KR20200070094A - Method and system for generating 3d image of character - Google Patents

Abstract

The present invention provides a method and system for generating a 3D image of character. The method includes the steps of: receiving an identifier of a template and a 2D image of the character generated based on the template corresponding to a type of the character; acquiring template information of the template by using the identifier, and extracting an effective parameter for machine learning to be performed to generate a 3D image based on the template information and the effective parameter; and generating the 3D image of the character by performing machine learning based on the effective parameter and the template information.

Description

Method and system for generating 3D images of characters {METHOD AND SYSTEM FOR GENERATING 3D IMAGE OF CHARACTER}

The present description relates to a method and system for generating a 3D image of a character from a 2D image using a template stored in advance.

Generally, 3D image restoration is performed manually by a designer using 3D modeling software. Therefore, in order to generate a high-quality 3D image of a character, a lot of time and professional know-how are required for 3D modeling. In addition, the quality of the three-dimensional image varies depending on the skill level of the designer.

The 3D image of the character can be created using computer vision and image processing. According to a method using computer vision and image processing, feature points are extracted through analysis of multiple images at any point in time, camera positions in 3D space are calculated based on the extracted feature points, and a plurality of 2D images The character's three-dimensional image can be restored by connecting a plurality of two-dimensional images based on the context before and after.

Recently, as artificial intelligence technologies such as large-scale databases and deep learning have developed, computer vision problems that are difficult to interpret using conventional methods have begun to be solved, and deep learning techniques have been applied to the restoration of 3D models (characters, etc.). . A typical example is a TL-embedding network that combines structures of autoencoders and alexnets that learn shape information of two-dimensional images and corresponding three-dimensional images. When the shape information of a plurality of images and 3D images represented by voxels is input, the TL-embedding network can learn the input data and generate the most similar 3D voxel model. The 3D Recurrent Reconstruction Neural Network is a deep learning network based on a long short-term memory structure that performs machine learning by inputting 2D images captured at various points in time and reconstructs a 3D voxel model. The 3D regression reconstruction neural network showed a relatively good reconstruction performance for a live image, compared to a conventional single 2D image-based reconstruction technique.

Deep learning-based 3D reconstruction technology using 2D images has been actively studied since the release of Shapenet's data set, and showed excellent results compared to traditional technologies. However, the reconstructed 3D model's voxel resolution is still low resolution (eg, 64X64X64).

One embodiment provides a method for generating a three-dimensional image of a character.

Another embodiment provides a system for generating a three-dimensional image of a character.

Another embodiment provides an apparatus for pre-processing image information for generating a three-dimensional image of a character.

According to one embodiment, a method of generating a three-dimensional image of a character is provided. The method for generating a 3D image includes receiving a 2D image of a character and an identifier of a template generated based on a template corresponding to a character type, obtaining template information of a template from a template library using the identifier, and Extracting valid parameters for machine learning to be performed to generate a 3D image based on information and a 2D image, and performing machine learning based on valid parameters and template information to generate a 3D image of a character Steps.

In the 3D image generation method, the template information includes the gender of the character and the age group of the character, and the 2D image may indicate a character of the age group having gender.

In the method of generating a 3D image, the template information further includes a character style, and the step of generating a 3D image of the character by performing machine learning based on valid parameters and template information is to transform the 3D image according to the style. And performing machine learning using an inference engine.

In the 3D image generation method, the template includes template information and a silhouette, and the 2D image may be an image of a character created based on the silhouette.

In the 3D image generation method, an effective parameter may include a plurality of template images mapped to each body part of the silhouette.

The 3D image generation method further includes determining a correlation between the 3D template of the template and the body part of the character based on the 2D image, and performing machine learning based on valid parameters and template information to perform character learning. The step of generating a 3D image of may include performing machine learning to synthesize a plurality of template images using correlation.

According to another embodiment, a system for generating a three-dimensional image of a character is provided. The 3D image generation system includes a template library for storing a plurality of templates representing a character type, template information of a first template among a plurality of templates, and a 2D image of a character generated based on the first template, 3 It includes a pre-processing unit for extracting valid parameters for machine learning to be performed to generate a dimensional image, and a modeling unit for generating a three-dimensional image of a character by performing machine learning based on the valid parameters.

In the 3D image generation system, the template information includes the gender of the character and the age group of the character, and the 2D image may indicate a character of the age group having gender.

In the 3D image generation system, the template information further includes a character style, and the modeling unit may perform machine learning using an inference engine that transforms the 3D image according to the style.

In the 3D image generation system, the first template includes template information and a silhouette, and the 2D image may be an image of a character produced based on the silhouette.

In the 3D image generation system, the effective parameter may include a plurality of template images mapped to each body part of the silhouette.

In the 3D image generation system, the preprocessor determines a correlation between the 3D template of the template and the body part of the character based on the 2D image, and the modeling unit uses the correlation to synthesize a plurality of template images. Machine learning can be performed.

According to another embodiment, an apparatus for preprocessing image information of a character is provided. The pre-processing apparatus includes a processor, a memory, and a transmission/reception interface, and the processor executes a program stored in the memory, and transmits and receives the identifier of the two-dimensional image and template of the character, generated based on the template corresponding to the character type. Receiving through, obtaining template information of a template using an identifier, extracting valid parameters for machine learning to be performed to generate a 3D image, using the template information and a 2D image, and a template Structure the information and valid parameters into a data structure, and transfer the data structure through a transmission/reception interface to a modeling unit performing machine learning.

In the pre-processing device, the processor further executes a program to determine a correlation between the 3D template of the template and the body part of the character based on the 2D image, and the processor data the template information and valid parameters. When performing a step of structuring into a structure and transmitting a data structure through a transmission/reception interface to a modeling unit performing machine learning, a step of structuring a correlation with a template information and valid parameters into a data structure may be performed.

In the pre-processing device, the template information includes a character style, the data structure includes instructing information about the inference neural network of the modeling unit, and the instructing information can instruct the modeling unit of the inference neural network corresponding to the body part.

In the pre-processing device, the template information includes the character's style, the data structure includes instructing information about the inference neural network of the modeling unit, and the instructing information can instruct the modeling unit of the inference neural network corresponding to the style.

Effective parameters that can increase the accuracy and speed of the inference neural network can be extracted through the template information of the 2D image and the template produced based on the template stored in the library in advance. Before the step of generating a 3D image by the inference neural network, by performing pre-processing to extract valid parameters suitable for machine learning, the optimal inference neural network can be selected, and the multiplexing of the inference engine within the inference neural network can be optimized.

1 is a block diagram showing an image generation system for generating a 3D image of a character according to an embodiment.
2A is a conceptual diagram showing the silhouette and template information of a 2D template according to an embodiment, and FIG. 2B is a conceptual diagram showing a 2D image of a character drawn on the silhouette of a 2D template according to an embodiment.
3 is a flowchart illustrating a method of generating a 3D image of a character according to an embodiment.
4 is a flowchart illustrating a method of generating a 3D image of a character according to another embodiment.
5 is a block diagram showing a pre-processing apparatus for pre-processing image information for generating a three-dimensional image of a character according to an embodiment.

Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those skilled in the art to which the present invention pertains may easily practice. However, the present description can be implemented in many different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present description in the drawings, parts irrelevant to the description are omitted, and like reference numerals are assigned to similar parts throughout the specification.

1 is a block diagram showing an image generating system for generating a 3D image of a character according to an embodiment, FIG. 2A is a conceptual diagram showing silhouette and template information of a 2D template according to an embodiment, and FIG. 2B is a It is a conceptual diagram showing a two-dimensional image of a character drawn on the silhouette of a two-dimensional template according to an embodiment.

Referring to FIG. 1, an image generation system for generating a 3D image of a character according to an embodiment includes a pre-processing unit 100, a template library 200, and a modeling unit 300. The user may obtain a 3D image from the 2D image of the character using the image generation system according to an embodiment. The user may create a digital 2D image or an analog 2D image using the template obtained from the template library 200. For example, a user downloads a template from the template library 200 through a communication network, and uses a tablet PC and a stylus pen to create a digital two-dimensional image of the character on the downloaded template silhouette. Alternatively, the user outputs the silhouette of the template downloaded from the template library 200 using a printer, draws a two-dimensional image of the character on the output silhouette, and then draws a hand-drawn two-dimensional image using a scanner or a photographing device. It can be converted into a digital 2D image. Since the 2D image is fabricated to be superimposed on the silhouette of the template, the modeling unit 300 may generate a more accurate 3D image by performing machine learning such as deep learning.

When the identifiers of the templates used in the creation of the digital two-dimensional image and the two-dimensional image of the character are input to the image generation system, the image generation system receives three characters of the character based on the identifier of the two-dimensional image and the template of the character received from the user. You can create dimensional images. The user may input the identifiers of the two-dimensional image and the template of the character into the image generation system, or the two-dimensional image of the character and the downloaded template into the image generation system. The pre-processing unit 100 may receive a 2D image of a character and an identifier of a template used in the production of the 2D image from the 2D image generation device 10 through the first interface IF1. When the identifier of the template is input to the pre-processing unit 100, the pre-processing unit 100 of the image generation system may use the identifier of the template to search for the template in the template library 200 and use the template information of the searched template. .

The pre-processing unit 100 according to an embodiment includes a first analysis unit 110, a second analysis unit 120, and a structured unit 130.

The first analysis unit 110 of the pre-processing unit 100 according to an embodiment may extract an effective parameter based on the identifier of the template used to produce the two-dimensional image and the two-dimensional image of the character received from the user. . The valid parameters may include a plurality of template images mapped to each body part of the silhouette of the template. For example, when the user draws a two-dimensional image of a character on the silhouette of the template as shown in FIG. 2B, the first analysis unit 110 divides the two-dimensional image into a head portion, a body portion, and a lower body portion, thereby cutting the silhouette of the template. Using can generate a plurality of template images from the two-dimensional image. At this time, the plurality of template images respectively correspond to each part of the silhouette. The template used by the first analysis unit 110 may be a 2D template or a 3D template.

The second analysis unit 120 of the preprocessing unit 100 according to an embodiment may additionally analyze a 2D image based on a 3D template. When there is a 3D template corresponding to the identifier of the template used in the production of the 2D image, the second analysis unit 120 may determine a correlation between the body parts of the character based on the 3D template and the 2D image. .

The structuring unit 130 of the pre-processing unit 100 according to an embodiment may structure template information of a template used to produce a two-dimensional image of a character and effective parameters extracted by the first analysis unit 110. . The structurer 130 may also structure the original two-dimensional image of the character using a predetermined format along with template information and valid parameters. The template information structured by the structuring unit 130 and the valid parameters may constitute a data structure used to be transmitted to the modeling unit 300. The data structure may further include a two-dimensional image of the character. When the pre-processing unit 100 and the modeling unit 300 are connected to each other through a wired/wireless network, the character's 2D image, template information, and valid parameters are wrapped in a secure manner through structuring to pre-process the 100 In the can be delivered to the modeling unit 300. The structurer 130 may structure two-dimensional images, template information, and valid parameters in a format such as Extensible Markup Language (XML). The structuring unit 130 according to another embodiment may further structure a correlation between a character's body part along with a character's two-dimensional image, a template, and valid parameters.

The data structure structured by the structurer 130 may include instructional information about the reasoning neural network of the modeling unit 300. The indication information about the inference neural network included in the data structure may indicate a reasoning engine or a reasoning neural network corresponding to each body part. For example, when the silhouette includes the head, the trunk, and the lower body as body parts, and the lower body part does not correspond to the two-dimensional image, the indication information about the inference neural network is the head to generate the three-dimensional image of the head part (頭部) The inference engine and the body inference engine that generates a three-dimensional image of the torso can be directed. Or, when a two-dimensional image is associated with each body part included in the silhouette, and an image that does not correspond to each body part (for example, a weapon held in the hand) is included in the two-dimensional image of the character, the inference neural network The instructional information regarding the periphery inference engine for generating a three-dimensional image of an image that does not correspond to a body part (eg, an equipment inference engine for generating a three-dimensional image of equipment such as weapons) may be indicated. Alternatively, the instruction information about the reasoning neural network may indicate a reasoning engine or a reasoning neural network corresponding to a style (cute style, elegant style, etc.). Thereafter, the modeling unit 300 selects an inference engine or an inference neural network that transforms the 3D image according to the style, and performs machine learning by multiplying the weight of the character's style determined by the template with the input valid parameter. It can be done. For example, when a character's style is pre-determined as'cute' by a template, the instruction information about the inference neural network may indicate an inference engine or an inference neural network suitable for transforming a 3D image according to the style. At this time, the type of inference neural network for generating a 3D image of the character may be determined according to the instruction information included in the data structure. The predetermined format used by the structuring unit 130 for structuring may be a data format that can be used for a newly developed reasoning engine and a cloud-based AI service platform.

In the template library 200 according to an embodiment, various templates are stored according to a character's gender, age, and style (fat, cute, etc.). That is, the template corresponds to each type of character, such as the character's gender, age group, and style, and the type of character represented by one template is included in the template as template information. That is, the template information may include gender, age group, style, and the like of the character represented by the template. In order to create an in-game character, the template library 200 may store different templates for each race (race or tribe) in the game. That is, different templates may exist depending on the character's race. An identifier is assigned to each template, and a 2D template and a 3D template can be associated with one identifier. Each template includes at least one silhouette representing the front, back, side, plane, bottom, etc. of the character. The 3D template having the same identifier as the 2D template used in the production of the 2D image of the character may be used when the correlation between the body parts of the character is determined by the second analysis unit 120. The template library 200 may be improved through values obtained in the process of extracting valid parameters.

In the silhouette of the template shown in FIG. 2A, M001 is an identifier of the template, and the M001 template is a template for producing a character who is an adult man. For example, according to the silhouette of the M001 template, the user draws the character's front face over the head silhouette, draws the character's upper coat over the body silhouette, and draws the character's lower coat over the lower body silhouette. The two-dimensional image may be drawn by the user in the silhouette in the other direction of the identifier M001.

The modeling unit 300 according to an embodiment includes a distribution unit 310, a first reasoning neural network 320, and a second reasoning neural network 330. The modeling unit 300 may further include another reasoning neural network according to the type of the 3D image. The modeling unit 300 may generate a 3D image by receiving the data structure from the pre-processing unit 100 through the second interface IF2 and driving the inference neural network using the data structure.

The distribution unit 310 may determine an inference neural network on which machine learning for generating a 3D image is performed based on the instruction information included in the data structure. Alternatively, the distribution unit 310 may determine an inference neural network in which machine learning is to be performed according to the type of template used in the production of the 2D image.

The first inference neural network 320 includes multiplexed inference engines NN ₁ to NN _m , and performs machine learning by driving a plurality of inference engines based on valid parameters and template information included in a data structure. . For example, the first inference neural network 320 may perform a weighted sum operation to generate a 3D image suitable for the style of the template among template information. That is, the first reasoning neural network 320 may be a weighted neural network. The first inference neural network 320 may be suitable for generating a 3D character image when a plurality of template images included in the data structure are distinguished as one collective characteristic. The learning data may be specialized for each reasoning engine included in the first reasoning neural network 320, and the optimal reasoning engine for generating a 3D character image suitable for the style of the template transmitted through the data structure is machine learning. Can be chosen for.

The second inference neural network 330 includes a plurality of inference engines P ₁ to P _n multiplexed, and through a plurality of template images and three-dimensional templates each corresponding to a body part among effective parameters included in the data structure. Based on the determined correlation, a 3D image of the character can be generated by driving the inference engine. That is, the second inference neural network 330 may be a partial 3D reconstruction neural network. The second reasoning neural network 330 may perform machine learning to synthesize a template image corresponding to each body part using a correlation between each body part determined through a 3D template.

The 3D image of the character generated by the first inference neural network 320 or the second inference neural network 330 of the modeling unit 300 may be fed back to the user who generated the 2D image. Alternatively, the 3D image of the character, generated by the modeling unit 300, may be transmitted to a display device that displays content to be used by the character. When the 3D image is transmitted to the display device, the data of the 3D image may be converted according to the performance profile of the 3D image receiving device 20, wherein the data conversion is performed by major elements such as the shape or texture of the 3D image. It can be performed within a range that is not lost.

3 is a flowchart illustrating a method of generating a 3D image of a character according to an embodiment.

Referring to FIG. 3, the pre-processor 100 receives a 2D image generated by a user using a template of the template library 200 and an identifier of a template used when a 2D image of a character is generated. (S110). The pre-processing unit 100 may query template information of the template used when the 2D image of the character is generated based on the identifier of the template from the template library 200. Referring to FIG. 2A, the preprocessor 100 may obtain template information-gender: man, age: adult-from the template library 200 using the identifier M001. Next, the first analysis unit 110 of the pre-processing unit 100 extracts an effective parameter for machine learning to generate a 3D image based on the 2D image and template information of the character (S120). The first analysis unit 110 according to an embodiment may extract an effective parameter by dividing a two-dimensional image into a plurality of template images using template information.

Thereafter, the structuring unit 130 of the pre-processing unit 100 constructs a data structure by structuring valid parameters and template information in a predetermined format, and delivers the generated data structure to the modeling unit 300 (S130).

The distribution unit 310 of the modeling unit 300 selects an inference neural network for generating a 3D image among a plurality of inference neural networks included in the modeling unit 300 or connected to the modeling unit 300 based on the data structure. Can be. When selecting the inference neural network, the distribution unit 310 may use indication information included in the data structure or template information included in the data structure. When the correlation between each body part corresponding to the plurality of template images is not included in the data structure, the distribution unit 310 may select the first inference neural network 320 performing the weight sum operation. Thereafter, the first inference neural network 320 generates a three-dimensional image of the character through machine learning that performs a weighted sum operation based on valid parameters and template information (S140).

4 is a flowchart illustrating a method of generating a 3D image of a character according to another embodiment.

Referring to FIG. 4, the pre-processing unit 100 receives a two-dimensional image of a character generated by a user using a template and an identifier of a template used for generating the two-dimensional image from the user (S210). The method in which the pre-processing unit 100 obtains template information based on the identifier of the template is the same as in FIG. 3, and thus will not be described again. Next, the pre-processing unit 100 extracts an effective parameter for machine learning to generate a 3D image using the first analysis unit 110 and the second analysis unit 120 (S220). As illustrated in FIG. 3, the first analysis unit 110 may extract an effective parameter by dividing a two-dimensional image into a plurality of template images using template information. The second analysis unit 120 may determine a correlation between body parts of the character using a 3D template having the same identifier or corresponding to the identifier received from the user (S230). Correlation may correspond to each template image.

Thereafter, the structuring unit 130 of the pre-processing unit 100 constructs a data structure by structuring valid parameters, template information, and correlation between each body part in a predetermined format, and modeling the generated data structure 300 It is delivered to (S240).

The distribution unit 310 of the modeling unit 300 may select an inference neural network for generating a 3D image among a plurality of inference neural networks included in or connected to the modeling unit 300 based on the data structure. have. Since the correlation between each body part corresponding to the plurality of template images is included in the data structure, the distribution unit 310 partially reconstructs the 3D image by selecting the second inference neural network 330, and the 3D part The final three-dimensional image can be determined by combining the images. That is, the second reasoning neural network 330 generates a 3D partial image for each body part of the character by performing machine learning based on the correlation between valid parameters, template information, and body parts (S250), and correlation between body parts Based on the relationship, a 3D partial image may be combined to generate a final 3D image of the character (S260).

As described above, effective parameters capable of increasing the accuracy and speed of the inference neural network may be extracted through the template information of the 2D image and the template produced based on the template stored in the library in advance. Before the step of generating a 3D image by the inference neural network, by performing pre-processing to extract valid parameters suitable for machine learning, the optimal inference neural network can be selected, and the multiplexing of the inference engine within the inference neural network can be optimized.

5 is a block diagram showing a pre-processing device for pre-processing image information according to an embodiment.

The pre-processing apparatus according to an embodiment may be implemented as a computer system, for example, a computer-readable medium. Referring to FIG. 5, the computer system 500 includes a processor 510 communicating through a bus 570, a memory 530, an input interface device 550, an output interface device 560, and a storage device 540 ). Computer system 500 may also include a communication device 520 coupled to the network. The processor 510 may be a central processing unit (CPU) or a semiconductor device that executes instructions stored in the memory 530 or the storage device 540. The memory 530 and the storage device 540 may include various types of volatile or nonvolatile storage media. For example, the memory may include read only memory (ROM) and random access memory (RAM). In the embodiments of the present disclosure, the memory may be located inside or outside the processor, and the memory may be connected to the processor through various known means. Memory is a volatile or non-volatile storage medium of various types, and for example, the memory may include read-only memory (ROM) or random access memory (RAM).

Thus, embodiments of the present invention may be implemented as a computer-implemented method or as a non-transitory computer-readable medium having computer-executable instructions stored thereon. In one embodiment, when executed by a processor, computer readable instructions may perform a method according to at least one aspect of the present disclosure.

The communication device 520 may transmit or receive a wired signal or a wireless signal.

Meanwhile, the embodiment of the present invention is not implemented only through the apparatus and/or method described so far, and may be implemented through a program that realizes a function corresponding to the configuration of the embodiment of the present invention or a recording medium in which the program is recorded. There is, such an implementation can be easily implemented by those skilled in the art to which the present invention pertains from the description of the above-described embodiment. Specifically, a method according to an embodiment of the present invention (eg, a network management method, a data transmission method, a transmission schedule generation method, etc.) is implemented in the form of program instructions that can be performed through various computer means, and is used for computer-readable media. Can be recorded. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the computer-readable medium may be specially designed and configured for an embodiment of the present invention, or may be known and usable by those skilled in the computer software field. The computer-readable recording medium may include hardware devices configured to store and execute program instructions. For example, computer readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and floptical disks. It may be the same magneto-optical media, ROM, RAM, flash memory, and the like. The program instructions may include machine language codes such as those produced by a compiler, as well as high-level language codes that can be executed by a computer through an interpreter or the like.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concept of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

Claims (16)

As a method of generating a 3D image of a character,
Receiving a two-dimensional image of the character and an identifier of the template, generated based on a template corresponding to the character type,
Obtaining template information of the template from a template library using the identifier, and extracting valid parameters for machine learning to be performed to generate the 3D image based on the template information and the 2D image, and
Generating the 3D image of the character by performing the machine learning based on the valid parameter and the template information
3D image generation method comprising a. In claim 1,
The template information includes the gender of the character and the age group of the character, and the two-dimensional image represents a character of the age group having the gender, a 3D image generating method. In claim 2,
The template information further includes the style of the character,
Generating the three-dimensional image of the character by performing the machine learning based on the valid parameters and the template information,
Performing the machine learning using an inference engine that transforms a 3D image according to the style
Including, 3D image generation method. In claim 1,
The template includes the template information and a silhouette, and the two-dimensional image is an image of the character produced based on the silhouette. In claim 4,
The effective parameter includes a plurality of template images mapped to each body part of the silhouette, the method of generating a three-dimensional image. In claim 5,
Determining a correlation between the 3D template of the template and the body part of the character based on the 2D image
Further comprising,
Generating the three-dimensional image of the character by performing the machine learning based on the valid parameters and the template information,
Performing the machine learning to synthesize the plurality of template images using the correlation.
Including, 3D image generation method. A system that generates a 3D image of a character,
Template library for storing a plurality of templates indicating the type of the character,
Extracting valid parameters for machine learning to be performed to generate the 3D image, from the template information of the first template among the plurality of templates and the 2D image of the character generated based on the first template Pre-treatment unit, and
A modeling unit that generates the 3D image of the character by performing the machine learning based on the effective parameter
3D image generation system comprising a. In claim 7,
The template information includes the gender of the character and the age group of the character, and the two-dimensional image represents a character of the age group having the gender, a three-dimensional image generation system. In claim 8,
The template information further includes the style of the character,
The modeling unit,
A system for generating a three-dimensional image, performing the machine learning using an inference engine that transforms a three-dimensional image according to the style. In claim 7,
The first template includes the template information and a silhouette, and the two-dimensional image is an image of the character produced based on the silhouette, a three-dimensional image generating system. In claim 10,
The effective parameter includes a plurality of template images mapped to each body part of the silhouette, a three-dimensional image generation system. In claim 5,
The pre-processing unit,
The correlation between the three-dimensional template of the template and the body part of the character is determined based on the two-dimensional image,
The modeling unit,
A system for generating a 3D image to perform the machine learning so that the plurality of template images can be synthesized using the correlation. A device for pre-processing character image information,
A processor, a memory, and a transmission/reception interface, wherein the processor executes a program stored in the memory,
Receiving a two-dimensional image of the character and an identifier of the template generated through a template corresponding to the type of the character through the transmission/reception interface,
Obtaining template information of the template using the identifier, and using the template information and the two-dimensional image, extracting valid parameters for machine learning to be performed to generate the three-dimensional image, and
Structuring the template information and the effective parameter into a data structure, and passing the data structure through the transmission/reception interface to a modeling unit performing the machine learning
Pre-processing device. In claim 13,
The processor executes the program,
Determining a correlation between the 3D template of the template and the body part of the character based on the 2D image
Do more,
When the processor performs the step of structuring the template information and the valid parameters into a data structure, and passing the data structure through the transmission/reception interface to a modeling unit performing machine learning,
Structuring the correlation with the template information and the valid parameters into the data structure
Pre-processing device. In claim 13,
The template information includes the style of the character, the data structure includes instruction information about the reasoning neural network of the modeling unit, and the instruction information pre-processes the modeling unit to indicate the reasoning neural network corresponding to the body part. Device. In claim 13,
The template information includes the style of the character, the data structure includes instructional information about the speculative neural network of the modeling unit, and the instructional information instructs the modeling unit of the speculative neural network corresponding to the style. .

Images