JPH05173537A - Character font generating method by neural network - Google Patents

Abstract

PURPOSE:To generate another character font having good quality from one character font. CONSTITUTION:This character font generating method divides one character font into thin KANA (Japanese syllabary) parts and substitutes the parts with parts of another character font to generate another character font; and the neural network 7 learns the position between the positions of the parts and the position of the parts of another character font by using a rough shape of one character font, the congestion extent of the components of the character font, and the kinds of the parts as parameters, and the positions of the parts are inputted to the neural network 7 by using the rough shape of one character font, the congestion extent of the parts of the character font, and the kinds of the parts as parameters to obtain the positions of the parts of the new character font.

Description

Detailed Description of the Invention

[Table of Contents] Industrial Application Field of the Prior Art (FIG. 7) Problem to be Solved by the Invention Means for Solving the Problem (FIG. 1) Action Example (a) Description of One Example (FIG. 2) To FIG. 6) (b) Description of another embodiment Effect of the invention

[0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network character font generation method for generating a character font of another typeface from a character font of one typeface by using a neural network.

Various Japanese character fonts are used in word processors, electronic typesetting machines, display devices, printer devices, etc. Such Japanese character fonts have one typeface or 7,000 characters. It takes a lot of work.

Generally, in order to create a font to be used, in the case of a character font that already exists as a type in the world, a printed matter in a type is read by a scanner or the like, a human makes minor corrections, and a character font that does not exist. In case of, write the character font and read it with a scanner,
Make corrections manually or create from the beginning on a character font editing system.

As described above, since the Japanese character font has a large number of characters and requires a great deal of labor and creation cost, the typeface type of the Japanese character font provided is in English in a general word processor. Compared to about 20 types of characters, there are very few 2-3 types.

[0006] As described above, when the typefaces of Japanese character fonts are small, it is difficult for the author's personality to appear in the output,
Also, the typefaces that can be selected according to the type of output material are limited. Therefore, it is desired to easily generate Japanese character fonts for many typefaces.

[0007]

2. Description of the Related Art FIG. 7 is an explanatory diagram of a conventional technique. Each character is divided into a crease and a make-up, further divided into strokes, and further divided into parts such as a straight line, a left end decoration, and a right end decoration.

That is, it can be said that a character font is a set of such parts. Therefore, an attempt has conventionally been made to represent each character as a set of parts and generate a character font.

For example, as shown in FIG. 7 (A), a stroke word "word" as a basic character has a "word" as its edge as shown in FIG. 7 (B), and its horizontal stroke is As shown in 7 (C), the parts are divided into the left end, the straight line, and the right end.

The left end of this part is the leftmost part of the Mincho body
When the right end is replaced with the right end part of Mincho type, a character pattern of Mincho type can be generated as shown in FIG. 7 (D). In this way, by having one reference character pattern, the parts can be replaced with the parts of each typeface to generate character patterns of various typefaces.

[0011]

However, the prior art has the following problems. Although the shape of the parts that make up the character font changes,
Since the relative position between the parts does not change, the shape of the whole character is different from the original Mincho body shown in FIG. 7 (E), which is a simple replacement of the parts shown in FIG. 7 (D). I can't get other good typeface characters.

There is also proposed a method in which the positional relationship between parameters such as character font characteristics and parts is made into a rule, and characters are generated using an artificial intelligence method. However, each parameter is complicatedly related to each other. Therefore, it is difficult to make rules, and the quality of the generated character font is not good.

Accordingly, it is an object of the present invention to provide a character font generation method by a neural network capable of generating a high quality character font of another typeface from a character font of one typeface.

[0014]

FIG. 1 shows the principle of the present invention. According to claim 1 of the present invention, a character font generation is performed in which a character font of one typeface is divided into fine parts and the parts are replaced with parts of a character font of another typeface to generate a character font of another typeface. In the method, the rough shape of some character fonts of the one typeface, the degree of congestion of the parts constituting the character font, and the kind of the part are used as parameters, and the position and the part of the part are manually created. , The neural network 7 is made to learn the relationship between the font of the same character as the character selected in one typeface and the position of the parts of the character font of the other typeface, and the rough outline of the character font that the one typeface has not learned. By inputting the position of the part to the neural network 7 using the shape, the congestion degree of the parts constituting the character font, and the kind of the part as parameters, And wherein the obtaining the position of the character font parts that do not learn.

According to claim 2 of the present invention, in claim 1,
The rough shape of the character font of the one typeface, the degree of crowding of the parts forming the character font, and the kind of the part are used as parameters, and the position of the part is used as the input side teacher data of the neural network 7. , The neural network 7 is learned by using the difference between the positions of the character font parts of the one typeface and the character font parts of the other typeface as the output side teaching data of the neural network 7.

A third aspect of the present invention is based on the first aspect.
The rough shape of the character font of the one typeface, the degree of crowding of the parts forming the character font, and the kind of the part are used as parameters, and the position of the part is used as the input side teacher data of the neural network 7. , Learning of the neural network 7 is performed by using the positions of the character font parts of the other typeface as output side teaching data of the neural network 7.

According to claim 4 of the present invention, in claim 1 or claim 2 or 3, the character font of the one typeface is a rough font having a reduced resolution of the character font of the one typeface. Is characterized by.

According to a fifth aspect of the present invention, in the first, second, third, or fourth aspect, a mask having a high resolution at the position of the part and a low peripheral resolution is used as the crowded condition of the parts constituting the character font. It is characterized by using the data through.

[0019]

It is considered that the position of a certain part of a character is determined by the approximate shape of the character and the degree of congestion around the part. Further, if the typefaces are the same, it is considered that the arrangement features of the parts are common. For example, the Mincho type has a decoration called a scale, so the horizontal bar is located slightly below and the vertical bar is located slightly to the left, compared to bone characters.

On the other hand, the neural network 7 can easily form a rule by learning the relationship between the input and the output. Therefore, in claim 1 of the present invention, the rough shape of the character font of the one typeface, the degree of crowding of the parts constituting the character font, and the kind of the part are used as parameters, and the position of the part and the Let the neural network 7 learn the relationship with the positions of the parts of the character font of other typefaces,
With the rough shape of the character font of the one typeface, the degree of crowding of the parts forming the character font, and the kind of the part as parameters, the position of the part is input to the neural network 7, By obtaining the positions of parts of character fonts of other typefaces, the arrangement positions of parts according to other typefaces are generated, and character fonts of other typefaces with good balance and quality are obtained. The neural network will learn the characteristics of the difference in the arrangement of the parts of bone characters and created characters.

According to a second aspect of the present invention, the position of the part is determined by using the rough shape of the character font of the one typeface, the degree of congestion of parts constituting the character font, and the kind of the part as parameters. , The neural network 7
The input side teacher data is used as the input side teacher data, and the difference between the positions of the character font parts of the one typeface and the character font parts of the other typeface is used as the output side teacher data of the neural network 7, and the neural network 7 is learned. So
As the output of the neural network 7, the relative positional difference of the parts between the fonts can be obtained, and the arrangement position with less error can be obtained.

According to a third aspect of the present invention, the position of the part is determined by using the rough shape of the character font of the one typeface, the degree of congestion of parts constituting the character font, and the kind of the part as parameters. , The neural network 7
As the input side teacher data, and the position of the part of the character font of the other typeface as the output side teacher data of the neural network 7, the neural network 7 is learned, so that the output of the neural network 7 is
It is possible to directly obtain the arrangement position of other typeface parts.

In claim 4 of the present invention, since the character font of the one typeface has a reduced resolution of the character font of the one typeface, the rough shape of the character font can be easily obtained. Can be obtained.

According to the fifth aspect of the present invention, since the data through a mask having a high position resolution and a low surrounding resolution is used as the congestion degree of the components forming the character font, the component It is possible to obtain data on the degree of congestion.

[0025]

【Example】

(a) Description of an embodiment FIG. 2 is a block diagram of an embodiment of the present invention, FIG. 3 is an illustration of a character font of an embodiment of the present invention, and FIG. 4 is an illustration of font data of an embodiment of the present invention. FIG. 5 is an explanatory diagram of pattern cutting according to an embodiment of the present invention, and FIG. 6 is an explanatory diagram of a generation operation of an embodiment of the present invention.

In FIG. 2, reference numeral 1 is a reference font data storage unit which stores the position of each component of the reference font as shown in FIGS. 3 (A) and 4 (A), for example, 7000. What stores character font data, 2 is a dot pattern expansion unit, which is shown in FIG. 3 from the font data of FIG. 4A of the reference font data storage unit.
What is developed in the dot pattern font of (A), 3
Is a dot pattern primary storage unit for temporarily storing the dot pattern font developed by the dot pattern development unit 2.

Reference numeral 4 denotes a resolution conversion unit, which is font data (for example, 256 × 25) in the dot pattern primary storage unit 3.
6) is resolution-converted into a 16 × 16 low-resolution font to form an approximate shape of the font. Reference numeral 5 is a pattern cutout portion, and as shown in FIG. the central a ₉ of squares a ₁ ~a ₉ of × 3, 3 × 3
Cells b _{1 to} b ₉ and the center b ₉ is 3 ×
It is divided into cells of 3 and is composed of a mask having a higher resolution in the center and a lower resolution in the surroundings, and creates congestion degree data around the target component from the font data in the dot pattern primary storage unit 3. Is.

Reference numeral 7 is a neural network, which has a parts identification unit 6, and is used to roughly determine the arrangement coordinates (X, Y) of the parts in the reference font data storage unit 1, the kind of the parts, and the font from the resolution conversion unit 4. Shape data and the congestion degree data around the part from the pattern cutout portion 5 are input, and the learning data is used as teacher data to learn the relationship between the position of the part of the reference font and the position of the part of the created font. It is a thing.

Reference numeral 8 is a learning data storage unit, which is shown in FIG.
The difference between the coordinates of the component data of the generated font of FIG. 4B obtained by obtaining a part of the font to be created as shown in FIG.
What is stored as output side teacher data as in (C),
Reference numeral 9 denotes a character data generation unit, which generates the arrangement coordinates of the parts of the generated font from the offset values X and Y from the neural network 7 and the arrangement coordinates of the parts of the reference font, and the kind of the parts (right end of stroke, etc.). The character data storage unit 10 outputs the type of parts of the generated font from the character data generation unit 9 and the arrangement coordinates thereof.

Next, the operation will be described. First, learning character data is created. The reference font is divided into parts as shown in FIG. 3 (A), and as shown in FIG. 4 (A),
It is assumed that each component is expressed by the origin coordinate data (X, Y) of each component and the communication relationship between the components, and for example, 7,000 characters are stored in the reference font data storage unit 1.

For example, referring to the character "word" in FIG. 3A, the stroke "-" of the word "word" is the right end, a straight line,
As shown in FIG. 4 (A), the stroke origin is divided into the leftmost parts, and as the stroke "-" parts, the origin origin coordinates (absolute coordinates) X and Y of the right end, the straight line, and the left end, and the communication relationship between the parts are shown. It is stored as data.

On the other hand, among the characters of the font to be created, some of the characters (for example, 100 characters) that include all the parts for expressing the characters and clearly represent the characteristics of the font to be created. , Create as before.

At this time, the size of the coordinate system used is the same as that of the reference font, and the expression format of the character data is also the same as that of the reference font data. For example, as shown in FIG. 3 (B), when the created font is Mincho type, "Word" that expresses the characteristics of Mincho type is selected, this pattern is created, and "Word" of the standard font is created. Similarly,
The stroke "-" of the word "word" is divided into right end, straight line, and left end parts, and as shown in FIG. Absolute coordinates) X, Y and the communication relationship between the parts create data.

Then, as shown in FIG. 4 (C), the arrangement coordinates (X1m to X3m, Y1m to
Y3m) and the placement coordinates of each part of the created font (X1n
To X3n, Y1n to Y3n) are created for each type of each part and stored in the learning data storage unit 8 as output side teaching data.

Next, the neural network is used to learn the relationship between the font characteristics and the placement positions of the parts. It is considered that the arrangement position of the part of a certain character font is determined by the approximate shape of the character and the degree of congestion around the part.

Therefore, as shown in FIG. 5B, the approximate shape of the character, which is obtained by reducing the resolution of the dot pattern of the reference font, is shown as the degree of congestion in the surroundings.
Learning the dot pattern of the standard pattern when the center of the mask is aligned with the coordinates of the same part of the standard font, the coordinates of the same part of the standard font, and the kind of the part as the input side teacher data of the neural network 7. Using the difference data of the relevant part in the data storage unit 8 as the output side teaching data, learning is performed by the known back propagation as many as the number of the created font parts.

That is, data of the same character as the created character is taken out from the reference font data storage unit 1, and the dot pattern expansion unit 2 expands the character into a dot pattern.
It is stored in the dot pattern storage unit 3.

Then, the resolution conversion unit 4 causes the dot pattern (for example, 256 × 2) of the dot pattern storage unit 3 to be generated.
56) is a dot pattern with reduced resolution (for example, 1
6x16) and input.

Furthermore, when the coordinates of the part to be learned and the dot pattern of the dot pattern storage unit 3 are input to the pattern cutout unit 5, as shown in FIG. 5B, the coordinates of the input component are centered. , A pattern whose resolution decreases as the distance from the coordinates is increased.

The output of the resolution conversion unit 4, the output of the pattern cutout unit 5, and the coordinates of the parts of the standard font are input to the neural network 7 and correspond to the type of the parts of the standard font of the part identification unit 6. The unit is set to “1”, the input side teacher data of the neural network 7 is used, and the difference data of the relevant part of the learning data storage unit 8 is used as the output side teacher data of the neural network 7, and learning by known back propagation is performed.

When the neural network 7 is made to learn as many as the parts of the created character font, a function (rule) for forming the difference between each part position of the standard character font and each part position of the created character font is formed. It

After learning the neural network 7 in this way, the placement position of the component is determined by the neural network 7. Reference font data storage unit 1
From the above, the reference font data of the character to be generated is extracted, and similarly to the learning, the dot pattern expansion unit 2 expands the character into a dot pattern, and the dot pattern storage unit 3
The dot pattern (for example, 256 × 256) of the dot pattern storage unit 3 stored in the resolution conversion unit 4 has a reduced resolution (for example, 16 × 16).
When the coordinates of the target component and the dot pattern of the dot pattern storage unit 3 are input to the pattern cutout unit 5, the coordinates of the input component are converted into the coordinates of the target component, as shown in FIG. 5B. As a center, a pattern whose resolution decreases as the distance from the coordinates is increased is generated.

The output of the resolution conversion unit 4, the output of the pattern cutout unit 5, and the coordinates of the parts of the standard font are input to the neural network 7 and correspond to the type of the parts of the standard font of the part identification unit 6. When the unit is set to "1" and input to the neural network 7,
In the output of the network 7, the difference between the part position of the standard font and the same part of the same character as the input standard font in the generated character font is output.

However, since this output is normalized between 0 and 1, the character data generator 9 converts it to the size of the original coordinate system and adds the position information of the relevant part of the reference font. As a result, it becomes the layout position information of the generated font.

With reference to FIG. 6, in order to generate a Mincho typeface "Emoji" from the standard font "Emoji" as shown in FIG. 6A, as shown in FIG. When the type of a font component (such as the left end of a stroke), the coordinates (X1m, Y1m) of that component, the standard pattern with reduced resolution, and the cutout pattern are input to the neural network 7, the standard pattern “Pen The positional difference ΔX, ΔY between the arrangement position of the part of “Mincho” and the arrangement position of the part of the Mincho body pattern “詔” is output, and this is the size of the original coordinate system (for example, 256 ×
256), and the data generating unit 9 subtracts from the arrangement positions X1m and Y1m of the corresponding part of the standard pattern “詔”, the arrangement position X1 of the corresponding part of the Mincho type pattern “詔”.
n and Y1n are obtained in the same manner as in FIG.

If this is repeated for each part of the standard pattern "詔", the arrangement position of each part of the Mincho body pattern "詔" is obtained, and the character data storage unit 10 shows the kind of parts and the communication relationship between the parts. The character data of the Mincho body pattern "詔" can be obtained by storing the character data in.

Here, the character data storage unit 10 stores only the coordinate data of each component, but since the shape pattern of the component of the typeface is stored separately depending on the type of the component,
Depending on the type of the part, if the shape pattern of the part of the typeface is drawn out and placed at the coordinate position, it can be expanded to the character pattern of the typeface.

In this way, the standard font is divided into each part, and the neural network 7 learns the relationship between the position of each part of the standard font and the position of the corresponding part of the created font to create another part. The font can be generated with the data of the corresponding standard font, and it is possible to generate a high-quality font having the position of the part corresponding to the typeface.

Further, in this embodiment, since the arrangement position difference between each part of the standard font and each part of the generated font is output, the influence of the error of the neural network 7 is small, and the generation of a font of higher quality is possible. It will be possible.

(B) Description of Other Embodiments In addition to the above embodiments, the present invention can be modified as follows. In the above-described embodiment, the arrangement position difference between each component of the standard font and each component of the generated font is output, but the component arrangement position of the generated font is calculated by the neural network 7.
If the output teacher data is output, the component arrangement position of the generated font is output from the neural network 7, and addition / subtraction with the component position of the standard font is not necessary, and this may be done.

As standard font data, each component is shown in absolute coordinates, but as is well known, the character "word" is used.
Is expressed at the character level of "Word", "Five", and "Mouth", and this level is expressed at the level of parts such as peculiarity and making (for example, "Word" is "-", "-"). , "-",
“−”, “Mouth”), and this may be expressed at the stroke component level as shown in FIG. 4, and the position may be indicated by relative coordinates.

Although the coordinate data of each part is stored in the character data storage unit 10, the shape pattern of each part may be allocated to form a character pattern and stored. Although the standard font is the skeleton pattern and the creation pattern is the Mincho type pattern, other typefaces may be used.

The present invention has been described above with reference to the embodiments.
Various modifications are possible within the scope of the invention, and these modifications are not excluded from the scope of the invention.

[0054]

As described above, according to the present invention,
It has the following effects. You can create fonts for all characters automatically by creating fonts for some of the character fonts you want to create.

Some character fonts among the fonts to be created are manually created, the character fonts are divided into parts, and the relationship between each part position of the standard font and each part position of the created font is neural network. 7
If you use this network, you can obtain the position of each part of the created font from the position of each part of the standard font even for a character font that has not been created. Can be obtained from the font.

The neural network creates a conversion rule by simply creating some of the created fonts, so that even an amateur can easily create a character font of a desired typeface without requiring much specialized knowledge. , It is possible to easily create a unique character font that takes advantage of individual characteristics.

[Brief description of drawings]

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

FIG. 2 is a configuration diagram of an embodiment of the present invention.

FIG. 3 is an explanatory diagram of a character font according to an embodiment of this invention.

FIG. 4 is an explanatory diagram of font data according to an embodiment of the present invention.

FIG. 5 is an explanatory diagram of pattern cutting according to an embodiment of the present invention.

FIG. 6 is an explanatory diagram of a generation operation according to an embodiment of the present invention.

FIG. 7 is an explanatory diagram of a conventional technique.

[Explanation of symbols]

 1 Reference font data storage unit 2 Dot pattern expansion unit 3 Dot pattern temporary storage unit 4 Resolution conversion unit 5 Pattern cutout unit 6 Component identification unit 7 Neural network 8 Learning data storage unit 9 Character data generation unit 10 Character data storage unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shigemi Nagata 1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture Fujitsu Limited

Claims (6)

[Claims] 1. A character font generation method for generating a character font of another typeface by dividing a character font of one typeface into fine parts and replacing the part with a component of a character font of another typeface. The position of the part and the position of the part of the character font of the other typeface are set with the rough shape of the character font of the one typeface, the degree of congestion of the parts forming the character font, and the type of the part as parameters. The neural network (7) is made to learn the relationship between the character font of the one typeface, the rough shape of the parts constituting the character font, and the kind of the parts as parameters, and Character font generation by a neural network, characterized in that the position is input to the neural network (7) to obtain the position of the part of the character font of the other typeface. Method. 2. With the rough shape of the character font of the one typeface, the degree of congestion of the parts constituting the character font, and the kind of the part as parameters, the position of the part is determined by the neural network (7). ) Input side teacher data, the difference between the positions of the character font parts of the one typeface and the character font parts of the other typeface is used as the output side teacher data of the neural network (7), and the neural network ( 7. A method for generating a character font by a neural network according to claim 1, wherein the method 7) is learned. 3. The position of the part is determined by the neural network (7) with the rough shape of the character font of the one typeface, the degree of congestion of parts constituting the character font, and the kind of the part as parameters. ) Is used as the input side teaching data, and the position of the part of the character font of the other typeface is used as the output side teaching data of the neural network (7) to learn the neural network (7). A method for generating a character font by the neural network of 1. 4. The neural network according to claim 1, wherein the character font of the one typeface has a reduced shape of the character font of the one typeface. Character font generation method. 5. The data through a mask having a high position resolution and a low surrounding resolution is used as the congestion degree of the components forming the character font. Character font generation method by Neural Network of K. 6. A character font generation method for generating a character font of another typeface by dividing the character font of one typeface into fine parts and replacing the part with a component of a character font of another typeface. The position of the part and the position of the part of the character font of the other typeface are set with the rough shape of the character font of the one typeface, the degree of congestion of the parts forming the character font, and the type of the part as parameters. The neural network (7) having learned the relationship with the part, using the rough shape of the character font of the one typeface, the degree of crowding of the parts forming the character font, and the kind of the part as parameters, A character font generation method by a neural network, characterized in that the position of the character font component of the other typeface is obtained by inputting the position of the character font.