JP3162154B2 - Figure conversion method using neural network - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a graphic conversion method using a neural network, for example, a method of using a character font as a graphic.

[0002]

2. Description of the Related Art As a figure used in a human society, there is a figure in which a plurality of figures are decorated according to a certain rule. In ancient Greece, arabesque patterns are used to decorate various shapes. It is in the field of type that decorations according to such a certain rule typically appear. There are flower letters, italic, Gothic, Paika, elite, proportional, Mincho, textbooks, etc.

[0003] When texts are created and edited electronically as in the present case, prints are stored in an electronic format called a font for displaying or printing them. I have. These are the respective word processors,
It is created for each manufacturer such as a personal computer and a DTP, or even for one manufacturer, depending on the case. Therefore, there are numerous font sets in the world.

[0004]

Conventionally, font creation is performed manually, and in the case of Japanese,
In particular, since there are thousands of kanji, creating a set of fonts requires a great deal of labor. Mincho,
For textbooks and other fonts, veteran humans had to create characters one by one in order to represent the characters written by calligraphers with dots and functions. Therefore, it was enormous cost to make one set of characters, which was a major problem.

[0005] Such a problem is the same in the alphabet field. In the alphabet field, so-called typefaces are protected, and using typefaces created by others without permission raises copyright law problems. Therefore, there is a great need and demand for easily creating a typeface different from that of another person.

One of the objects of the present invention is to easily create such a character font. By the way, for example, italic is one type of italic in word processors,
Mincho style usually has only one type of Mincho style. Therefore, as long as a sentence created by a word processor is created by the same word processor, the same character is used by anyone.

[0007] Therefore, while being very easy to read and efficient, the individuality of the writer, which appears in handwritten letters and the like, is lost. Recently, in Japan, many people use word processors for personal letters and the like, and there are data indicating that most word processors sold to individuals are used for creating New Year's cards and New Year's cards. However, once a year's personal letter is written in bold print, it is very bland and I'm not happy to get it. However, using handwritten characters does not make sense using a word processor, which is a typesetting machine. This may be the case in Europe and the United States, where it is customary to send greeting cards such as Christmas cards.

[0008] Another object of the present invention is to allow a personal font, which is ideally close to a typeface, but which takes into account personal habits and tastes, to be used as a special character for that person's personal word processor. It is in.

However, as described above, fonts are created by hand, and are therefore very expensive, costing tens of millions to hundreds of millions of yen. Therefore, it was impossible for a very individual to have a font dedicated to his word processor.

Still another object of the present invention is to provide an inexpensive font for each individual.

[0011]

SUMMARY OF THE INVENTION The present invention automatically learns most of the unshown figures by learning several figures such as characters having a certain style as samples as samples. By creating in
They try to provide figures at low cost.

According to the present invention, as shown in FIG. 1, in a neural network, a group of deformed figures in which a desired decoration is applied to a group of original figures is prepared, and a deformation rule between the two is learned. A modified figure is obtained in accordance with the modified transformation rule.

[0013]

For example, when a character is treated as a graphic, a character font is obtained by deforming a prototype character according to one style. For example, the Mincho style is a variation of the stick on the right, which has a bulge called a helmet on the right end. In word processors of various companies, this deformation is slightly different, but it is also created by deforming all characters according to a certain rule.

Therefore, a set of fonts is completed by modifying the original characters in accordance with a certain transformation rule. Therefore, a desired decorative character group for the prototype character group is prepared, and the transformation rules between the two are learned by a neural network. When a character that has not been learned is input, a new decoration can be automatically performed according to the learned deformation rule. <Neural Network> The present invention can use various types of neural networks in addition to any hierarchical neural network. Examples of the neural network are shown in FIGS. The neural network is disclosed in Japanese Patent Application Laid-Open Nos. 2-64788 and 2-22638.
2, one disclosed in JP-A-2-22878 can be used. However, it is not limited to these.

[0015] The hierarchical neural network (NN) has at least two distinct layers, an input layer and an output layer. Originally, some processing is possible with only these two layers, but it is more effective to insert one or more intermediate layers called hidden layers between these layers.

In the neural network (NN), it is necessary to adjust the network so as to obtain a desired output with respect to an input pattern, that is, to learn, and a typical example of the method is a back propagation method.

In the present invention, the neural network (N
N) includes a learning unit (A), and a device that evaluates unlearned data using the weight of the weight updating unit 13 after the learning in the learning unit (A) is completed is preferably used.

Here, the learning unit (A) includes an input / output pattern holding unit 10 for holding an input to the neural network (NN) and a teacher pattern which is a desired output for the input / output, and an input / output pattern holding unit 10 for holding the input / output pattern. Is input to the neural network (NN) to calculate an output value, a learning parameter holding unit 12 for holding predetermined learning rules and learning parameters, and a learning rule and a learning parameter based on the learning rules and the learning parameters. The learning execution unit 11
A weight updating unit 13 that compares the output of the network with the teacher pattern of the input / output pattern holding unit 10 and updates the weight of the network, and a weight storage of the connection that stores the updated network weight at an arbitrary learning stage. A configuration including the unit 14 can be exemplified.

Further, a hierarchical neural network (N
As one form of N), an Elman-type recurrent network can be used. Although not shown, the Elman-type recurrent network includes an input layer, a hidden layer,
This is a neural network (NN) having an output layer and a context layer for copying and holding the output data of the hidden layer and inputting the held output data to the hidden layer again. Each of the input layer, the hidden layer, and the output layer has one or more nodes.

Then, the output value w of the hidden layer at time t
Is copied to the context layer, and the values of the context layer and the input layer become the output of the hidden layer at time t + 1. Therefore, this context layer holds the history of past input data. <Creation of Input Pattern and Teacher Pattern> In the present invention, a figure (input pattern) serving as a prototype and a figure (teacher pattern) obtained by decorating and deforming the pattern are image data expressing a certain electronic method and a diagram of the figure. And vector data are input to the neural network.

Therefore, in the present invention, as the input device, for example, an image scanner, a mouse, a trackball, a keyboard, a light pen, a handwriting input device, an image processing device having an imaging device, and the like can be used.

In the image scanner, the captured image information is used as it is as image data (dot data) or is converted into vector data for use. The mouse, trackball, and keyboard control the cursor on the screen and directly input vector data.

The light pen can draw characters directly on the screen. The handwriting input device has a pad or a screen for drawing, and takes in a handwritten figure drawn on the pad or the screen as dot image data. The captured dot image data may be used as it is, or may be used after being converted into vector data.

As a handwriting input method, there is a method of inputting by a pen touch method on a stylus screen. When using an image processing apparatus having an image pickup device, a prototype figure or a decorated / deformed figure is taken by an image pickup apparatus such as a TV camera, and the image is processed to render the drawing as dot image data, and further, as vector data. Get as.

When a graphic such as a character is input as vector data to a neural network and subjected to decoration processing according to a desired deformation rule, the following two methods are conceivable. (Pattern creation method 1) (shown in FIGS. 6 to 8) Vector data on strokes of a figure such as a character given as n pieces of continuous line segment information is converted into a set of continuous line segments for each stroke. A neural network for converting a stroke of a character is constructed by learning as a segment which is a minimum unit of processing, and learning it in correspondence with an input pattern and a teacher pattern of the neural network.

In the method 1, since one stroke is treated as one segment, when the stroke is approximated by a plurality of line segments by the polygonal line approximation method, the plurality of line segments are specified by their starting coordinates, length, and angle. Then, a line segment following the previous line segment is specified by the angle and length information. Therefore, the number of input / output nodes of the neural network is the number of lengths and angles of each of the n line segments plus the number corresponding to the starting point coordinates of the first line segment.

In the method 1, since the plurality of one strokes correspond to the fixed segment number, when the stroke is approximated by a plurality of straight lines, as shown in FIG. (The number of straight lines). In this case, it is possible to search for a corresponding stroke and linearly map according to the length of the stroke.

In this method 1, as shown in FIG. 8, all the data of one stroke is made to correspond to the input and output of the neural network, so that the desired stroke information can be obtained by one process.

(Pattern creating method 2) (shown in FIGS. 9 to 12) Vector data relating to a stroke of a figure such as a character given as n continuous line segment information is processed for each line segment constituting the stroke. Is considered as a sequence of pattern data representing n line segments, and the neural network learns each vector data of each line segment, so that the entire character is a set of n strokes. Construct a neural network that converts strokes.

In method 2, since each line segment constituting one stroke is treated as one segment,
When a stroke is approximated by a plurality of line segments by the polygonal line approximation method, the plurality of line segments are specified by their starting point coordinates, length, and angle.

In the method 2, in order to reduce the input and output vector data processed at one time to the vector data of each line segment constituting one stroke, the number of input / output nodes of the neural network is set to the starting coordinates, length, and angle, respectively. , The network can be simplified, and the processing time per operation can be reduced. However, it is necessary to operate the network n times by the number of line segments in order to obtain desired stroke information.

In the method 2, when two strokes are processed, the starting point of the second stroke is preferably obtained from an output indicating the first stroke. For example, as shown in FIG. 12, the first stroke is defined as X ₀ , Y ₀ ,
When the outputs for the inputs of L ₀ and A ₀ are X ₁ , Y ₁ , L ₁ and A ₁ , the inputs X ₂ , Y ₂ , L ₂ , and
Among A ₂ , X ₂ and Y ₂ indicating the starting point coordinates are preferably obtained by the following formula.

X ₂ = X ₁ + L ₁ × cos (A ₁ ) (1) Y ₂ = Y ₁ + L ₁ × sin (A ₁ ) (2) The two methods have been described, but FIG. 13 shows a comparison between the two methods. It is not possible to argue which is better in method 1 or method 2 because which segment is more advantageous in the stroke recognition output depends on how many segments are approximated and the performance of the execution machine.

As a technique for extracting feature points of an input pattern and a teacher pattern, Japanese Patent Laid-Open Nos. 62-216092 and
Techniques described in JP-A-2-164183, JP-A-62-208191, JP-A-62-92089, JP-A-62-92084 and the like can be used. <Method of Setting Teacher Pattern Using Neural Network> In the present invention, the rule of changing a figure is learned using a neural network. In this case, the learning method is divided into the following two types depending on how the teacher pattern is selected.

A case where a group of prototype figures is used as an input pattern, and a group of figures obtained by applying a desired decoration or the like to the group of figures is used as a teacher pattern (FIGS. 14 to 30).

The learning method will be described more specifically.
First, a prototype figure group (A1, A2,... An) is prepared. The prototype figure is represented in a certain electronic way. For example, image data or vector data expressing a diagram of a figure.

Next, a group of graphic samples (B1, B2,... Bn) obtained by decorating and deforming the prototype graphic is prepared. This is also represented by a certain electronic method such as image data or vector data.

Then, the subsets A1, A2,... Of the prototype figure group corresponding to the samples B1, B2,. . Give An. The teacher patterns include B1, B2,. . Give Bn. When learning is performed in this manner, a transformation rule from A to B is generated as network data.

Next, all the rest of the set of A, which have not been learned, are successively given to the trained neural net.
Then, a figure group B according to the learned deformation rule is obtained. A case where a group of prototype figures is used as an input pattern, and a difference between this group of figures and a group of figures obtained by applying a desired decoration or the like to a group of teachers is used as a teacher pattern (FIGS. 31 to 35).

This learning method will be described more specifically.
First, a prototype figure group (A1, A2,... An) is prepared. The prototype figure is represented in a certain electronic way. For example, image data or vector data expressing a diagram of a figure.

Next, a group of graphic samples (B1, B2,... Bn) obtained by decorating and deforming the prototype graphic is prepared. This is also represented by a certain electronic method such as image data or vector data.

Then, the subsets A1, A2,... Of the prototype figure group corresponding to the samples B1, B2,. . Give An. When information on the difference between the prototype graphic A and the deformed graphic B is given, it may be converted as a polynomial based on the term T = BA (3). That is, the simplest conversion equation is equation (3).

However, since T generally exceeds the range of the output layer unit of the neural network, equation (3) cannot be applied to generation of all learning patterns.

Therefore, it is usually preferable to use the equation (4) that can adapt any data. However, By giving equation (4) as a teacher pattern for the input pattern, if no change is made to the input, C: (upper limit−lower limit of sigmoid) / 2 is obtained as the output.

An example of the sigmoid function is shown in equation (5). w _i (i = 1, 2,..., n): connection between neurons θ: threshold This equation is expressed as shown in FIG. As shown in this figure, the sigmoid function often has a value range from 0 to 1.
The value of C: (upper limit of sigmoid-lower limit) / 2 is 0.5
Often becomes. In the present invention, since it is desired to learn the deviation from the prototype pattern, the teacher pattern is 0.5 + α except in extreme cases. Since the initial value of the weight of the neural network is usually given by a random number having a uniform distribution centered on 0, the initial output is often around 0.5. This indicates that the convergence point is close to the initial value, that is, that the convergence is fast.

In the case where the teacher pattern is not given by the difference as described above, that is, when the desired output value is given as it is as described above, in order to obtain the desired operation, it is necessary to prepare a considerably large number of learning patterns. No. The method of using the difference as the teacher pattern may be in terms of a prepared learning pattern or stability of operation at the time of execution. When converting the output from the neural network to the desired output, the above-described giving method simplifies the calculation process. <Input of Control Parameter> As shown in FIG. 16, the neural network preferably includes a node for inputting a control parameter in addition to an input node for inputting the input pattern.

The control parameters are usually given in the range of 0 to 1. Since the sigmoid threshold is usually from 0 to 1, it is given from 0 to 1 in the example. Therefore, the value itself has no meaning.

The value set for the control parameter is 1
, The output is closer to the teacher pattern, and as the value set for the control parameter is closer to 0, the output is closer to the input pattern. That is, an intermediate output between the input pattern and the teacher pattern can be obtained depending on the value of the control pattern.

A node for inputting a control parameter is a new node.
One or two or more may be provided in the neural network.
More specifically, a control parameter for inputting a control parameter
While at least one data input node is provided, the input
Prepare two or more pairs of tongue and teacher pattern, and input each
In the set of patterns and teacher patterns, the input patterns are the same,
Teacher patterns are different from each other, and the first
When learning a set, the control parameter is set to an arbitrary X, and the second
When learning a set, set the control parameter to any Y
When the input pattern of X and Y is given,
By giving an intermediate value between
Output an intermediate value. <Acceleration of learning convergence>
When it is considered that set B has an unknown conversion rule, the prototype
Subset A of set A of figures₀ For each element
Of deformed figures having a one-to-one correspondence₀ Choose
And a subset A of set A₀ Is an input pattern, and set B₀ 
Is learned by the neural network as a teacher pattern.
A ₀ Sets A to B of₀ Unknown transformation into a population B
The rule is estimated, and the transformed figure is
To obtain the set B of
Network speeds up learning.

First, a subset A ₀ of the set A of prototype figures is set as an input pattern, and a subset A ₀ identical to the input pattern is given to the neural network as a teacher pattern for learning. Such a teacher pattern is not inconsistent with the input pattern. Therefore, they always converge. In addition, the actual number of times of learning can be reduced. With the weight obtained by this learning as an initial value, the set A ₀ is used as an input pattern, and the set B ₀ is used as a teacher pattern and given to the neural network, and the unknown transformation rule from A ₀ to B ₀ is learned. Let it.

As described above, the learning stage is divided into two stages.
By saving the network structure after the first-stage learning and setting it as the initial value of the second-stage learning, the number of times of the second-stage learning can be substantially reduced.

As a result, a subset A of the set A of prototype figures
₀ may be used as an input pattern, and the set B ₀ may be used as a teacher pattern from the beginning to learn the unknown transformation rule from A ₀ to B ₀ . Many.

FIG. 36 shows the principle of speeding up learning. The first-stage learning has an energy space as shown by the dotted line in the figure, and the weight is at the position of point 1 when the convergence is achieved. The starting point of the second-stage learning is point 2, the convergence point of the second-stage learning stage is point 3, and learning is started from a generally given random weight (for example, a location like point 4). Convergence is faster than in the case.

The first-stage learning has the advantage that the energy space is simple and the convergence is easy because the number of learning patterns is small. As a result, as shown in FIGS. 37 and 38, learning is not performed in two stages as described above.
In the case of (1), the number of times of learning up to learning convergence is 10007, and it takes a long time. On the other hand, in this method in which the initial value is set at the first stage, the number of times until learning convergence is 752, which is short.

Note that this learning method is limited to the case where the control parameters are input simultaneously with the input pattern. This is because the following two conditions are required when applying this method. There is a guarantee that the learning of the first learning group will easily converge. The data structures of the first learning group and the second learning group are similar. (That is, the energy spaces of the learning first stage and the learning two stage are similar.) In this method, since these two conditions are clearly satisfied, it is possible to effectively use it. <Study of some common data> In standard data, a part of the data may be common due to the normalization of the data. May be the same.

For example, there is a case where a character is learned as graphic data. More specifically, suppose that when learning human personal characters and constructing a character font creation system, “silver” and “copper” are selected as learning patterns, for example, as shown in FIG. In this case, individual characters are registered and learned for each character, but the input pattern is exactly the same because the "gold" portion is the same.

On the other hand, an individual input is not input in exactly the same pattern even with the same “gold”. Therefore, a phenomenon occurs in which different teacher patterns correspond to the same input pattern.

In this case, there are two teachers for the input, and it is not clear whether one of them is a true teacher. Even if the number of times of learning is increased, learning is not achieved. Therefore, it is not possible to determine at which point the learning can be ended.

Therefore, in a neural network learning apparatus for learning a variety of graphic data groups from a standard graphic data group, when a part of the graphic pattern is common to a plurality of input graphic patterns, the common It is preferable to select only one graphic pattern from a group of input graphic patterns having a part as a learning pattern and to make the neural network learning device learn.

In this case, the input graphic patterns are compared with each other, the distance between one graphic pattern and the other graphic pattern is calculated, and the graphic patterns whose distances are equal to or less than a certain value are grouped into one group. Can be exemplified.

FIG. 39 shows a process when a plurality of teacher patterns exist for the same input pattern. In this figure, the vertical and horizontal axes have no meaning other than indicating that the space is a pattern. (10) to (10) show each input pattern. //
Since the pattern groups of and / (10) are close to each other, they are regarded as one group. In this case, one of // and one of / (10) are selected and used for learning.

[0062]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below. Before describing individual embodiments, first, a neural network used in the embodiments will be described.

The hierarchical neural network (NN) used in this embodiment has a learning unit (A) as shown in FIG. 2, and after the learning by this learning unit (A) is completed, the weight updating unit Evaluate the unlearned data using the 13 weights.

First, a neural network (NN)
Has an input layer, a hidden layer, and an output layer, and each layer has a plurality of nodes (units). A plurality of units constituting each layer are connected to each other with a certain weight. By giving a pair of an input pattern and a desired output pattern (teacher pattern) to this network, the weight of the network can be learned.

The learning proceeds as follows. First, an input pattern is given to the network to obtain an output. Extract information on the difference between the actual output and the correct output. Then, the network adjusts the internal structure of the network (the strength of connection = weight) so that the difference between the correct output value and the actual output value is reduced. By repeating this many times, the network automatically searches for weights that satisfy a certain input / output relationship. This learning algorithm is the back propagation method.

When the network trained in this way is used, the displayed correct output is returned for the learned input pattern, but an output pattern similar to the learned input / output pattern is returned for the input pattern that has not been further learned.
This is a major feature of the neural network (NN).

As described above, the neural network (N
N) has a learning unit (A), and after learning in the learning unit (A) is completed, processes unlearned data using the weight of the weight updating unit 13.

Here, the learning unit (A) comprises an input / output pattern holding unit 1 for holding an input to the neural network (NN) and a teacher pattern which is a desired output for the input.
0, a learning execution unit 11 for inputting the input pattern of the input / output pattern storage unit 10 to the neural network (NN) and calculating an output value, a learning parameter storage unit 12 for storing predetermined learning rules and learning parameters, A weight update unit 13 that compares the output of the network in the learning execution unit 11 with the teacher pattern of the input / output pattern holding unit 10 based on the learning rules and the learning parameters and updates the weight of the network. A connection weight storage unit 14 for storing network weights at any learning stage is provided.

The learning execution section 11 will be described in more detail.
As described above, the neural network (NN) has a multilayer network structure. Each layer is composed of many units, and a connection weight W is defined between each unit.

Each unit calculates the output value of the network as described below. When a certain unit receives inputs from a plurality of units, an input value net is obtained by adding the threshold value θ of each unit to the weighted sum (see FIG. 3).

Net = Σw _ij O _j + θ w _ij : weight of connection from unit U _j to unit U _i O _j : output of unit U _{j As} for the output value of the unit, the activation function is applied to the sum net of this input. Is calculated. A sigmoid function, which is a differentiable nonlinear function, is used as the activation function. Then, the output value O _i of the unit U _i becomes _{O i = 1 / {1 +} exp (-net i)}.

The network used in the back propagation method is a multi-layer network as described above, but here, the three layers of the input layer, the hidden layer, and the output layer as shown in FIG.
The case of a layer will be described.

Each unit in the hidden layer is associated with every unit in the input layer. Each unit of the output layer is connected to each unit of the input layer and the hidden layer. And there is no connection between units in each layer.

Each unit in the input layer is provided with input data to the network. Therefore, the output value h of each unit in the hidden layer is h _j = 1 / {1 + exp (−net _i )} d _K : output value of k-th input unit h _j : output value of j-th hidden unit w _jk : between k-th input unit and j-th hidden unit Weight _{j j} : threshold value of the j-th hidden unit.

The output value o of each unit in the output layer is h _j : output value of the j-th hidden unit o _i : output value of the i-th output unit w _ij : weight of the connection between the j-th hidden unit and the i-th output unit θ _i : of the i-th output unit It becomes the threshold.

Next, the weight updating unit 13 will be described in more detail.

The weight updating unit 13 changes the weight of the network so that the output of the network becomes a desired output. Actual output value when a certain pattern p is given ( _opi )
If, taking the mean square error E _p of the desired output value (t _pi).

[0078] E _p = (1/2) in order to (t _pi -o _pi) ² is learning to reduce this error, change all weights in the network.

The learning rules for the output layer are as follows: i) Change in weight between hidden layer unit and each unit in the output layer ii) Change in weight between unit in input layer and unit in output layer n: Number of learning α: Momentum δ _pi = (t _pi −o _pi ) ｛ _opi (1− _opi )｝ The learning rule for the hidden layer is a change in weight between the unit of the input layer and the unit of the hidden layer. It is.

Next, the connection weight storage unit 14 stores the weight of the network at an arbitrary learning stage. The output device loads the weight of the network stored in the weight storage unit 14 and, when an input pattern is given, calculates the output of the network. Using the network learned as described above,
Returns the correct output for the trained input pattern,
Further, an input pattern which has not been learned is returned as an output pattern based on the learned input / output pattern. <Embodiment 1> An example in which a character font is created by the above neural network will be described. The present embodiment is an example of a case in which a group of deformed figures (desired output patterns) in which a desired decoration or the like is applied to the group of primitive figures (input patterns) is directly learned as a teacher pattern. Here, the above-described method 2 is used as a pattern creation method, control parameter input is also used, and the above-described learning speed-up processing is also performed. (Procedure of Implementation) The procedure of this embodiment will be described with reference to the flowchart of FIG.

(A) First, a set of characters to be obtained (for example, educational kanji, JIS first level, etc.) is converted to a prototype font by a certain electronic method. In this case, a curve representing a character is represented by a plurality of continuous straight lines. , And each line segment (segment) of the approximated broken line is represented by a vector (coordinate data of the starting point, length data of the line segment, and angle data indicating the direction of the line segment) on the coordinate screen. .

The prototype font is an essence-like one with the decoration of characters removed. An existing font, for example, a Gothic font, may be used as a font to be used as a prototype. The character expressed in this manner is referred to as a prototype character group A.

(B) Next, a sample of a character having a style to be converted is prepared. That entire set is the character set we want to get right now. This is called a style character group B. Samples of style character groups are B1, B2,. . Bn
And this is also stored as vector data as in (a).

(C) For input to the neural network,
The subsets A1, A2,... Of the prototype character group A corresponding to the samples B1. . Give An (Step 1
01).

(D) The teacher patterns include B1, B2,. .
Bn is given (step 102). The learning section of the neural network learns the transformation rule from A to B based on the input data (step 103). Even if the learning does not converge, it is possible to end the learning (step 106). (E) After the learning is completed, all the remaining sets of the unlearned A are sequentially input to the trained neural net. (Step 107). Then, a deformation according to the above-described deformation rule is applied to all of the remaining sets of the unlearned set A (step 108), and all desired style character sets B are obtained (step 109).

In this way, character data is converted into an appropriate data expression and learned by the neural network, whereby a character reflecting the characteristics of an individual character can be generated. According to the above procedure, a simulation experiment was performed under the following learning conditions. Here, for the same prototype character, Mr. T and Mr. Y input desired handwritten characters as desired modified characters as teacher patterns.
After the learning was completed, the unlearned prototype characters were input, and a modified character corresponding to the prototype characters was obtained as an output of the neural network. The simulation results are shown in FIGS.
3 is shown. Teacher Pattern and Input Pattern Creation Method In this example, the above-described pattern creation method 2 is used, and FIG.
9, “Country”, “field”, “solid”, “a”,
Input patterns were created for the kanji of "brother" and "self."

In the creation, as shown in FIG. 10, first, a curve representing a character serving as a prototype is approximated to a polygonal line composed of a plurality of continuous straight lines. Then, of the approximated polygonal lines, the starting point coordinates (X _I0 , Y _I0 ) of the first line segment, the length of the first line segment (L _I0 ), and the first line segment are converted into a horizontal line including the starting point coordinates. The angle (A _I0 ) made with respect to this is obtained. Then, this first
The origin coordinates (X) of the second line segment following the line segment
_{I 1} , Y _I1 ), length (L _I1 ), and angle (A _I1 ). Then, also for the n-th line segment, its starting point coordinates (X _In ,
Y _In ), length (L _In ), and angle (A _In ) are determined in advance.

Next, as shown in FIGS. 19 and 20, two persons, Mr. T and Mr. Y, are assigned “country”, “field”, “solid”, “a”,
I asked them to write the kanji of "brother" and "self" by hand, and made them the desired modified characters.

For this deformed character, a curve consisting of one stroke constituting the character was approximated to a polygonal line consisting of a plurality of continuous straight lines, and vector data was obtained. In other words, the starting coordinates (X
_O.sub.0 , _{Y.sub.O.sub.0} ), the length of the first line segment ( _L.sub.O0 ), and the angle ( _A.sub.O0 ) formed by the first line segment with respect to the horizontal line including the origin coordinates. Next, for the second line segment following the first line segment, the origin coordinates (X _O1 , Y _O1 ), length (L _O1 ), angle (A _O1 ), and n-th line segment , Its starting point coordinates (X _On , Y _On ), length (L _On ), and angle (A _On ) are obtained in advance. Neural Network The neural network used has four input nodes, one control parameter input node, and four output nodes as shown in FIG.

・ Structure units = (5 5 4) ・ Learning coefficient epsilon = (5 1) alpha = (0.8 0.3) ・ Learning tolerance allowance = 0.01 ・ Maximum number of learning iteration = 10000 ・ Random sheet random_seed = 5 ・ Initial value weight = random (0.5 -0.01 0.
01) threshold = random (0.5 -0.01 0.01) Learning During learning, the transformed figures themselves were used as teacher patterns.

In the neural network having this configuration, (X
_I0 , Y _I0 , L _I0 , A _I0 ) as input patterns, and (X _O0 , Y
_O0 , L _O0 , A _O0 ) as a teacher pattern,
(X _I1 , Y _I1 , L _I1 , A _I1 ) as input patterns,
(X _O1 , Y _O1 , L _O1 , A _O1 ) are given as teacher patterns,
This is input to the input pattern (X _In , Y _In ,
L _In , A _In ) and the output pattern (teacher pattern) (X _On ,
Y _On , L _On , and A _On ) are repeated to learn about one stroke information.

As described above, in the method 2, the input and output vector data are narrowed down to four parameters, and the input and output nodes of the neural network are each set to four, thereby simplifying the network and reducing the processing time per one time. Although the size is reduced, it is necessary to operate the network n times (n: the number of line segments obtained by approximating a character by polygonal line approximation) several times in order to obtain desired stroke information.

In learning, “country”, “field”,
The following two conversions were learned for each of the characters "Ku", "A", "Brother", and "Shi". Prototype character → Prototype character (Control parameter = 0) Prototype character → Deformation character (Control parameter = 1) Here, the reason for learning Prototype character → Prototype character (Control parameter = 0) is as described above. That's why.

In the above neural network, the learning did not converge, and was therefore terminated at the maximum number of times of learning of 10,000. At this time, up to 2 of Mr. T's network
The squared error was 0.005320 and the average error was 0.000419.

The maximum square error of Mr. Y's network was 0.007625, and the average error was 0.000623. When learning one stroke information is completed,
When learning the next stroke information, from the data of the previous first line segment (X _O0 , Y _O0 , L _O0 , A _O0 ) as the starting point coordinates serving as the first reference position of the stroke, the above equation (1) is obtained. According to (2), the starting coordinates of the second line segment are obtained. As a result, since the starting points of the character strokes match, the resulting character does not become unnatural (FIG. 30B).
reference). If the origin coordinates of the previous first line segment (X _O0 , Y
_{If O0} ) is not inherited, the result may be that the line segments that should originally intersect do not intersect, as shown in FIG. Font generation After learning the learning pattern, unlearned "times", "medium",
The three characters of "neck" were input, and the characters after deformation based on the learned deformation rules were obtained. The results are shown in FIG. 21 and FIG.
2, shown in FIG. However, for each character,
The control parameters are changed from 0 to 1 every 0.2 in six ways. The results are shown in the figure.

The value of the control parameter is c_
para = ?????? To obtain an output closer to the personal characters trained by Mr. T and Y, set the control parameter to a value close to 1; to obtain an output closer to the prototype character, input a value close to 0. Fig. 21 and Fig. 2
2 and FIG. <Example 2> A character font was automatically generated under the same conditions as in Example 1 except for the following conditions. Neural Network The alphabet was learned using the following network.

・ Structure units = (5 2 4) ・ Learning coefficient epsilon = (5.00000 1.00000) alpha = (0.80000 0.30000) ・ Learning tolerance allowance = 0.1 ・ Maximum number of learning iteration = 5000 ・ Random sheet random_seed = 5 ・ Weight weight = random (0.2 -0.1 0.
1) ・ Threshold value threshold = random (0.2 -0.1 0.
1) random (abc) indicates that the weight is set by a random number from b to c by the ratio a (otherwise it is set to 0). Learning Patterns As learning patterns, as shown in FIGS. 24 and 25, letters of alphabets “A”, “T”, “H”, “M”
The following two transforms were trained.

Prototype character → personal character (deformed character) (control parameter = 1) Prototype character → prototype character (control parameter = 0) Creation of font After learning the learning pattern, unlearned “K”, “E”,
Input the four characters “I” and “L”, and
As shown in FIG. 7, FIG. 28, and FIG. 29, characters after deformation based on the learned deformation rules were obtained. However, for each character, the control parameter is changed from 0 to 1 in six increments of 0.2. By changing the control parameters, a deformed character located between the input pattern and the teacher pattern can be obtained as described above. Example 3 A character font was automatically generated under the same conditions as in Example 1 except for the following conditions. Learning Pattern Creation Method The above-described learning pattern creation method 1 was used. First, as shown in FIG. 7, a curve representing a character serving as a prototype is approximated to a polygonal line composed of a plurality of continuous straight lines. Then, of the approximated polygonal lines, the starting point coordinates (X _I0 , Y _I0 ) of the first line segment and the first
Of the line segment (L _I0 ), and the angle (A _I0 ) formed by the first line segment with respect to the horizontal line including the origin coordinates. Next, also for the second line segment following the first line segment, the length (L
_I1 ), the angle (A _I1 ) formed by the second line segment with respect to the horizontal line including the starting point is obtained. This is continued for the n-th line segment, and the length (L _In ) and angle (A _In ) are determined.

Similarly, for the desired deformed character
Also composes a curve consisting of one stroke that constitutes a character.
Approximate a polyline consisting of a series of straight lines, and
Of the approximated polygonal lines, the starting point coordinates (X_O0, Y
_O0), The length of the first line segment (L_O0), The first line segment is
Angle (A) to the horizontal line containing the point coordinates_O0Get)
Next, a second segment following the first segment is also
Length (L_O1), Angle (A _O1) And calculate the nth line segment
The length (L_On), Angle (A_On).

Then, (X _I0 , Y _I0 , L _I0 , A _I0 ,
L _I1 , A _I1 ,... L _In , A _In ) are input patterns,
(X _O0 , Y _O0 , L _O0 , A _O0 , L _O1 , A _O1 ,...
L _On , A _On ) as output patterns (teacher patterns),
Learn character stroke information.

In the method 1, since all the input and output vector data are made to correspond to the input and output of the neural network, the target stroke information can be obtained by one process.

Therefore, as shown in FIG. 8, the number of input nodes and output nodes of the neural network is 2
n + 2. "Country",
The following two transformations were learned for "field", "solid", "a", "brother", and "self".

Prototype character → personal character (control parameter = 1) Prototype character → prototype character (control parameter = 0) Then, after learning of the learning pattern, an unlearned “times” is input and based on the learned deformation rule. I got the character after it was deformed. The characters obtained are shown in FIG. <Embodiment 4> This embodiment is an example of learning a difference between a group of prototype figures (input patterns) and a group of deformed figures (desired output patterns) on which a desired decoration or the like has been applied as a teacher pattern. .

In this embodiment, as shown in FIGS. 31 and 32, as the teacher pattern T, T = {(BA) / 2} + C (4) B: deformed figure (teacher pattern) A: prototype figure (input pattern) C: (upper limit-lower limit of sigmoid function) / 2

FIG. 33 shows the configuration of the neural network in this case. The neural network has a post-processing unit 50 and a bypass 51 that bypasses the neural network from the input and sends input vector data to the post-processing unit.

The processing in the post-processing unit 50 is shown in FIG.
This will be described with reference to (c) and (d). First, the conversion equation T = (O
-I) / R + C. T: Output of the neural network O: Original desired output I: Input of the neural network R: Conversion rate C: Constant (const.) In order to obtain O, T = O / RI-I / R + CO / R = T + I / R-CO = I + (TC) * R.

Accordingly, each input / output relationship of each node in FIG. 33 (a) is as shown in FIG. 33 (b). In FIG. 33 (b), when R = 1, the shape becomes the simplest one shown in FIG. 33 (c), and when R = 2, which is often used, the shape becomes as shown in FIG. 33 (d).

FIG. 3 shows an example of the teacher pattern used in this embodiment.
It is shown in FIG. After the learning, when an input was made for the unlearned “neck” character, an output as shown in FIG. 35 was obtained. This is compared with FIG. 18 in which the teacher pattern is used as it is. <Embodiment 5> This embodiment is an embodiment in the case where a character recognition system is constructed by utilizing the present invention.

Although an optical character reader (OCR) has recently been used as a character input device, it does not always have a good recognition rate. Conventional character recognition systems incorporate existing character fonts and can recognize read characters if they are the same or similar to the character font.
In many cases, personal characters cannot be recognized.

Therefore, if a group of characters that cannot be recognized, such as personal characters, is incorporated into the system as a recognizable character font by the above-described method of the present invention, personal characters and the like can be recognized thereafter.

As shown in FIG. 41, the image scanner 61 and the image scanner 6 have the same configuration.
A character recognition system 63 for recognizing the character input in 1;
If it is determined that character recognition by this character recognition system is not possible, a character to be recognized input from the image scanner is used as a teacher pattern, a recognizable existing character font is used as an input pattern, and a recognizable character font group And a neural network NW for generating

Here, the character recognition system includes a character font storage unit 65 for storing a reference character font group used as a reference for character recognition, and character data input by an image scanner. A recognition unit 67 for recognizing characters in comparison with fonts.

The neural network uses the above-described configuration. That is, when it is considered that there is an unknown conversion rule from the set A of the prototype figures to the set B of the deformed figures, the subset A ₀ of the set A of the prototype figures is
We picked a set B ₀ corresponding relationship deformation shape of one-to-one for each element, an input pattern subset A ₀ of the set A, the set B ₀ as a teacher pattern, is trained in the neural network, A ₀ The unknown conversion rule from the mother set A to the mother set B of B ₀ is estimated, and a set B of deformed figures is obtained using the estimated conversion rule.

When a character group that cannot be recognized by the character recognition system is input,
And ₀ a is the teacher pattern, in response to the teacher pattern each, given to the neural network a recognizable character font by the recognition system as an input pattern consisting of the subset A _0, unknown from the A ₀ to B ₀ Is estimated, and a character font group B corresponding to the set A of the recognizable prototype characters is finally obtained.

The procedure is as specifically described in the above embodiments. The finally obtained character font group is stored in the character font storage unit 65 of the character recognition system as a new recognizable character font group.

The character recognition system itself can be constituted by a neural network. With the character recognition system having such a configuration, the recognition rate can be significantly improved.

[0117]

According to the present invention, with respect to the subset A ₀ of the set A prototype figure, only provide a set B ₀ corresponding relationship figure of one to one for each element to the neural network, The entire set B of deformed figures having a one-to-one correspondence with each element of the set A of prototype figures can be easily obtained.

Therefore, character fonts and the like can be obtained easily and inexpensively, and individual character fonts and typefaces can be used on an individual level.

[Brief description of the drawings]

FIG. 1 is a principle diagram of the present invention.

FIG. 2 is a block diagram showing an example of a neural network.

FIG. 3 is a diagram showing input / output characteristics of a network;

FIG. 4 is a diagram showing a network structure;

FIG. 5 is a graph showing a sigmoid function.

FIG. 6 is a diagram showing a method 1 for creating an input pattern and a teacher pattern.

FIG. 7 is a diagram showing a method 1 for creating an input pattern and a teacher pattern.

FIG. 8 is a configuration diagram of a neural network corresponding to the pattern creation method 1.

FIG. 9 is a diagram showing a method 2 for creating an input pattern and a teacher pattern.

FIG. 10 is a diagram showing a method 2 for creating an input pattern and a teacher pattern.

FIG. 11 is a configuration diagram of a neural network corresponding to the pattern creation method 2.

FIG. 12 is a diagram showing a method for obtaining the starting point of the next stroke in pattern creation method 2;

FIG. 13 is a comparison diagram of pattern creation methods 1 and 2;

FIG. 14 is a configuration diagram when a desired output pattern is used as a teacher pattern as it is.

FIG. 15 is a flowchart of character font creation according to the present invention.

FIG. 16 is a configuration diagram of a neural network according to the first embodiment.

FIG. 17 is a learning pattern when a desired output pattern is used as it is as a teacher pattern.

FIG. 18 is an output result of the neck in the case of the first embodiment.

FIG. 19 shows a learning pattern 1 according to the first embodiment,
9 (a) is a prototype character, FIG. 19 (b) is a personal character (Mr. T) as a teacher pattern, and FIG. 19 (c) is a personal character (Mr. Y) as a teacher pattern.

20 shows a learning pattern 2 according to the first embodiment, and FIG.
9 (a) is a prototype character, FIG. 19 (b) is a personal character (Mr. T) as a teacher pattern, and FIG. 19 (c) is a personal character (Mr. Y) as a teacher pattern.

FIG. 21 is an execution result 1 of an unlearned pattern according to the first embodiment.

FIG. 22 shows an execution result 2 of an unlearned pattern according to the first embodiment.

FIG. 23 shows an execution result 3 of an unlearned pattern according to the first embodiment.

FIG. 24 shows an input pattern and a teacher pattern 1 according to the second embodiment.

FIG. 25 shows an input pattern and a teacher pattern 2 in the second embodiment.

FIG. 26 is an execution result 1 of an unlearned pattern according to the second embodiment.

FIG. 27 is an execution result 2 of an unlearned pattern according to the second embodiment.

FIG. 28 shows an execution result 3 of an unlearned pattern according to the second embodiment.

FIG. 29 is an execution result 4 of an unlearned pattern according to the second embodiment.

FIG. 30 is a comparison diagram of the results obtained by the input pattern and teacher pattern creation methods 1 and 2;

FIG. 31 is a principle diagram of a neural network using a difference between an input pattern and a desired output pattern as a teacher pattern.

FIG. 32 is a diagram showing difference data between an input pattern and a desired output pattern for a human version;

FIG. 33 is a specific configuration diagram of a neural network in which a difference is used as a teacher pattern in the fourth embodiment.

FIG. 34 shows an example of a teacher pattern composed of a difference between an input pattern and a desired output pattern.

FIG. 35 is an execution example when a difference between an input pattern and a desired output pattern is used as a teacher pattern;

FIG. 36 is a diagram showing a weight space of a network;

FIG. 37: When the same pattern is not used as a teacher pattern (a)
Comparison of the learning process with (b)

FIG. 38: When the same pattern is not used as a teacher pattern (a)
Comparison of learning time between (b) and (b)

FIG. 39 is a conceptual diagram showing processing when a plurality of teacher patterns exist for the same input pattern.

FIG. 40 is a specific example in the case where a plurality of teacher patterns exist for the same input pattern.

FIG. 41 is a character recognition system using the graphic creation system of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 ... Input / output pattern holding part 11 ... Learning execution part 12 ... Learning parameter holding part 13 ... Weight updating part 14 ... Weight storage part 50 ... Post-processing part 51 ... Bypass Road 61: Image scanner 63: Character recognition system 65: Character font storage 67: Recognition unit NW: Neural network

──────────────────────────────────────────────────続 き Continuation of the front page (51) Int.Cl. ⁷ Identification code FI G09G 5/24 G09G 5/24 (56) References JP-A-3-189876 (JP, A) JP-A-3-273761 (JP (A) JP-A-4-109299 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G06G 7/60 G06F 15/18 G06K 9/66 G06K 9/68 G06T 1/00 G09G 5/24

Claims (10)

(57) [Claims] 1. A character from a set A of a prototype of a character font
When it is considered that there is an unknown conversion rule in the set B of the deformed figures of the font, the set B _{0 of the} deformed figures having a one-to-one correspondence with each element with respect to the subset A ₀ of the set A of the prototype figures. Is selected, and the neural network is trained using the subset A ₀ of the set A as an input pattern and the set B ₀ as a teacher pattern, and from the mother set A of the subset A ₀
To estimate an unknown conversion rule from mother to set B of the set B _0, using the conversion rule the estimated geometric transformation method using a neural network, characterized in that to obtain a set B of deformation shape. 2. A graphic conversion method using a neural network according to claim 1 , wherein said desired deformed graphic group is learned as a teacher pattern as it is. 3. The graphic conversion method according to claim 1, wherein difference information between the desired deformed graphic and a prototype graphic as an input pattern is learned as a teacher pattern. 4. The graphic conversion system according to claim 1, wherein the input pattern is input to the neural network and a control parameter for changing the output pattern is input. 5. A set A of a prototype figure and a set B of a deformed figure
Is considered to have an unknown conversion rule, a set B _{0 of} deformed figures having a one-to-one correspondence with each element is selected for the subset A ₀ of the set A of the prototype figure. When the neural network is trained using the subset A ₀ as an input pattern and the set B ₀ as a teacher pattern, an input pattern consisting of a subset A ₀ of a set A of prototype figures, a control parameter, and the input pattern A training pattern consisting of the same subset A ₀ is given to a neural network for learning. The weights obtained by this learning are used as initial values, the subset A ₀ of the set A is used as an input pattern, and the set B ₀ is used as a teacher pattern.
The figure conversion method using a neural network according to claim 4, wherein the neural network is made to learn. 6. Vector data relating to a stroke of a graphic such as a character given as n continuous line segment information is regarded as a segment which is a minimum unit of processing for each stroke which is a set of continuous line segments. 2. The graphic conversion method according to claim 1, wherein a stroke of a character is converted by learning in correspondence with an input pattern and a teacher pattern of the neural network. 7. Vector data relating to a stroke of a graphic such as a character given as n continuous line segment information is regarded as a segment which is a minimum unit of processing for each line segment constituting the stroke, and n line segments are processed. The entire character stroke, which is a set of n strokes, is converted by assuming a series of pattern data representing the set of n and learning the neural network for each vector data of each line segment. Figure conversion method using neural network. 8. When a part of a graphic pattern is common to a plurality of input graphic patterns, only one graphic pattern is selected as a pattern from an input graphic pattern group having the common part, and learning is performed on the neural network. 2. A graphic conversion method using a neural network according to claim 1, wherein: 9. The input graphic patterns are superimposed and compared with each other, a distance between one graphic pattern and another graphic pattern is calculated, and a graphic pattern whose distance is equal to or less than a certain value is set as one group, and one of the groups in the group is determined. 9. The graphic conversion method using a neural network according to claim 8, wherein one is selected. 10. An image scanner, a character recognition system for recognizing characters input by the image scanner,
If it is determined that character recognition by this character recognition system is not possible, a character to be recognized input from the image scanner is used as a teacher pattern, a recognizable existing character font is used as an input pattern, and a group of recognizable character fonts is used. And a neural network that creates an A. The neural network, when it is considered that there is an unknown conversion rule from the set A of the prototype figures to the set B of the deformed figures, for a subset A ₀ of the set A of the prototype figures, A set B of deformed figures having a one-to-one correspondence with each element
₀ picked, an input pattern subset A ₀ of the set A, the set B ₀ as a teacher pattern, is trained in the neural network, an unknown conversion rule from the mother set A of A ₀ to mother set B of B ₀ was estimated, using the conversion rule estimated, which obtain a set B of deformation figure, when the character recognition system according to impossible to recognize a character group is input, it is the set B ₀ this teacher and pattern, in response to the teacher pattern each, given to the neural network a recognizable character font by the recognition system as an input pattern consisting of the subset a _0, the a
A character font group B corresponding to a set A of prototype characters that can be finally recognized by estimating an unknown transformation rule from ₀ to B ₀
And a new character font group that can be recognized.