KR20230105814A - Apparatus and method for generating images using generative adversarial network and quality filtering - Google Patents

Abstract

One embodiment of the present invention provides a method for generating images by using an image generating device. The method comprises the steps in which: the image generating device generates a camouflage image based on an object image by using a first artificial intelligence model trained to generate the camouflage image corresponding to an input image; the image generating device determines whether the camouflage image represents a preset object included in a training data set, by using a second artificial intelligence model trained to determine whether the input image represents an object included in the training data set including the object image; and when it is determined that the camouflage image represents the preset object included in the training data set, the camouflage image is added to the training data set to create a new training data set, and based on the new training data set, the second artificial intelligence model is trained. Therefore, the method can reduce the time, cost, and labor required to filter out low-quality fake images.

Description

Apparatus and method for generating images using generative adversarial networks and image filtering

The present invention relates to an apparatus and method for generating images using a generative adversarial neural network and image filtering, and more particularly, to generating high-quality simulated images by performing image filtering on simulated images generated through a generative adversarial neural network. It relates to an apparatus and method for doing.

In order to improve performance by training a deep learning network, the data collection procedure used for training is important. Although a sufficient number of images are required in the data collection procedure, it is difficult to secure a sufficient amount of data set depending on the situation. In addition, inefficient time and cost may be consumed in processing the collected data due to various reasons such as a lack of experts in the field in the data pre-processing process. This soon leads to problems of insufficient data set and class imbalance, and eventually overfitting may occur, resulting in poor performance of deep learning networks.

Various methods are being used as a way to solve the problem of insufficient data set and class imbalance. Conventional data enlargement methods have limitations in increasing the number of images because most of them use a method of repeating only geometric transformation functions such as rotation, flipping, and color change as image enlargement. Accordingly, a method of supplementing a data set by artificially generating an imitation image for an original image using a Generative Adversarial Network (GAN) has been proposed.

However, a data set including fake images generated through GAN includes many low-quality images, and performance may deteriorate when a model is trained with such a data set. Thus, there was a need to filter low-quality fake images. In the conventional case, a person manually performs image filtering on each of the fake images, but in this case, objectivity for quality standards may be a problem, and it is very inefficient in terms of time and labor requirements.

An object of the present invention is to provide an apparatus and method for generating high quality fake images by performing image filtering on fake images generated through a generative adversarial neural network.

The technical problems to be achieved by the present invention are not limited to the above technical problems, and other technical problems of the present invention can be derived from the following description.

As a technical means for solving the above technical problem, an embodiment according to an aspect of the present invention provides a method of generating an image using an image generating device. The method includes generating, by the image generating device, a camouflage image based on an image of an object by using a first artificial intelligence model learned to generate a camouflaged image corresponding to an input image; , The camouflage image is a preset object included in the training data set, by using a second artificial intelligence model trained to determine whether the input image represents an object included in the training data set including the object image. determining whether the camouflage image represents a preset object included in the training data set, adding the camouflage image to the training data set to create a new training data set, and the new learning data set and training the second artificial intelligence model based on the data set.

In addition, an embodiment according to another aspect of the present invention provides an apparatus for generating an image using an artificial intelligence model. The apparatus includes a memory for storing data used for learning and an image generating program, and a processor for executing the program stored in the memory. The processor executes the image generation program to generate a camouflage image based on an object image using a first artificial intelligence model learned to generate a camouflage image corresponding to an input image, and the input image is the object It is determined whether the stomach image represents a preset object included in the training data set using a second artificial intelligence model trained to determine whether the camouflage image represents an object included in a training data set including an image, and the When it is determined that the camouflage image represents a preset object included in the training data set, the camouflage image is added to the training data set to generate a new training data set, and the second artificial intelligence model based on the new training data set. It is configured to perform learning.

According to the present invention, it is possible to artificially generate fake images for an original image using a generative adversarial neural network, and automatically perform image filtering on the fake images to filter low-quality fake images. and reduce labor.

In addition, according to the present invention, it is possible to solve problems of insufficient data set and class imbalance that may occur in model learning using deep learning by generating a training data set including high-quality imitation images with a small number of original images.

1 is a block diagram showing the configuration of an image generating apparatus using a generative adversarial neural network and image filtering according to an embodiment of the present invention.
2 to 5 are diagrams for explaining processes of generating an image using the image generating apparatus shown in FIG. 1 .
6 is a flowchart illustrating a sequence of an image generation method using a generative adversarial neural network and image filtering according to another embodiment of the present invention.
7 and 8 are diagrams illustrating detailed steps for some steps of the image generating method shown in FIG. 6 .
9 is a diagram for explaining in detail an image filtering process applied to an embodiment of the present invention.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings. However, the present invention may be implemented in many different forms, and is not limited to the embodiments described herein. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, and the technical ideas disclosed in this specification are not limited by the accompanying drawings. All terms including technical terms and scientific terms used herein should be interpreted as meanings commonly understood by those of ordinary skill in the art to which the present invention belongs. The terms defined in the dictionary should be interpreted as having additional meanings corresponding to the related technical literature and the currently disclosed content, and are not interpreted in a very ideal or limiting sense unless otherwise defined.

In order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and the size, shape, and shape of each component shown in the drawings may be variously modified. Same/similar reference numerals are assigned to the same/similar parts throughout the specification.

The suffixes "module" and "unit" for the components used in the following description are given or used interchangeably in consideration of ease of writing the specification, and do not have meanings or roles that are distinguished from each other by themselves. In addition, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related known technologies may obscure the gist of the embodiments disclosed in this specification, the detailed description is omitted.

Throughout the specification, when a part is said to be “connected (connected, contacted, or combined)” with another part, this is not only the case where it is “directly connected (connected, contacted, or coupled)”, but also has other members in the middle. It also includes the case of being "indirectly connected (connected, contacted, or coupled)" between them. In addition, when a part "includes (provides or provides)" a certain component, it does not exclude other components, but "includes (provides or provides)" other components unless otherwise specified. means you can

Terms indicating ordinal numbers such as first and second used in this specification are used only for the purpose of distinguishing one element from another, and do not limit the order or relationship of elements. For example, a first element of the present invention may be termed a second element, and similarly, the second element may also be termed a first element. Singular forms used herein should be construed to include plural forms as well, unless the meaning is clearly indicated to the contrary.

1 is a block diagram showing the configuration of an image generating device (hereinafter referred to as “image generating device 100”) using a generative adversarial neural network and image filtering according to an embodiment of the present invention, and FIGS. 5 is diagrams for explaining processes of generating an image using the image generating device 100 . Hereinafter, with reference to FIGS. 1 to 5 , the image generating device 100 will be described in detail.

Referring to FIG. 1 , an image generating device 100 is a device for generating an image using an artificial intelligence model, and includes a memory 120 and a processor 140 . Also, the image generating device 100 may further include a communication module 110 and a database 130 . The image generating device 100 may be formed in the form of a device such as a server or a terminal, and in a cloud computing service model such as Software as a Service (SaaS), Platform as a Service (PaaS), or Infrastructure as a Service (IaaS). can work The image generating device 100 may be built in the form of a server such as a private cloud, public cloud, or hybrid cloud system. In addition, the image generating device 100 may include, for example, a laptop, a desktop, a laptop equipped with a web browser, a wireless communication device that ensures portability and mobility, a smartphone, a tablet PC, and the like. It can be implemented in all types of handheld-based wireless communication devices.

The communication module 110 may transmit/receive information with an external device through a communication network. The communication module 110 may receive or transmit image data, a deep learning algorithm, and a learning model generated based on the deep learning algorithm used for learning of the image generating device 100 from the outside. Here, the external device may refer to a device such as a server or a terminal separate from the image generating device 100 . The communication module 110 is a wired network such as a local area network (LAN), a wide area network (WAN) or a value added network (VAN), a mobile radio communication network, or a satellite It can be connected to external devices through all kinds of wireless networks such as communication networks. The communication module 110 may include a device including hardware and software necessary for transmitting and receiving a signal such as a control signal or a data signal to and from other network devices through a wired or wireless connection.

The memory 120 stores data used for learning and an image generating program. The name of the program is set for convenience of explanation, and the name itself does not limit the function of the program. In addition, the memory 120 may store at least one of information and data input to the communication module 110, information and data required for functions performed by the processor 140, and data generated according to execution of the processor 140. can be saved. For example, the memory 120 may store images and data used for learning of the image generating device 100 and artificial intelligence learning models generated by the processor 140 . The memory 120 should be interpreted as collectively referring to a non-volatile storage device that continuously maintains stored information even when power is not supplied and a volatile storage device that requires power to maintain stored information. Also, the memory 120 may temporarily or permanently store data processed by the processor 140 . The memory 120 may include magnetic storage media or flash storage media in addition to volatile storage devices that require power to maintain stored information, but the scope of the present invention is not limited thereto. no.

The database 130 may include information and data generated by the image generating device 100 . The database 130 may be built by the processor 140 and may be configured to include camouflage images according to image filtering described below. The database 130 may form part of the memory 120 or may be formed separately outside the memory 120 .

Processor 140 is configured to execute programs stored in memory 120 . The processor 140 may include various types of devices that control and process data. The processor 140 may refer to a data processing device embedded in hardware having a physically structured circuit to perform functions expressed by codes or instructions included in a program. In one example, the processor 140 includes a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, an application-specific integrated circuit (ASIC), an FPGA ( field programmable gate array), etc., but the scope of the present invention is not limited thereto. Processor 140 is configured to execute the image generating program to perform the following functions and procedures.

The processor 140 generates a camouflage image based on the image of the object by using a first artificial intelligence model learned to generate a camouflage image corresponding to an input image. The processor 140 determines whether the spoof image represents a predetermined object included in the training data set by using the second artificial intelligence model trained to determine whether the input image represents an object included in the training data set. When the processor 140 determines that the camouflage image represents a predetermined object included in the training data set, the processor 140 updates the training data set by adding the camouflage image to the training data set. That is, processor 140 creates a new training data set. Then, the processor 140 trains the second artificial intelligence model based on the new training data set. A camouflage image included in the training data set and not determined to represent a preset object may be discarded or deleted. The process of updating the training data set may be performed a plurality of times.

The object image may include a real object image and a fake object image, and the spoof image may include a real spoof image and a fake spoof image. Images described in this specification, including an object image and a stomach image, may each include a plurality of images. The real object image and real camouflage image refer to images acquired by the processor 140 from a photographing means such as a camera. The fake object image and the fake camouflage image refer to images acquired through an artificial intelligence model in which the processor 140 receives a real image and generates a fake image. In one example, the object image may include a plurality of face images, and the camouflage image may include images in which at least one of a hat, a mask, glasses, sunglasses, and a scarf is disguised for each face in the plurality of face images. In another example, the object image may include a plurality of images belonging to the healthy leaf category, and the camouflage image may include images masquerading as disease on each leaf in the plurality of images belonging to the healthy leaf category.

The processor 140 may execute an image generation program and further perform a process of generating a first artificial intelligence model using a Cycle Generative Adversarial Network (CycleGAN). At this time, the processor 140 generates a fake camouflage image based on the real object image, creates a fake object image based on the real object image, compares the difference between the real object image and the fake object image, and compares the fake camouflage image with the fake camouflage image. A process of training the first artificial intelligence model to minimize periodic consistency loss calculated by comparing differences between real camouflage images may be further performed.

The processor 140 may assign a preset label to an object image included in the training data set. For example, different labels may be set for each of the object images. The processor 140 determines whether the camouflage image represents the same object as the object image to which a preset label has been assigned, and learns the camouflage image only when it is determined that the camouflage image represents the same object as the target image to which a preset label has been assigned. A process of generating a new training data set by adding to the data set and training a second artificial intelligence model based on the new training data set may be further performed. If it is not determined that the camouflage image represents the same object as the object image to which a predetermined label is assigned, the processor 140 may discard or delete the corresponding image. Additionally, the processor 140 may further perform a process of constructing the camouflage image database 130 including camouflage images added to the training data set.

Referring to FIGS. 2 and 3 , the image generating apparatus 100 may generate an image based on an artificial intelligence model based on a circular generative adversarial network including a first model and a second model. Each of the first model and the second model includes a generator and discriminator. The generator of the first model learns a transition from class A to class B and the generator of the second model learns a transition from class B to class A. In this way, the artificial intelligence model can generate a natural fake image by learning to minimize periodic consistency loss through bidirectional transition.

For example, referring to FIG. 2 , the first model 210 inputs a first input image 201 including an undisguised real face image and an undisguised fake face image generated by the second model 220 . and generate a disguised face image as a first output image. The second model 220 receives the second input image 202 including the disguised real face image and the disguised fake face image generated by the first model 210 as input, and serves as the second output image, a face that is not disguised. create an image In this process, an artificial intelligence model can be trained in the direction of reducing the periodic consistency loss calculated by comparing the real face image that is not disguised with the imitation face image that is not disguised, and the real face image that is disguised and the fake face image that is disguised. there is.

Referring to FIG. 3 , the first model 310 receives a first input image 301 including an undisguised real leaf image and an undisguised imitation leaf image generated by the second model 320 and produces 1 Generate a camouflaged leaf image as an output image. The second model 320 receives the second input image 302 including the camouflaged real leaf image and the camouflaged imitation leaf image generated by the first model 310, and receives the second output image as an undisguised leaf. create an image Here, camouflage means that a region corresponding to a leaf in a healthy leaf image is disguised as a diseased area.

4 is a diagram illustrating the operation process of the first artificial intelligence model 410 and the second artificial intelligence model 420 described above with reference to FIG. 1, and FIG. 5 is a label set in the above-described image and It is a drawing shown to explain camouflage. Since the operation process of the first artificial intelligence model 410 is substantially the same as the artificial intelligence model described above with reference to FIGS. 2 and 3, a description thereof will be omitted.

Referring to FIG. 4 , real images including real images that are not disguised and real images of an object and disguised images of the object generated by the first AI model 410 are generated by a second AI model 410 . Used in model 420. For example, the real images may be used as training data of the second artificial intelligence model 420, and the disguised images may be used as test data of the second artificial intelligence model 420. The second artificial intelligence model 420 performs a process of determining whether an object included in the disguised images represents an object included in the real images. This process is performed using labels assigned to each of the actual images.

Referring to FIG. 5 related to labels, the same label is assigned to a real face image of an object and a simulated face image of the object generated by an artificial intelligence model. A label such as Subject 1 may be assigned to the first object, and Subject 2 may be assigned to the second object. after. The image generating device 100 performs a camouflage process 520 including hat camouflage, glasses camouflage, mask camouflage, scarf camouflage, sunglasses and mask camouflage, etc., for the labeled image, thereby generating camouflage images 530. can be obtained. The second artificial intelligence model 420 recognizes disguised objects in the camouflage images 520 and performs label comparison. The image generating apparatus 100 may include spoofed images whose labels have been recognized through label comparison in an existing training data set, and discard spoofed unknowns whose labels are not. This test process may be performed multiple times. Also, the image generating device 100 may build a database storing only camouflage images for which labels have been recognized.

An example of training the face recognition model, which is the second artificial intelligence model, by the image generating device described above is as follows. The image generating device 100 performs learning with a normal image and a fake image of a real face using a SqueezeNet-based face recognition model. The face recognition model has never seen the image generated by the generative adversarial network, but the distribution between the camouflage image generated by the generative adversarial network and the actual camouflage image is similar due to periodic consistency loss, so the trained face recognition model Both images can be correctly recognized. thus. The face model can be used to discriminate between good and bad creations of synthesized face images. The correctly predicted images in the first iteration are saved and included in the training data in the second iteration. In the second iteration, the face recognition model is trained with the real image and the generated camouflage image. In this way, the trained face recognition model is tested against misrecognized images in the first iteration. Accordingly, correctly recognized images are saved and the rest are discarded.

6 is a flow chart showing the sequence of an image generation method (hereinafter referred to as “image generation method”) using a generative adversarial neural network and image filtering according to another embodiment of the present invention, and FIGS. 7 and 8 are image generation methods. These are drawings showing detailed steps for some steps of. An image generating method to be described below may be performed by the image generating device ( 100 of FIG. 1 ) previously described with reference to FIGS. 1 to 5 . Therefore, the contents of the embodiments of the present invention described above with reference to FIGS. 1 to 5 can be equally applied to the embodiments to be described below, and descriptions overlapping with those described above will be omitted. The steps described below do not necessarily have to be performed in order, the order of the steps may be set in various ways, and the steps may be performed almost simultaneously.

Referring to FIG. 6 , the image generation method is a method of generating an image using an image generation device, and includes a camouflage image generation step (S110), a camouflage image test step (S120), and a training data set update step (S130). do.

The camouflage image generation step ( S110 ) is a step in which the image generating device generates a camouflage image corresponding to the object image by using the first artificial intelligence model learned to generate a camouflage image corresponding to the input image.

In the camouflage image test step (S120), the camouflage image is learned by using the second artificial intelligence model trained to determine whether the input image represents an object included in the training data set including the object image, by the image generating device. This is a step of determining whether the data set represents a predetermined object included in the data set. In the training data set updating step (S130), when the image generating device determines that the spoof image represents a preset object included in the training data set, a new training data set is generated by adding the spoof image to the training data set, and a new training data set is generated. This step is to train the second artificial intelligence model based on the training data set. Here, the object image may include a real object image and a fake object image, and the camouflage image may include a real camouflage image and a fake camouflage image.

The camouflage image generating step ( S110 ) may include generating, by the image generating device, a first artificial intelligence model using a Cycle Generative Adversarial Network (CycleGAN). In addition, referring to FIG. 7 , generating a camouflage image (S110) includes the steps of generating, by the image generating device, a fake camouflage image based on a real object image and generating a fake object image based on the real camouflage image (S111) and the actual camouflage image. The method may further include comparing the difference between the object image and the fake object image and comparing the difference between the fake spoof image and the actual spoof image to minimize periodic consistency loss calculated by training the first artificial intelligence model (S112). .

In one example, the above-described object image may include a plurality of face images. The camouflage image may include images in which at least one of a hat, a mask, glasses, sunglasses, and a scarf is camouflaged for each face in a plurality of face images.

In another example, the above-described object image may include a plurality of images belonging to a healthy leaf category. The camouflage image may include images camouflaging disease on each leaf among a plurality of images belonging to the healthy leaf category.

Referring to FIG. 8 , a preset label may be assigned to an object image included in a training data set. At this time, in the updating of the training data set (S130), the image generating device determines whether the camouflage image represents the same object as the object image to which a preset label has been assigned (S131), and the image generating device determines whether the camouflage image is a basic generating a new training data set by adding a camouflage image to a training data set and training a second artificial intelligence model based on the new training data set only when it is determined that the set label represents the same object as the assigned object image; (S132) may be included.

Although not shown in the figure, the image generating method may further include, by the image generating device, constructing a camouflage image database 130 including camouflage images added to the learning data set in step S130.

The image generation method described above includes the steps of generating a camouflage image corresponding to an object image by using a first artificial intelligence model learned to generate a camouflage image corresponding to an input image by an image generation device; The device uses a second artificial intelligence model trained to determine whether an input image represents an object included in a training data set including the object image, and the camouflage image is included in the training data set. determining whether the input image represents an object; learning, by the image generating device, a second artificial intelligence model for determining whether an input image represents an object included in a training data set including the object image; 2 Determining whether the camouflage image represents a preset object included in the learning data set using an artificial intelligence model; Adding a camouflage image to the training data set to create a new training data set, and holding the fake camouflage image when it is determined that it does not represent a preset object; and, based on the new training data set, the second training data set is created. It may also be implemented by learning an artificial intelligence model to determine whether the withheld simulated gastrointestinal image represents a predetermined object, and discarding images that cannot be represented.

9 is a diagram for explaining in detail an image filtering process applied to an embodiment of the present invention. Referring to the descriptions of the embodiments of the present invention described above together with FIG. 9 , according to an embodiment of the present invention, the image generating apparatus 100 first trains a face recognition model on a training data set of real face photos. . The image generating apparatus 100 trains a cognitive model through an existing training data set, and then tests a simulated fake image generated through a recursive adversarial neural network. The image generating apparatus 100 forms a new training data set by including only images correctly recognized according to the test to the existing training data set, and retains data that are not. The image generating device 100 trains a cognitive model for a new data set. The image generating apparatus 100 tests the retained images again with the trained model, and includes correctly recognized images in the data set, and discards images that are not.

The method of generating an image according to an embodiment of the present invention described above may be implemented in the form of a recording medium including instructions executable by a computer, such as program modules executed by a computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer readable media may include computer storage media. Computer storage media includes both 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.

Those skilled in the art to which the present invention pertains will be able to understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention based on the above description. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. The scope of the present invention is indicated by the following claims, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts should be interpreted as being included in the scope of the present invention. The scope of the present application is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and equivalent concepts thereof should be construed as being included in the scope of the present application.

Claims (15)

In the method of generating an image using an image generating device,
a) generating, by the image generating device, a camouflage image corresponding to an image of an object by using a first artificial intelligence model learned to generate a camouflage image corresponding to an input image;
b) The image generating device uses a second artificial intelligence model trained to determine whether an input image represents an object included in a training data set including the object image, so that the camouflage image is selected from the training data set. determining whether it represents a predetermined object included in; and
c) When the image generating device determines that the spoofed image represents a preset object included in the training data set, a new training data set is created by adding the spoofed image to the training data set, and the new training data set is added to the training data set. Training the second artificial intelligence model based on a data set. According to claim 1,
In step a),
The image generating method comprising generating, by the image generating device, the first artificial intelligence model using a Cycle Generative Adversarial Network (CycleGAN). According to claim 2,
The object image includes a real object image and an imitation object image,
The camouflage image includes a real camouflage image and an imitation camouflage image, and
In step a),
a-1) generating, by the image generating device, the fake camouflage image based on the real object image, and generating the fake object image based on the real camouflage image; and
a-2) comparing the difference between the real object image and the simulated object image, and training the first artificial intelligence model to minimize periodic consistency loss calculated by comparing the difference between the simulated gastrointestinal image and the real gastrointestinal image A method of generating an image, comprising steps. According to claim 1,
The object image includes a plurality of face images, and
The camouflage image includes images in which at least one or more of a hat, a mask, glasses, sunglasses, and a scarf are camouflaged for each face in the plurality of face images. According to claim 1,
The object image includes a plurality of images belonging to the health leaf category, and
Wherein the camouflage image includes images camouflaging disease on each leaf in a plurality of images belonging to the healthy leaf category. According to claim 1,
A preset label is assigned to an object image included in the learning data set,
In step c),
c-1) determining, by the image generating device, whether the camouflage image represents the same object as the object image to which the preset label is assigned; and
c-2) When the image generating device determines that the camouflage image represents the same object as the object image to which the preset label is assigned, a new training data set is created by adding the camouflage image to the training data set. Generating and learning the second artificial intelligence model based on the new training data set. According to claim 1,
d) constructing, by the image generating device, a camouflage image database including camouflage images added to the training data set according to step c); A device for generating an image using an artificial intelligence model,
a memory for storing data used for learning and an image generating program; and
A processor for executing a program stored in the memory;
The processor executes the image generating program,
To generate a camouflage image based on an object image using a first artificial intelligence model learned to generate a camouflage image corresponding to an input image, and to determine whether the input image represents an object included in the training data set Determining whether the camouflage image represents a preset object included in the training data set using the learned second artificial intelligence model, and determining that the camouflage image represents a preset object included in the training data set if the camouflage image is added to the training data set to generate a new training data set, and to perform training of the second artificial intelligence model based on the new training data set. According to claim 8,
The processor executes the image generating program,
The image generating device further performing a process of generating the first artificial intelligence model using a Cycle Generative Adversarial Network (CycleGAN). According to claim 9,
The object image includes a real object image and an imitation object image,
The camouflage image includes a real camouflage image and an imitation camouflage image, and
The processor executes the image generating program,
The fake camouflage image is generated based on the real object image, the fake object image is created based on the real object image, and a difference between the real object image and the fake object image is compared and the fake camouflage image and the real object image are compared. Further performing a process of learning the first artificial intelligence model to minimize periodic consistency loss calculated by comparing differences between camouflage images. According to claim 8,
The object image includes a plurality of face images, and
The camouflage image includes images in which at least one of a hat, a mask, glasses, sunglasses, and a scarf is camouflaged for each face in the plurality of face images. According to claim 8,
The object image includes a plurality of images belonging to the health leaf category, and
The camouflage image includes images masquerading as disease on each leaf in a plurality of images belonging to the healthy leaf category. According to claim 8,
A preset label is assigned to an object image included in the learning data set,
The processor executes the image generating program,
It is determined whether the gastrointestinal image represents the same object as the object image to which the preset label is assigned, and the gastrointestinal image is displayed only when it is determined that the gastrointestinal image represents the same object as the object image to which the preset label is assigned. An image generating device further performing a process of generating a new training data set in addition to a training data set and learning the second artificial intelligence model based on the new training data set. According to claim 8,
Wherein the processor executes the image generating program to further perform a process of constructing a camouflage image database including camouflage images added to the training data set. A non-transitory computer-readable recording medium on which a computer program for performing the image generating method according to claim 1 is recorded.

Images