handwriting, it is still unnatural transformation.
It was difficult to generate an effective character pattern that leads to a significant improvement in cognitive ability.

[0008]

As described above, in the conventional character pattern generating apparatus for generating a character pattern which is deformed by adding noise to a typical character pattern, what is called deformation by human handwriting? When applied to character recognition, it is not possible to expect an improvement in recognition ability because an unrelated unnatural character pattern is generated, and when the generated character pattern is normalized, the actual character pattern There was also the problem that the size would change.

The object of the present invention is to solve the above-mentioned problems of the prior art, and it is possible to generate a natural character pattern close to human handwritten characters, and to normalize the character pattern. It is an object of the present invention to provide a character pattern generation device in which the substantial size of the character pattern does not change. It is another object of the present invention to provide a character recognition device that uses this character pattern generation device to improve the recognition ability for handwritten characters.

[0010]

In order to solve the above-mentioned problems, a character pattern generation device of the present invention is a reference character for generating a reference character trajectory of a two-dimensional coordinate expression showing a representative pattern of a designated character type. Trajectory generating means, a character type designating means for designating a character type for generating the reference character trajectory with respect to the reference character trajectory generating means, and a dynamic characteristic model simulating a dynamic characteristic of a human hand are inherent, The reference character trajectory generated by the character trajectory generating means is simulated by handwriting to generate a deformed character trajectory of a two-dimensional coordinate representation as a character pattern, and a hand dynamic characteristic simulating means.

Further, the present invention is characterized in that the dynamic characteristic model is provided with a plurality of parameters in the dynamic characteristic model, and the parameter setting means is provided for variably setting these parameters individually.

Further, according to the present invention, for a plurality of character types designated by the character type designating means, a character pattern generated by the dynamic characteristic simulating means in accordance with a plurality of dynamic characteristic models each having a parameter variably set by the parameter setting means. Is stored as data.

The character recognition apparatus of the present invention is a learning means for performing learning for character recognition on the above-mentioned character pattern generation apparatus by using the data of the character pattern accumulated in the accumulating means, and the learning means of this learning means. It is configured by combining recognition means for recognizing a character to be recognized based on the learning result.

[0014]

According to the dynamic characteristic model of the human hand, that is, the dynamic characteristic model simulating the characteristic of the hand movement when the human actually writes the character, the handwriting is performed by handwriting with respect to the reference character trajectory of the designated character type. The deformation is imitated so that a character pattern having a natural deformation close to that of a handwritten character is generated. In addition, by variably setting a plurality of parameters included in the dynamic characteristic model individually, various character patterns can be obtained for the same character type, and it is possible to sufficiently cope with regional characteristics of handwritten characters.

Therefore, a large number of the character pattern data thus generated are accumulated as character samples for producing learning data for character recognition, learning is performed by a neural network, etc., and the learning result is used for character recognition. By doing so, the recognition ability for handwritten characters is improved.

Further, since the character pattern generated in this manner does not involve extremely large deformation such as the method of adding noise, even if the normalization is performed prior to learning, the character pattern is substantially converted. The size does not change, and the recognition rate increases from this point as well.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a character pattern generation device and a character recognition device using the character pattern generation device according to an embodiment of the present invention. The character pattern generation device 10 includes a reference character trajectory generation unit 11, a character type designation unit 12, and a hand dynamic characteristic simulation unit 1.
3, a parameter setting unit 14, a display unit 15, and a character pattern storage unit 16.

The reference character trajectory generation unit 11 is represented by two-dimensional coordinates indicating a typical pattern of the character type designated by the character type designation unit 12 (for example, a typeface, a regular typeface or a typeface pattern corresponding thereto). Generate a reference character trajectory. More specifically, as shown in FIG. 2, when the trajectory of a character is represented by a two-dimensional coordinate composed of a horizontal axis coordinate x and a vertical axis coordinate y, the reference character trajectory 11 is a trajectory of a typical character pattern ( (Reference trajectory) is generated in the form of a time function indicating the abscissa and ordinate coordinates of the reference trajectory. It should be noted that the character type means the type of each character, and for example, if it is a number, it is “0”, “1”, “2”, ... Character type designation part 1
Reference numeral 2 selectively designates the character types of the reference character trajectories to be generated by the reference character trajectory generation unit 11 one by one.

The reference character trajectory generated by the reference character trajectory generator 11 is input to the hand dynamic characteristic simulation unit 13 (hereinafter, simply referred to as dynamic characteristic simulation unit). This dynamic characteristic simulation unit 13
Is a dynamic model that simulates the dynamic characteristics of the human hand, that is, the characteristics of the hand movement when a person writes a character. The deformed character trajectory expressed by the two-dimensional coordinates is generated as a character pattern. Further, the dynamic characteristic model in which the dynamic characteristic simulating unit 13 is included includes a plurality of parameters, and the setting of these parameters is variable by the parameter setting unit 14.

The character pattern generated by the dynamic characteristic simulation unit 13 is displayed on the display unit 15 so that the user can confirm it.
The data is displayed on the RT and the like, collected in the character pattern storage unit 16 for use in character recognition, and stored as data together with the attribute information indicating the character type from the character type designation unit 12.

The character recognition device is a character pattern generation device 1
It is configured by combining the character reading unit 17, the recognition unit 18, and the learning unit 19 with respect to 0. These will be described later.

FIG. 3 shows a concrete example of the configuration of the reference character trajectory generator 11. In FIG. 3, the representative character locus memory 3
In 1, a typical character locus of each character type, for example, in the case of a numeral, a typical locus of each character of “0” to “9” is stored as a discrete value sequence. That is, in the representative character locus memory 31, the horizontal axis coordinate and the vertical axis coordinate x, which are the reference trajectories, are set.
Numerical sequence {(x _r (iΔt), y _r (iΔt)); i = 1 to N) obtained by discretizing _r and y _r with respect to time with a discretization width Δt is 2
Digital data represented by a base number (hereinafter referred to as locus data) is stored. Here, i represents the discretization time number, and N represents the number of data.

When a character type is designated by the character type designation unit 12, the representative character locus memory 31 transfers the locus data corresponding to the character type to the temporary memory 32. The locus data stored in the temporary memory 32 is input to the multiplexer 33, and according to the clock from the clock generator 34,
i = 1, 2, ... N are transferred to the registers 35 and 36 in this order. That, x _r (iΔt) to the register 35, y _r (iΔt) are transferred respectively to the registers 36. Register 3
The data held in 5, 36 are stored in the D / A converters 37, 3
Are converted into analog values by 8 and the time function x _r
The data is output as (t) and y _r (t) data.

FIG. 4 shows an example of the time function x _r (t). As shown in this figure, the function indicating the reference character trajectory output from the reference character trajectory generation unit 11 changes stepwise, but if the discretization width Δt is made sufficiently small, a smooth change can be obtained. it can.

Next, the dynamic characteristic simulation section 13 will be described in detail. Now, when the trajectory of the character is represented by the two-dimensional coordinates of x and y as shown in FIG. 2, the dynamic characteristic model in which the dynamic characteristic simulating unit 13 is present is expressed by the following equation.

Mx ″ = ε (x _r −x) −rx ′ (1) my ″ = ε (y _r −y) −ry ′ (2) where x: horizontal axis coordinate of the character pattern y: the character Vertical axis coordinate of the pattern x _r : Function indicating the horizontal axis coordinate of the reference trajectory y _r : Function indicating the vertical axis coordinate of the reference trajectory m: Coefficient indicating the inertial strength of the hand r: Between writing instrument and paper Friction coefficient ε: coefficient indicating restoring force to the reference trajectory x ′: first-order differential coefficient of x y ′: first-order differential coefficient of y x ″: second-order differential coefficient of x y ″: second-order differential coefficient of y

Here, the functions x _r , y _r (hereinafter, x _r (t), y _r (t) only when it is clearly specified that they are time functions.
Is a function of the reference trajectory given from the reference trajectory generation unit 11, and this represents the trajectory of the character virtually drawn on the head when a person writes a character. Since the second-order differential coefficients x ″ and y ″ of x and y correspond to the acceleration when the writing instrument moves on the paper, the left sides mx ″ and my ″ of the equations (1) and (2) are
Represents the force applied manually to a piece of paper through a writing instrument. Also,
The first term on the right side of Eqs. (1) and (2) represents the force with which a human hand returns to the reference trajectory when writing a character. Since the first-order differential coefficients x ′ and y ′ of x and y correspond to the speed at which the writing implement moves on the paper, rx ′ and ry ′ in the second term on the right side represent the frictional force between the writing implement and the paper. It will be.

That is, when a person writes a character on paper by hand, he / she is aware of the shape of the character he / she wants to write (reference character trajectory), and tries to move the writing instrument along the shape.
Restorative force (the first term on the right side) that always tries to return to the reference trajectory is applied to the writing instrument. Then, the frictional force (second term on the right side) acts in a direction to reduce this restoring force. That is, equations (1) and (2) show that the force actually applied to the paper by the writing instrument when a person writes a character is the value obtained by subtracting the frictional force from the restoring force. You can see that it reflects well.

FIG. 5 is a concrete configuration example of the dynamic characteristic simulation unit 13 in which the dynamic characteristic model represented by the equations (1) and (2) is embedded.
In this example, it is realized by an analog arithmetic circuit.
In FIG. 5, integrators 41 and 51 respectively integrate the second-order differential coefficients x ″ and y ″ of x and y to obtain the first-order differential coefficient x ′,
Output y '. The integrators 42 and 52 further integrate these first-order differential coefficients x ′ and y ′, and output x and y.
The inverting amplifiers 43 and 53 have a gain of -r / m, and x ',
Amplifies y'and outputs -rx '/ m and -ry' / m. The gain -1 inverting amplifiers 44 and 54 are -x,
-Y is output respectively. The adders 45 and 55 are -x,
Each of the -y outputs and each of the horizontal axis and the vertical axis coordinates x _r , y _r of the reference trajectory given from the reference trajectory generator 11 are added to output x _r -x, y _r -y, respectively. Gain ε
/ M amplifiers 46 and 56 amplify x _r −x and y _r −y, respectively, to obtain ε (x _r −x) / m, ε (y _r −y) / m.
Is output. The adders 47 and 57 respectively output the outputs ε (x _r −x) / m and ε (y _r −y) / m of the amplifiers 46 and 56 and the outputs −rx ′ / m and −ry of the amplifiers 43 and 53. ′ /
Each of m is added to obtain ε (x _r −x) / m−rx ′ /
Output m, ε (y _r −y) / m−ry ′ / m, respectively. The outputs of these adders 47 and 57 are equal to the second-order differential coefficients x ″ and y ″ of x and y, respectively.
It becomes 1 input.

Each parameter (m, r, ε) of the dynamic characteristic model existing in the dynamic characteristic simulating section 13 can be individually changed by the parameter setting section 14. The parameter setting unit 14 also performs initial setting of each amplifier in FIG.

With such a configuration, the horizontal axis coordinate x and the vertical axis coordinate y in the equations (1) and (2) are used as the information of the character pattern in which the deformation by handwriting is simulated according to the dynamic characteristic model. It is output from 13.

From the horizontal axis coordinate x and the vertical axis coordinate y output from the dynamic characteristic simulating unit 13, the locus of the generated character pattern is displayed on the display unit 15 in a form that can be seen and understood by humans. On the other hand, the character pattern accumulating unit 16 converts the horizontal axis coordinate x and the vertical axis coordinate y output from the dynamic characteristic simulating unit 13 into digital data by an A / D converter, contrary to the reference trajectory generating unit 11, for example. And store it in memory.

FIG. 6 shows an example in which the character "4" is generated from the character pattern generator 10 according to the configuration of the present embodiment, and (a) is the reference character generated by the reference character trajectory generator 11. Trajectories, (b) and (c) are examples of character patterns output from the dynamic characteristic simulating unit 13 and displayed on the display unit 15. Here, two types of character patterns are shown, but the respective parameters used in the dynamic characteristic simulating unit 13 are variously changed by the control of the parameter setting unit 14 to make the same character type similar to various handwritten characters. It is possible to generate character patterns.

Next, the character pattern generation device 10 described above.
The case where is applied to the character recognition device will be described. Figure 1
The character reading unit 17 in is configured by using an image scanner or the like, reads a character to be recognized, and sends it as image data to the recognition unit 18. In the recognition unit 18, the learning unit 1
The character to be recognized is recognized based on the learning result of 9. Here, a case where a neural network is used as the recognition unit 18 and the learning unit 19 in FIG. 1 will be described.

The learning unit 19 uses the character pattern data (given by the two-dimensional coordinates of x and y) from the character pattern storage unit 16 and the attribute information indicating the character type of the character pattern data to recognize the character recognition in the recognition unit 18. Learn for. Specifically, when the neural network forming the learning unit 19 has, for example, a three-layer perceptron structure including S, A, and R layers, the character pattern data from the character pattern storage unit 16 is input to the S layer, which is an input layer. Then, the learning operation is performed by a well-known algorithm such as the back propagation method using the attribute information indicating the character type as the teaching input to the R layer which is the output layer.

That is, since the learning unit 19 is taught that the character type of the input character pattern data is the teaching input to the R layer, the learning unit 19 indicates that the character pattern data input from the S layer is the taught character type. Each time new character pattern data is input, learning is performed so that the output shown can be obtained from the R layer. The result of this learning is from S layer to A
Weighting coefficient (W _ij ) of variable coupling from layer A and layer R
Reflected in. That is, the weighting factor _Wij becomes learning data for character recognition.

The weighting coefficient W thus obtained by the learning unit 19
_ij is assigned to the recognition unit 18 having the same structure as that of the learning unit 19 (in this case, the three-layer perceptron structure), and the recognition information from the A layer to the R
It is directly ported as a weighting factor for variable coupling to layers.
In the recognition unit 18, the image data from the character reading unit 17 is input to the S layer, and the recognition result is output from the R layer.

Since the character pattern generator 10 can generate a character pattern similar to a human handwritten character as described above, by incorporating this into the character recognition device as described above, noise is added to the conventional representative character pattern. In particular, the recognition rate for handwritten characters is dramatically improved as compared with the case where a character pattern generation device using such a method is used for character recognition.

Although one embodiment of the present invention has been described above, the present invention can be implemented in various modifications as follows.
For example, in the above embodiment, the dynamic characteristic simulating unit 13 is realized by an analog arithmetic circuit, but as is known in programming of an analog computer, the circuit configuration in the case of converting the above equations (1) and (2) into an analog arithmetic circuit. Is not unique, and various configurations are conceivable. The point is that the dynamic characteristic model of the hand given by the motion equations such as equations (1) and (2) can be incorporated.

Further, at present, it is easy to simulate the characteristics of an analog calculation circuit with a digital computer, and there are many existing programs classified as a continuous system simulation language. It is also possible to realize the dynamic characteristic simulating unit 11 by using.

Further, in the case where the recognition unit 18 and the learning unit 19 are realized by a neural network in the above embodiment, the recognition unit 18 and the learning unit 19 may be configured by separate hardware such as an integrated circuit. The unit 19 may be implemented by software using a general-purpose computer, and only the recognition unit 19 may be configured by hardware. The recognition unit 18 and the learning unit 19 may be configured as one piece of hardware, and the mode may be switched between the learning mode and the recognition mode and used.

[0042]

As described above, according to the present invention, by using the dynamic characteristic model of the human hand, a natural and smooth pattern close to that of an actual handwritten character can be obtained, and a wide variety of characters of the same character type can be obtained. Patterns can be generated. Further, as a side effect, the problem that the substantial size of the character pattern changes when the generated character pattern is normalized is also solved.

Further, in the present invention, many character patterns can be obtained by using the dynamic characteristic model of the hand only by preparing a relatively small number of reference character trajectories. It is not necessary to collect a large number of patterns of many handwritten characters, and the great effort required for the collection is unnecessary. Therefore, by using the character pattern generation device of the present invention, the recognition ability of the character recognition device can be easily and dramatically increased.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a character pattern generation device and a character recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram showing that the trajectory of a character is represented by two-dimensional coordinates.

FIG. 3 is a block diagram showing a specific configuration example of a reference character trajectory generation unit in FIG.

FIG. 4 is a diagram showing an example of a temporal change in coordinates of a reference trajectory.

5 is a block diagram showing a specific configuration example of a dynamic characteristic simulation unit of the hand in FIG.

FIG. 6 is a diagram showing a specific example of a reference character trajectory and a character pattern obtained by modifying the reference character trajectory in the same embodiment.

FIG. 7 is a diagram showing a conventional character pattern generation method.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 10 ... Character pattern generation device 11 ... Reference character trajectory generation unit 12 ... Character type designation unit 13 ... Hand dynamic characteristic simulation unit 14 ... Parameter setting unit 15 ... Display unit 16 ... Character pattern storage unit 17 ... Character reading unit 18 ... Recognition unit 19 ... Learning Department

Claims (5)

[Claims] 1. A reference character trajectory generation means for generating a reference character trajectory of a two-dimensional coordinate expression representing a representative pattern of a designated character type, and the reference character trajectory generation means for generating the reference character trajectory. A character type designating means for designating a character type to be used and a dynamic characteristic model simulating the dynamic characteristic of a human hand are provided internally, and a deformation by handwriting is simulated for the reference character trajectory generated by the reference character trajectory generating means. A character pattern generation device comprising: a dynamic characteristic simulating means for generating a deformed character trajectory of a given two-dimensional coordinate expression as a character pattern. 2. A reference character trajectory generating means for generating a reference character trajectory of a two-dimensional coordinate expression showing a representative pattern of a designated character type, and the reference character trajectory for the reference character trajectory generating means. A character type designating means for designating a character type to be used and a dynamic characteristic model having a plurality of parameters imitating the dynamic characteristic of a human hand are inherent, and handwritten with respect to the reference character trajectory generated by the reference character trajectory generating means. A dynamic characteristic simulating means for generating a deformed character trajectory of a two-dimensional coordinate expression simulated as a character pattern, and a parameter setting means for individually variably setting the plurality of parameters. A character pattern generation device characterized by: 3. A reference character trajectory generation means for generating a reference character trajectory of a two-dimensional coordinate expression showing a representative pattern of a designated character type, and the reference character trajectory generation means for generating the reference character trajectory. A character type designating means for designating a character type to be used and a dynamic characteristic model having a plurality of parameters imitating the dynamic characteristic of a human hand are inherent, and handwritten with respect to the reference character trajectory generated by the reference character trajectory generating means. Of the deformed character trajectory of a two-dimensional coordinate expression simulated by the above-mentioned method, a dynamic characteristic simulating means for the hand that generates a character pattern, a parameter setting means for individually variably setting the plurality of parameters, and the character type designation A plurality of dynamic characteristic models each having a parameter variably set by the parameter setting means for a plurality of character types designated by the means And a storage means for storing the character pattern generated by the dynamic characteristic simulation means as data. 4. The dynamic characteristic model in which the dynamic characteristic simulating means is present is represented by the following equation:
The character pattern generation device according to 2 or 3. mx ″ = ε (x _r −x) −rx ′ my ″ = ε (y _r −y) −ry ′, where x: horizontal axis coordinate of the character pattern y: vertical axis coordinate of the character pattern x _r : the above Function indicating the horizontal axis coordinates of the reference trajectory y _r : Function indicating the vertical axis coordinates of the reference trajectory m: Coefficient indicating the strength of inertia of the hand r: Friction coefficient between the writing instrument and the paper ε: Restoration to the reference trajectory Coefficient indicating force x ′: first-order differential coefficient of x y ′: first-order differential coefficient of y x ″: second-order differential coefficient of x y ″: second-order differential coefficient of y 5. A reference character trajectory generation means for generating a reference character trajectory of a two-dimensional coordinate expression showing a representative pattern of a designated character type, and the reference character trajectory for the reference character trajectory generation means. A character type designating means for designating a character type to be used and a dynamic characteristic model having a plurality of parameters imitating the dynamic characteristic of a human hand are inherent, and handwritten with respect to the reference character trajectory generated by the reference character trajectory generating means. Of the deformed character trajectory of a two-dimensional coordinate expression simulated by the above-mentioned method, a dynamic characteristic simulating means for the hand that generates a character pattern, a parameter setting means for individually variably setting the plurality of parameters, and the character type designation A plurality of dynamic characteristic models each having a parameter variably set by the parameter setting means for a plurality of character types designated by the means Accumulating means for accumulating the character pattern generated by the dynamic characteristic simulating means as data, learning means for performing learning for character recognition using the data of the character pattern accumulated in the accumulating means, and the learning means. And a recognition means for recognizing a character to be recognized based on the learning result of 1.