KR102348368B1 - Device, method, system and computer readable storage medium for generating training data of machine learing model and generating fake image using machine learning model - Google Patents

Abstract

A method of generating training data for a machine learning model that performs machine learning-based similar image generation is described. An exemplary method of generating training data includes: acquiring an image source; analyzing the image source to set a learning section in the image source; extracting a label image from the set learning section; extracting at least one adjacent image based on the set learning section and the label image; and based on the label image and the at least one adjacent image, generating training data for input to the machine learning model.

Description

DEVICE, METHOD, SYSTEM AND COMPUTER READABLE STORAGE MEDIUM FOR GENERATING TRAINING DATA OF MACHINE LEARING MODEL AND GENERATING FAKE IMAGE USING MACHINE LEARNING MODEL}

The present disclosure relates to an apparatus, method, system, and computer-readable storage medium for generating training data of a machine learning model and generating a similar image using the machine learning model.

Unless otherwise stated herein, the subject matter described in this section is not prior art to the claims in this application, and should not be admitted as prior art on the grounds that it is recited in this section.

As the use of personal communication devices, such as smart phones, is generalized and transmission speed increases due to the development of communication technology, available online contents are explosively increasing. Unlike the text-oriented conventional online content, most of the recent online content includes an image or video data composed of a plurality of images. As such, as the number of images online increases dramatically, the problem of copyright infringement with the use of such images follows.

With the recent development of machine learning techniques, it is considered to generate images using machine learning techniques. Using machine learning techniques to generate images provides cost and time benefits as users do not have to take or create photos themselves. It may also be possible to avoid the problem of copyright infringement if machine learning techniques are used appropriately.

"Generative Adversarial Nets, Ian J. Goodfellow et al." (hereinafter, "Prior Art Document 1") proposes a model for evaluating an image generation model using machine learning through an adversarial process. According to this model, the performance of the generative network is improved by generating a similar image in the generative network according to the machine learning technique, comparing the generated similar image with the real image, and evaluating whether the generated similar image is a real image and feedback. make it In "StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation, Choi, et al." (hereinafter, "Prior Art Document 2"), using Prior Art Document 1 as a basic structure, We present a model that can generate similar images reflecting various domains.

However, in using this machine learning-based image generation technique, in order for the machine learning model to exhibit sufficient performance, it is necessary to input a very large amount of training data, ie, images, to the machine learning model to perform learning. This is essential for building machine learning models, but it is a very cumbersome task and can take a lot of time and money.

The present disclosure is intended to solve the above problems, and provides an apparatus, a method, and a computer-readable storage medium for generating training data of a machine learning model and generating a similar image using a machine learning model.

In some embodiments of the present disclosure, a method of generating training data is described. The exemplary method of generating training data may be performed under the control of a computing device and may generate training data of a machine learning model that performs machine learning-based similar image generation. The training data generation method includes: acquiring an image source; analyzing the image source to set a learning section in the image source; extracting a label image from the set learning section; extracting at least one adjacent image based on the set learning section and the label image; and based on the label image and the at least one adjacent image, generating training data for input to the machine learning model.

In some examples, the machine learning model may include a Generative Adversarial Network. The machine learning model may use at least one adjacent image to generate a similar image, and use the label image as a real image used to determine the generated similar image.

In some other embodiments, a similar image generating apparatus is described. The exemplary similar image generating apparatus may include an image source manager, a learning section setting unit, a label image extracting unit, an adjacent image extracting unit, and a similar image generating unit. The image source manager may acquire an image source. The learning section setting unit may analyze the image source and set the learning section in the image source. The label image extractor may extract a label image from the set learning section. The adjacent image extractor may extract at least one adjacent image based on the set learning section and the label image. The similar image generator includes a machine learning model that generates a machine learning-based image, generates a similar image using the machine learning model, and trains the machine learning model using the label image and at least one adjacent image. have.

In some still other embodiments, a computer-readable storage medium storing a program for generating learning data is described. The training data generation program may generate training data of a machine learning model that performs machine learning-based image generation. An exemplary learning data generating program, when executed by a computing device, includes: obtaining an image source; analyzing the image source to set a learning section in the image source; extracting a label image from the set learning section; extracting at least one adjacent image based on the set learning section and the label image; and one or more instructions to perform an operation of generating training data for input to the machine learning model based on the label image and the at least one neighboring image.

The above brief summary and description of effects are merely exemplary and are not intended to limit the technical matters intended in the present disclosure. In addition to the above-described exemplary embodiments and technical features, additional embodiments and technical features may be understood by referring to the following detailed description and accompanying drawings.

The foregoing and other features of the present disclosure will become sufficiently apparent from the following description with reference to the accompanying drawings. With the understanding that these drawings illustrate only a few embodiments in accordance with the present disclosure, and therefore should not be considered limiting of their scope, the present disclosure is described in greater detail and detail through the use of the accompanying drawings. will be
1 is a block diagram illustrating an apparatus for generating a similar image, according to at least some embodiments of the present disclosure.
FIG. 2 is a block diagram illustrating an example of the similar image generator of FIG. 1 .
3 illustrates an example in which a label image and an adjacent image used to generate training data are extracted, according to some embodiments of the present disclosure.
4 depicts an example process for generating training data, in accordance with at least some embodiments of the present disclosure.
5 illustrates a computer program product that may be used to generate training data of a machine learning model for generating similar images, in accordance with at least some embodiments of the present disclosure.
6 is an exemplary block diagram of a computing device arranged in accordance with at least some embodiments of the present disclosure.

Hereinafter, embodiments and examples of the present application will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art to which the present disclosure pertains can easily carry out. However, the present application may be embodied in many different forms and is not limited to the embodiments and examples described herein.

The present disclosure relates generally to a method, apparatus, and a computer-readable storage medium and a program stored thereon for generating training data used in a machine learning model that generates a similar image.

1 is a block diagram illustrating a similar image generating apparatus 100 according to at least some embodiments of the present disclosure. The image-like device 100 may be a single device such as, for example, a desktop computer, a laptop computer, a smartphone, a tablet computer, a mobile phone, a personal digital assistant (PDA), a special purpose device, or a fusion device comprising any of the above functions. computing device; Alternatively, it may be a multiple computing configuration, such as a plurality of computing devices, such as a server farm, a distributed network, or a cloud computing configuration. Accordingly, the similar image generating apparatus 100 may be implemented as a server that receives a request from a personal terminal device and transmits an output in response to the request, or may be implemented in a personal terminal device.

The similar image generating apparatus 100 includes an image source manager 110 , a learning section setting unit 120 , a label image extracting unit 130 , an adjacent image extracting unit 140 , and a similar image generating unit 150 . In an additional or optional embodiment, the similar image generating apparatus 100 may further include at least one of an image preprocessor 160 and a metadata extractor 170 .

The similar image generating apparatus 100 may store one or more programs according to the present disclosure in an internal memory, and the one or more programs may be executed by one or more processors, for example, a central processing unit (CPU) or a graphics processing unit (GPU). and the processing to perform at least some of the operations of the respective components 110 , 120 , 130 , 140 and 150 and the selection/additional components 160 and 170 of the similar image generating apparatus 100 described below. can Although these components 110 , 120 , 130 , 140 and 150 and optional components 160 and 170 are shown as separate components, they may be separated into additional components or fewer components without departing from the scope of the disclosed subject matter. can be combined or eliminated. In addition, although the component is illustrated as being implemented by software in FIG. 2 , each function and/or operation of the component may be individually and/or collectively implemented through hardware, software, firmware, or any combination thereof, to those skilled in the art. will understand that

The image source management unit 110 is configured to acquire an image source. In some examples, the image source manager 110 may collect the image source in a manner predetermined by the user. Various methods, such as acquisition through a network address of an open source repository, reception from a predetermined repository, and web crawling, may be used for the collection of the image source. In some other examples, the image source manager 110 may directly receive an image source from a user of the similar image generating apparatus 100 . In some other examples, the image source manager 110 may receive the image source from an external computing device associated with the similar image generating apparatus 100 or from an address of a network input by a user. The acquired image source may be stored in an internal memory of the similar image generating apparatus 100 . The image source may be various types of image data from which a plurality of images may be extracted, as described in detail below. The image source may include, for example, video data having a plurality of frame images, a panoramic image such as a 360 degree image, 360 degree video data, and the like. Additionally, the image source manager 110 may classify and store the acquired image source using a predetermined technique.

In some optional examples, the image source management unit 110 may receive an image source input from a user, the metadata extraction unit 170 extracts metadata from the input image source, and the image source management unit 110 receives the metadata. Based on the data, acquisition of additional image sources may be performed. In this example, the image source manager 110 may classify and store the image source based on the metadata.

The learning section setting unit 120 may set a learning section in the image source by analyzing the image source obtained by the image source management unit 110 . The learning section setting unit 120 may perform various analyzes in order to derive a learnable section within the image source, and may set the learning section based on the result of the analysis. The learning section setting unit 120 may perform, for example, at least one of measuring a color change rate of an image source, measuring image complexity, and measuring a structural similarity of an object in the image source.

In some examples, the learning interval setting unit 120 may measure a color changing ratio in the image source. When the image source is video data, the learning interval setting unit 120 may measure a color change rate between a plurality of frames in the video data. When the image source is a panoramic image, the learning section setting unit 120 may extract an individual image from the panoramic image using a frame of a predetermined size and measure a color change rate between the extracted images. The learning section setting unit 120 may set a section in which the measured color change rate is within a predetermined range, and may determine this section as the learning section.

In some other examples, the learning interval setting unit 120 may measure image complexity in the image source. When a predetermined object, for example, a specific landmark is included in the image source, the image including the object may have different image complexity from the image not including the object. In this example, the learning section setting unit 120 may exclude a section in which a corresponding object is not included, for example, a section having a complexity outside a predetermined range from the learning section to be set.

In some other examples, the learning interval setting unit 120 may measure structural similarity within the image source. There may be obstacles within the image source. For example, in video data for a specific landmark such as Gyeongbokgung, there may be obstacles such as pedestrians or vehicles passing the landmark. These obstacles may be predefined by the user or pre-stored data, and the learning section setting unit 120 may detect an object having a structure similar to the predefined obstacles from the image source, and the obstacles are included. Sections can be excluded from the learning section.

The label image extractor 130 may extract a label image from the learning section set by the learning section setting unit 120 . The label image extractor 130 may determine a label image from the image source within a learning section set for the image source. The label image is an image including a main object included in an image source, and may refer to a representative image. In some examples, the label image extraction unit 130 may extract the label image in an arbitrary manner within the learning section. In some other examples, the label image extraction unit 130 may extract an image located at the center of the set learning section, that is, the center image as the label image. When the image source is video data, the label image extractor 130 may determine the image of the center frame as the label image. In some other examples, the label image extractor 130 may determine, as the label image, an image that satisfies one or more characteristics or falls within one or more characteristic ranges among a plurality of images extracted within the learning section.

The adjacent image extractor 140 may extract at least one adjacent image based on the learning section set by the learning section setting unit 120 and the label image extracted by the label image extractor 130 . In some examples, the adjacent image extraction unit 140 may extract an image separated by a predetermined distance from the label image as the adjacent image by using the label image as a reference image of the learning section. In one example, the adjacent image extraction unit 140 may extract two images separated by a predetermined interval as adjacent images. In some other examples, the adjacent image extraction unit 140 may extract an arbitrary image within the learning section, measure the structural similarity between the extracted image and the label image, and the measured structural similarity is within a predetermined range (eg, If the structural similarity is 0.9 or more), the corresponding image may be extracted as an adjacent image.

The similar image generator 150 may include a machine learning model that generates a similar image based on machine learning. The similarity image generator 150 may generate training data for learning the corresponding machine learning model. Accordingly, the similar image generating unit 150 may generate training data based on the label image extracted by the label image extracting unit 130 and at least one neighboring image extracted by the neighboring image extracting unit 140 . have. The machine learning model of the similar image generator 150 may be, for example, a generative adversarial network (GAN), a star GAN, a rule-based system, or a neural network-based system. system) (e.g. convolutional neural networks (CNN)), feedforward neural networks (FNNs), recurrent neural networks (RNNs), etc. In some examples, the machine learning model may use at least one adjacent image to generate a similar image, and use the label image as a real image used to determine the generated similar image. An example will be described in more detail with reference to Fig. 2. The similar image generator 150 may use the generated training data to learn the machine learning model.

In this way, the label image extracted by the label image extraction unit 130 and the at least one adjacent image extracted by the adjacent image extraction unit 140 may be used as an input of the similar image generation unit 150 . In some additional and optional examples, the image preprocessor 160 may generate an image for additional training data by performing preprocessing using the label image and/or at least one adjacent image. In some examples, the image preprocessor 160 adds intentional noise to the label image and/or at least one adjacent image, thereby preventing overfitting from occurring in the training of the similar image generator by the generated image. can do. In some examples, the image preprocessor 160 may extract an image contour and compare the similarity by using Open Source Computer Vision (OpenCV) to generate additional data for learning.

In some additional and optional examples, the metadata extraction unit 170 may extract metadata of an image based on at least one of an image source, a label image, and at least one adjacent image. Such metadata may include, for example, data about the context of the image, such as the name of an object in the image, the subject of the image, weather on the image, time on the image, color distribution of the object, and the like. Such metadata may be used as learning data in addition to the label image and at least one adjacent image input to the similar image generating unit 150 . For example, the similarity image generating unit 150 may learn to generate the similarity image based on the label image and at least one adjacent image, as well as the metadata extracted by the metadata extracting unit 170 . .

As described above, according to the present disclosure, data used for learning a machine learning model is automatically generated just by supplying an image source from which a plurality of images can be extracted, for example, video data, an image of a predetermined size, etc. Since it can be supplied to the learning model, it is possible to promote the convenience of learning the machine learning model compared to the case where learning data needs to be provided passively, while improving the efficiency and speed of learning the machine learning model.

FIG. 2 is a block diagram illustrating an example of the similar image generator 150 of FIG. 1 . As shown in FIG. 2 , the similar image generator 150 includes a generative adversarial neural network (GAN), but it will be understood that the embodiment according to the present disclosure is not limited to such an image generation model. As shown in FIG. 2 , the similar image generating unit 150 includes a similar image generating network 210 , a fake image discriminate network 220 , a feedback unit 230 , and an output unit 240 . include

As described with reference to FIG. 1 , the similar image generator 150 generates training data based on the label image and at least one adjacent image. In some examples, the similar image generating network 210 may generate a similar image by using at least one adjacent image. The similar image determining network 220 determines whether the image generated by the similar image generating network 210 is a fake image or a real image. In some examples, the similar image determination network 220 uses the label image as an actual image to determine whether the generated image is a similar image or an actual image. The feedback unit 230 may provide feedback to the similarity image generating network 210 and the similarity image determining network 220 based on a result of the determination of the similarity image determining network 230 . In this way, the machine learning model in the similar image generating unit 150 may perform learning. The output unit 240 may output a similar image generated by the similar image generating network 210 according to a predetermined request or condition.

3 illustrates an example in which a label image and an adjacent image used to generate training data are extracted, according to some embodiments of the present disclosure. According to some embodiments of the present disclosure, a learning section 310 as shown in FIG. 3 may be set. When the image source is video data, a plurality of frame images of a predetermined time period may be included in the learning period. When the image source is a single image such as a panoramic image, a spatial range of a predetermined size larger than a label image to be subsequently extracted and an adjacent image may be set as the learning section 310 .

According to some embodiments of the present disclosure, the label image 320 may be extracted from the learning section 310 . In some examples, the label image 320 may be a central image of the learning section 310 . In some other examples, the label image 320 may be an image satisfying a predetermined condition.

After the label image 320 is extracted, according to some embodiments of the present disclosure, at least one adjacent image 330 may be extracted based on the label image 320 . In some examples, an image having a predetermined distance from the label image 320 may be selected as the adjacent image 330 . In some other examples, an image having a difference within a predetermined range from the label image 320 may be selected as the adjacent image 330 .

4 depicts an example process 400 of generating training data, in accordance with at least some embodiments of the present disclosure. For example, the process 400 may be performed under the control of a computing device, such as the similar image generating device 100 of FIG. 1 . The process 400 illustrated in FIG. 4 may include one or more acts, functions, or acts as illustrated by blocks 410 , 420 , 430 , 440 and/or 450 . The schematic operations illustrated in FIG. 4 are provided by way of example only, and some of the operations may be optional, may be combined into fewer operations, or expanded into additional operations without departing from the essence of the disclosed embodiment. have. Process 400 may begin at block 410 of obtaining an image source.

At block 410 , the computing device may obtain an image source. In some examples, the computing device may in some instances collect the image source in a predetermined manner. In some other examples, the computing device may directly receive an image source from a user. In some other examples, the computing device may receive the image source from an external computing device or from an address in the network entered by the user. The image source may include, for example, video data having a plurality of frame images, a panoramic image such as a 360 degree image, 360 degree video data, and the like. Process 400 may continue from block 410 to block 420 of establishing a learning interval within the image source.

In block 420 , the computing device may analyze the image source obtained in block 410 to set a learning section in the image source. The computing device may perform various analyzes in order to derive a learnable section within the image source. The computing device may, for example, perform at least one of measuring a color change rate of a source of an image, measuring an image complexity, and measuring a structural similarity of an object in the image source. In some examples, the computing device may measure a color change rate in the image source, and set a section in which the measured color change rate is within a predetermined range. In some other examples, the computing device may measure image complexity in the image source, and may set a section in which the complexity is within a predetermined range as the learning section. In some other examples, the computing device may measure structural similarity within the image source. There may be predefined obstacles in the image source, and the computing device may set the learning interval so that these obstacles are not included. Process 400 may continue at block 420 to block 430 where the label image is extracted.

In block 430 , the computing device may extract a label image from the learning section set in block 420 . The computing device may determine a label image from the image source within a learning interval set for the image source. The label image is an image including a main object included in an image source, and may refer to a representative image. In some examples, the computing device may extract the label image in any manner. In some other examples, the computing device may extract a central image located at the center of the learning section as a label image. In some other examples, the computing device may determine, as a label image, an image satisfying a predetermined condition among a plurality of images extracted within the learning section. Process 400 may continue from block 430 to block 440 where at least one adjacent image is extracted.

In block 440 , the computing device may extract at least one adjacent image based on the learning section set in block 420 and the label image extracted in block 430 . In some examples, the computing device may use the label image as a reference image of the learning section to extract an image separated by a predetermined distance from the label image as an adjacent image. In some other examples, the computing device may extract an image having a structural similarity with the label image of a predetermined value (eg, 0.9) or more as an adjacent image. Process 400 may continue from block 440 to block 450 generating training data.

In block 450 , the computing device generates training data of a machine learning model that performs similar image generation based on the label image extracted in block 430 and the at least one adjacent image extracted in block 440 . can In some examples, the machine learning model may be stored on a computing device. The machine learning model may include various types of artificial intelligence networks including, for example, GANs, Star GANs, and the like. The machine learning model may perform learning using the generated training data. In some examples, the machine learning model, ie, the computing device, may use at least one adjacent image to generate a similar image, and use the label image as a real image used to determine the generated similar image.

5 illustrates a computer program product 500 that may be used to generate training data of a machine learning model for generating similar images, in accordance with at least some embodiments of the disclosure, and in accordance with at least some embodiments of the disclosure. . An exemplary embodiment of an exemplary computer program product is provided using a signal bearing medium 510 . In some embodiments, the signal bearing medium 510 of the one or more computer program products 500 may include a computer readable medium 530 and/or a recordable medium 540 .

The instructions 520 contained in the signal bearing medium 510 may be executed by a computing device, such as the similar image generating device 100 . Accordingly, the computing device may include the instruction 520 for generating training data of a machine learning model that performs machine learning-based image generation. When instructions 520 are executed, one or more instructions for obtaining an image source; one or more instructions for analyzing the image source to set a learning interval in the image source; one or more commands for extracting a label image from the set learning section; one or more instructions for extracting at least one adjacent image based on the set learning section and the label image; or one or more instructions for generating, based on the label image and the at least one adjacent image, training data for input into the machine learning model.

6 is a block diagram of an example embodiment of a computing device 600 arranged in accordance with at least some embodiments of the present disclosure. In one exemplary basic configuration 602 , computing device 600 may include one or more processors 604 and system memory 606 . A memory bus 608 may be used to communicate between the processor 604 and the system memory 606 .

Depending on the configuration required, the processor 604 may be of any type including, but not limited to, a microprocessor (μP), a microcontroller (μC), a digital signal processor (DSP), or any combination thereof. The processor 604 may include one or more levels of cache, such as a level 1 cache 610 , a level 2 cache 612 , a processor core 614 , and registers 616 . The processor core 614 may include an arithmetic logic unit (ALU), a floating point unit (FPU), a digital signal processing core (DSP Core), or any combination thereof. Memory controller 618 may also be used with processor 604 , or in some implementations memory controller 618 may be an internal part of processor 604 .

Depending on the configuration required, system memory 606 can be of any type including, but not limited to, volatile memory (such as RAM), non-volatile memory (such as ROM, flash memory, etc.), or any combination thereof. does not System memory 606 may include an operating system 620 , one or more applications 622 , and program data 624 . Application 622 is arranged to perform functions as described herein, including functional blocks and/or actions described with respect to process 400 of FIG. 4 and/or apparatus 100 for generating similar images of FIG. 1 . algorithm 626 . Program data 624 may include data 628 for use with algorithm 626 , eg, data corresponding to a static network environment, or the like. In some embodiments, application 622 may be arranged to operate with program data 624 on operating system 620 such that an implementation for determining an optimal transfer environment may be provided as described herein. For example, the similar image generating device 100 may include all or part of the computing device 600 and may include all or part of the application 622 , such that an implementation determining an optimal transmission environment may be provided as described herein. It may be possible to do some. This described basic configuration is illustrated by its components within dashed line 602 in FIG. 6 .

Computing device 600 may have additional features or functionality, and additional interfaces, to facilitate communication between basic configuration 602 and any desired devices and interfaces. For example, bus/interface controller 630 may be used to facilitate communication between basic configuration 602 and one or more data storage devices 632 over storage interface bus 634 . The data storage device 632 may be a removable storage device 636 , a fixed storage device 638 , or a combination thereof. Magnetic disk devices such as flexible disk drives and hard disk drives (HDDs), compact disk (CD) drives, or digital versatile disk (DVD) drives, to name a few examples of removable and fixed storage devices. such as optical disk drives, solid state drives (SSDs), and tape drives. Exemplary computer storage media includes volatile and nonvolatile removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, or other data. can do.

System memory 606 , removable storage 636 , and non-removable storage 638 are all examples of computer storage media. Computer storage media may include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage device, magnetic cassette, magnetic tape, magnetic disk storage device or other magnetic storage device, or including, but not limited to, any other medium that may be used to store the required information and that may be accessed by the computing device 700 . Any such computer storage media may be part of computing device 600 .

Computing device 600 also provides an interface bus 642 for facilitating communication from various interface devices (eg, output interfaces, peripheral interfaces, and communication interfaces) to basic configuration 602 via bus/interface controller 642 . ) may be included. Exemplary output device 642 may include a graphics processing unit 648 and an audio processing unit 650 , which may be configured to communicate via one or more A/V ports 652 to various external devices, such as displays or speakers. can be configured. Also, although graphics processing unit 648 is shown as a component of an output device in FIG. 6 , in addition to processor 604 , it is possible to perform operations according to some implementations. According to some embodiments according to the present disclosure, the graphic processing unit 648 may be configured to perform various neural network-related operations, such as, for example, learning of a neural network, processing according to the neural network, and the like. Exemplary peripheral interface 644 may include a serial interface controller 654 or a parallel interface controller 656 , which may include input devices (eg, keyboard, mouse, pen, It may be configured to communicate with an external device, such as a voice input device, a touch input device, etc.) or other peripheral device (eg, a printer, scanner, etc.). The exemplary communication device 646 includes a network controller 660 , which may be arranged to facilitate communication with one or more other computing devices 762 over network communications via one or more communication ports 664 .

Computing device 600 may also be implemented as a personal computer including both laptop computer and non-laptop computer configurations. Additionally, computing device 600 may be implemented as part of a wireless base station or other wireless system or device.

The foregoing description of the present application is for illustration, and those of ordinary skill in the art to which the present application pertains will understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present application. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and likewise components described as distributed may also be implemented in a combined form.

Above, the object to be claimed in the present disclosure has been described in detail. The claimed subject matter in this disclosure is not limited in scope to the specific implementations described above. For example, in some implementations it may be in the form of hardware used operatively on a device or combination of devices, in other implementations it may be implemented in the form of software and/or firmware, and in still other implementations it may be in the form of a signal bearing medium; may include one or more articles, such as storage media. Herein, a storage medium such as a CD-ROM, a computer disk, a flash memory, etc. is an instruction that, when executed by a computing device such as a computing system, a computing platform, or other systems, may cause the processor to execute according to the embodiments described above. can be saved. Such computing devices may include one or more processing units or processors, one or more input/output devices such as a display, keyboard and/or mouse, and one or more memory such as static random access memory, dynamic random access memory, flash memory and/or hard drives. may include

In the foregoing detailed description, various embodiments of apparatus and/or processes have been described by way of block diagrams, flow diagrams, and/or other examples. Such block diagrams, flow diagrams, and/or other examples will include one or more functions and/or operations, and those skilled in the art will recognize that each function and/or operation within the block diagrams, flow diagrams, and/or other examples may be implemented in hardware, software, firmware, Or it will be understood that they may be implemented individually or collectively by any combination thereof. In one embodiment, some portions of the subject matter described in the present disclosure may be implemented through an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), a Digital Signal Processor (DSP), or other form of integration. In contrast, some aspects of embodiments of the present disclosure include one or more computer programs running on one or more computers (eg, one or more programs running on one or more computer systems), one or more programs running on one or more processors ( writing code for software and/or firmware, which may be equally implemented in whole or in part as, for example, one or more programs running on one or more microprocessors), firmware, or substantially any combination thereof; and/or the design of the circuit is within the skill of those skilled in the art in light of the present disclosure. In addition, those skilled in the art will understand that the mechanisms of the subject matter of this disclosure may be distributed in various forms of program products, and the examples of the subject matter of this disclosure apply irrespective of the particular type of signal bearing medium used to actually perform the distribution. will understand

While specific exemplary techniques have been described and illustrated herein using various methods and systems, those skilled in the art will understand the possibility of various other modifications or equivalents permutations without departing from the claimed subject matter. Additionally, many modifications may be made to adapt a particular situation to the teachings of the claimed subject matter without departing from the central concepts described herein. Accordingly, it is intended that claimed subject matter not be limited to the particular examples disclosed, but that such claimed subject matter may also include all embodiments falling within the scope of the appended claims and their equivalents.

The scope of the present disclosure is indicated by the following claims rather than the above detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present application. do.

Claims (11)

A method for generating training data that is performed under the control of a computing device and generates training data of a machine learning model that performs machine learning-based similar image generation, comprising:
acquiring an image source;
analyzing the image source to set a learning section in the image source;
extracting a label image from the set learning section;
extracting at least one adjacent image based on the set learning section and the label image; and
generating training data for input to the machine learning model based on the label image and the at least one adjacent image;
A method of generating training data comprising a. According to claim 1,
Wherein the machine learning model includes a generative adversarial network (Generative Adversarial Network). According to claim 1,
The machine learning model uses the at least one adjacent image to generate a similar image, and uses the label image as a real image used to determine the generated similar image. According to claim 1,
The step of setting the learning section is
a measuring step of performing at least one of measuring a color change rate of the image source, measuring image complexity, and measuring a structural similarity of an object in the image source; and
Including the step of setting a section in the image source so that a result of the measurement step satisfies a predetermined condition, the learning data generating method. According to claim 1,
The step of extracting the label image is
The method of generating learning data, comprising extracting a central image as the label image within the learning section in the image source. According to claim 1,
The step of extracting the at least one neighboring image includes extracting the corresponding neighboring image from the learning section so that a structural similarity between each of the at least one neighboring image and the label image is within a predetermined range. Way. A similar image generating device comprising:
an image source management unit for acquiring an image source;
a learning section setting unit that analyzes the image source and sets a learning section in the image source;
a label image extraction unit for extracting a label image from the set learning section;
an adjacent image extractor configured to extract at least one adjacent image based on the set learning section and the label image; and
A similarity comprising a machine learning model for generating a machine learning-based image, generating a similar image using the machine learning model, and learning the machine learning model using the label image and the at least one adjacent image image generator
A similar image generating device comprising a. 8. The method of claim 7,
The learning section setting unit is configured to perform at least one of measuring a color change rate of the image source, measuring image complexity, and measuring a structural similarity of an object in the image source. 8. The method of claim 7,
and the adjacent image extracting unit extracts the adjacent image from the learning section so that a structural similarity between each of the at least one adjacent image and the label image is within a predetermined range. 8. The method of claim 7,
A metadata extraction unit for extracting metadata from the image source, the label image, and the at least one adjacent image
further comprising,
The similar image generating unit further uses the metadata to train a machine learning model. A computer-readable storage medium storing a training data generation program that generates training data of a machine learning model that performs machine learning-based image generation, wherein the training data generation program is executed by a computing device,
acquiring an image source;
analyzing the image source and setting a learning section in the image source;
extracting a label image from the set learning section;
extracting at least one adjacent image based on the set learning section and the label image; and
generating training data for input to the machine learning model based on the label image and the at least one adjacent image
A computer-readable storage medium comprising one or more instructions to cause

Images