KR102491057B1 - Device and Method for Image Style Transfer - Google Patents

Abstract

An image conversion device and method are disclosed.
According to one aspect of the present invention, a neural network trained to extract a plurality of feature maps for an image for each of a plurality of convolution layers; A style loss function is calculated based on a plurality of feature maps for style images extracted from the plurality of convolution layers and a plurality of feature maps for an input image, and a plurality of feature maps for the content image extracted from one convolution layer. a calculator for calculating a content loss function based on a feature map of and a plurality of feature maps of the input image; and a conversion unit that converts the style and content of the input image using the style loss function and the content loss function.

Description

Image conversion device and method {Device and Method for Image Style Transfer}

Embodiments of the present invention relate to an image conversion apparatus and method for applying the style of a style image and the content of a content image to an input image.

The information described in this section simply provides background information on the present invention and does not constitute prior art.

An image style transfer is an image conversion device that generates an image in which the style of a style image and the content of a content image are synthesized. An image conversion device using image style transfer mainly uses a convolutional neural network (CNN).

An image conversion device using image style transfer is divided into two methods. The first is a model optimization method, and the second is an image optimization method.

The model optimization method is a method of generating an output image in which the style of a style image and the content of a content image are mixed using a convolutional neural network (CNN) or a generative adversarial network (GAN). The neural network used in the model optimization method learns numerous images and updates parameters.

In the case of using a generative adversarial network, the generative neural network generates a composite image of a style image and a content image, and the adversarial network is trained to separate the composite image again. The image conversion device may synthesize two arbitrary images using a neural network trained on various image data sets.

In addition to this, there is an image conversion method using a learnable style transfer network. In this method, after adding a learnable style transfer network to a neural network trained to extract features from images, a process of updating the parameters of the style transfer network is performed. Then, a style image and a content image may be synthesized using a pre-trained neural network and a network that has undergone a learning process.

Meanwhile, the image optimization method is a method of generating an image in which a style image and a content image are synthesized by applying a style image of a style image and a style image of a content image to an input image using a pretrained neural network. At this time, the previously trained neural network is used to extract a feature map from the image using fixed parameters.

However, the image optimization method has the advantage of not requiring a separate training process because it uses a neural network that has been trained in advance. There is a problem that it takes a lot of time to generate an image. In addition, unlike the model optimization method, which computes a loss function for a large number of data, trains parameters, and synthesizes two random images, the image optimization method has to perform a calculation process for each input data, so the image is not stable. has a downside.

Therefore, in the image optimization method, research is needed to reduce the amount of computation to reduce the time required to generate a composite image, and to properly combine a style image and a content image.

Embodiments of the present invention add a third loss function in the process of synthesizing a style image and a content image, and apply the style image and content image to an input image using this, thereby converting the style and content of the input image. The main object is to reduce the time required for processing and to provide an image conversion device and method with high reliability.

Other embodiments of the present invention generate a mixed image by mixing an input image and a style image, and reflect the style of the style image and the content of the content image to the pre-generated mixed image, thereby converting the style and content of the input image. One object of the present invention is to reduce the time required for processing and to provide an image conversion device and method with high stability.

According to one aspect of the present invention, in the image conversion device for applying the style of a style image and the content of a content image to an input image, including a plurality of convolution layers, a neural network trained to extract a plurality of feature maps for an image for each convolutional layer; A style loss function is calculated based on a plurality of feature maps for style images extracted from the plurality of convolution layers and a plurality of feature maps for an input image, and the content extracted from one convolution layer a calculator for calculating a content loss function based on a plurality of feature maps of an image and a plurality of feature maps of the input image; and a conversion unit for converting the style and content of the input image using the style loss function and the content loss function (the converted input image is used as an input image for the trained neural network). do.

According to another aspect of this embodiment, in the operating method of an image conversion device for applying the style of a style image and the content of a content image to an input image, including a plurality of convolution layers, each convolution layer includes a plurality of images for the image. inputting a style image to a neural network trained to extract a feature map of and acquiring a plurality of feature maps for the style image; inputting the input image to the trained neural network and acquiring a plurality of feature maps of the input image; inputting the content image to the trained neural network and acquiring a plurality of feature maps of the content image; calculating a style loss function based on a plurality of feature maps for the style image and a plurality of feature maps for the input image among a plurality of feature maps extracted from the plurality of convolution layers; calculating a content loss function based on a plurality of feature maps for the content image and a plurality of feature maps for the input image among a plurality of feature maps extracted from one convolution layer; converting the style and content of the input image using the style loss function and the content loss function; and inputting the converted input image to the trained neural network.

As described above, according to an embodiment of the present invention, the style of the style image is applied to the style of the input image, and the content of the content image is applied to the content of the input image, thereby converting the style and content of the input image. It is possible to reduce the conversion time and increase the conversion stability.

1 is a configuration diagram showing components of an image conversion device according to an embodiment of the present invention.
2 is a diagram for explaining an operation process of an image conversion device according to an embodiment of the present invention.
3 is a diagram for explaining an operation process of an image conversion device using an additive loss function and mixed images according to an embodiment of the present invention.
4 is a flowchart illustrating an operating method of an image conversion device according to an embodiment of the present invention.

Hereinafter, some embodiments of the present invention will be described in detail through exemplary drawings. In adding reference numerals to components of each drawing, it should be noted that the same components have the same numerals as much as possible even if they are displayed on different drawings. In addition, in describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description will be omitted.

Also, terms such as first, second, A, B, (a), and (b) may be used in describing the components of the present invention. These terms are only used to distinguish the component from other components, and the nature, order, or order of the corresponding component is not limited by the term. Throughout the specification, when a part 'includes' or 'includes' a certain component, it means that it may further include other components without excluding other components unless otherwise stated. . In addition, terms such as '~unit' and 'module' described in the specification refer to a unit that processes at least one function or operation, and may be implemented by hardware, software, or a combination of hardware and software.

1 is a configuration diagram showing components of an image conversion device according to an embodiment of the present invention.

Referring to FIG. 1 , an image conversion device (not shown) includes a pre-trained neural network 100 , a calculation unit 110 and a conversion unit 120 . First, the image conversion device acquires a style image and a content image. In FIG. 1 , the input image may be a white noise image or the same image as the content image.

The trained neural network 100 includes a plurality of convolution layers, and is a pre-trained component to extract a plurality of feature maps of an image for each convolution layer. In FIG. 1, the trained neural network 100 generates a plurality of feature maps (F(s)) for the style image after receiving a style image, and after receiving a content image, a plurality of feature maps (F(s)) for the content image F(c)) is generated, and after receiving an input image, a plurality of feature maps F(i) for the input image are generated.

The operation unit 110 determines at least one of a style loss function, a content loss function, and relative entropy to be applied to the input image using the feature map extracted from the trained neural network 100. It is a component that computes one.

The calculator 110 calculates a style loss function based on a plurality of style feature maps extracted from a plurality of convolution layers for a style image and a plurality of input feature maps extracted from a plurality of convolution layers for an input image. In addition, the calculator 110 calculates a content loss function based on a plurality of content feature maps extracted from one convolution layer for the content image and a plurality of content feature maps extracted from one convolution layer for the input image. do. Here, according to an embodiment of the present invention, one convolution layer may mean any one of an arbitrary one of a plurality of convolution layers, a last output layer of a neural network, or a layer immediately before it.

The conversion unit 120 is a component that converts the style and content of an input image using a style loss function and a content loss function. Specifically, the conversion unit 120 converts the style of the input image to be similar to the style of the style image by adjusting the pixel value of the input image using a loss function, and transforms the content of the input image into the content image of the content image. Transform similarly to content. The converted image includes the style of the style image and the content of the content image.

However, since an input image converted once through the above-described process cannot be regarded as similar to a style image and a content image, the above-described process must be repeated. Accordingly, the transformed input image is used as an input image of the trained neural network 100 again. The trained neural network 100 extracts a plurality of feature maps for the converted input image, and the calculator 110 calculates a style loss function and a content loss function based on the plurality of feature maps for the converted input image, The conversion unit 120 applies a loss function to the converted input image to generate an input image converted twice.

The image conversion device repeats the above process until a condition set by at least one of a preset number of times, a preset time, and a preset loss function value is satisfied, and finally determines the converted image as the final output image. As a result, the output image is an image in which the style of the style image and the content of the content image are appropriately synthesized.

2 is a diagram for explaining an operation process of an image conversion device according to an embodiment of the present invention.

)와 컨텐트 손실 함수(

)를 연산하고, 입력 이미지에 적용하는 과정이 도시되어 있다.Referring to Figure 2, the image conversion device is a style loss function (

) and the content loss function (

) is calculated and applied to the input image.

The image conversion device provides a style image to the trained neural network 100 and obtains a plurality of feature maps extracted from each layer. One convolution layer includes a plurality of filters, and a feature map for input data is extracted using the plurality of filters. Among the plurality of convolution layers, the closer to the input layer extracts the main content of the image (shape, outline, etc.), and the closer to the output layer, the more detailed style (color, sharpness, etc.) of the image is extracted.

는 첫 번째 컨볼루션 레이어가 스타일 이미지에 대해 추출한 i개의 특징맵을 의미하고,

는 두 번째 컨볼루션 레이어가 첫 번째 컨볼루션 레이어의 특징맵에 대해 추출한 i개의 특징맵을 의미한다. 여기서, 각각의 컨볼루션 레이어는 i개의 특징맵을 추출한다고 설명했지만, 이는 하나의 실시예에 불과하며 각 컨볼루션 레이어는 같은 수의 특징맵을 생성할 수도 있고, 서로 다른 수의 특징맵을 생성할 수도 있다. 또한, 이미지 변환 장치는 훈련된 신경망(100)에 입력 이미지를 제공하고, 각각의 레이어로부터 추출된 복수의 특징맵을 획득한다. in Figure 2

Means i feature maps extracted for the style image by the first convolution layer,

denotes i feature maps extracted from the feature maps of the first convolution layer by the second convolution layer. Here, it has been described that each convolution layer extracts i feature maps, but this is only one embodiment, and each convolution layer may generate the same number of feature maps or a different number of feature maps. You may. In addition, the image conversion device provides an input image to the trained neural network 100 and obtains a plurality of feature maps extracted from each layer.

The image conversion device may apply detailed styles of the style image to the input image by using feature maps extracted from all convolution layers for the style image and the input image.

Thereafter, the image conversion device provides the content image to the trained neural network 100 and obtains a plurality of feature maps extracted from one convolution layer among a plurality of convolution layers. Here, one convolution layer may mean any one of an arbitrary one of a plurality of convolution layers, a last output layer of a neural network, or a layer immediately before that.

The image conversion apparatus may apply the main content (shape, contour, etc.) of the content image to the input image by using a feature map extracted from one convolution layer for the content image and the input image.

란 복수의 특징맵 간 상관 관계를 의미하며, 또한 복수의 특징맵을 추출한 복수의 필터 간 상관 관계를 의미하고, 복수의 특징맵에 대한 그램 행렬(Gram matrix)을 통해 연산된다.The calculator 110 calculates a correlation between a plurality of feature maps output from one convolution layer in order to calculate a style loss function. Here, the correlation (

means a correlation between a plurality of feature maps, and also means a correlation between a plurality of filters from which a plurality of feature maps are extracted, and is calculated through a Gram matrix for a plurality of feature maps.

는 l번째 컨볼루션 레이어가 스타일 이미지에 대해 추출한 i개의 특징맵으로부터 연산된 그램 행렬을 의미한다. 그램 행렬은 본 발명이 속하는 기술 분야에서 통상의 지식을 가진 자에게 자명한 기술이므로 이에 대한 자세한 설명은 생략한다.in Figure 2

denotes a gram matrix calculated from i feature maps extracted from the style image by the lth convolution layer. Since the gram matrix is a technology that is obvious to those skilled in the art, a detailed description thereof will be omitted.

The calculator 110 calculates a correlation between a plurality of feature maps extracted from the style image by the convolution layer and a correlation between a plurality of feature maps extracted from the input image by each convolution layer for each convolution layer. do. The calculator 110 calculates a mean square error (MSE) of the correlation for the style image and the correlation for the input image, and determines a sum of these values as the style loss function. The style loss function can be expressed as Equation 1.

는 각 컨볼루션 레이어가 추출한 특징에 대한 가중치(weight), G는 그램 행렬, F(S)는 스타일 이미지에 대한 특징맵, F(I)는 입력 이미지에 대한 특징맵을 의미한다.In Equation 1, l is the number of convolutional layers, i is the number of feature maps extracted from one convolutional layer,

is the weight of the feature extracted by each convolution layer, G is the gram matrix, F(S) is the feature map for the style image, and F(I) is the feature map for the input image.

In addition, the calculation unit 110 calculates the mean square error between a plurality of feature maps for the content image and a plurality of feature maps for the input image among a plurality of feature maps extracted from one convolution layer, and calculates the sum of all of them is determined as the content loss function. The style loss function can be expressed as Equation 2.

)를 결정한다. 총 손실 함수는 스타일 손실 함수에 가중치가 곱해진 값과 컨텐트 손실 함수에 가중치가 곱해진 값의 합일 수 있다. 두 가중치의 비율은 적절하게 조정될 수 있다.The calculation unit 110 calculates a total loss function (based on the style loss function and the content loss function).

) to determine The total loss function may be the sum of a value obtained by multiplying a style loss function with a weight and a value obtained by multiplying a weighted value with a content loss function. The ratio of the two weights can be adjusted appropriately.

The conversion unit 120 converts the style and content of the input image using the total loss function. The conversion process may be performed through Equation 3.

은 변환된 입력 이미지,

는 변환 가중치,

는 경사 하강(gradient descent)를 의미한다.in Equation 3

is the transformed input image,

is the transform weight,

stands for gradient descent.

The input image converted in FIG. 2 is an image to which Equation 3 is applied to the input image. That is, each pixel of the input image is converted into a pixel of the converted input image through Equation 3. Then, the converted input image is used as an input image of the trained neural network 100 .

As the above process is repeated, the style and content of the converted input image become closer to the style of the style image and the content of the content image, and the total loss function decreases with each process. The number of times the converted input image is used again as an input image is determined according to a preset number of times, a preset time period, and a preset loss function value. The image conversion device outputs the finally converted image as an output image.

3 is a diagram for explaining an operation process of an image conversion device using an additive loss function and mixed images according to an embodiment of the present invention.

Referring to FIG. 3 , an image conversion device (not shown) according to an embodiment of the present invention may use a mixed image as an input image in addition to a white noise image and a content image. A mixed image is an image in which pixel values of an input image and pixel values of a style image corresponding to the pixel values are mixed. Generation of the mixed image can be expressed as Equation 4.

는 혼합 이미지,

는 혼합 비율, I는 입력 이미지, S는 스타일 이미지를 의미한다. 혼합 비율은 실험적으로 결정되고, 실시예 마다 달라질 수 있다.in Equation 3

is a blended image,

is the blending ratio, I is the input image, and S is the style image. The mixing ratio is determined experimentally and may vary from example to example.

By mixing the input image with the style image before inputting it to the trained neural network 100, the time required to apply the style of the style image to the input image can be reduced. Also, when the input image is a content image, the time required to apply the style of the style image and the content of the content image to the input image can be significantly reduced.

Since the process of calculating the style loss function and the content loss function by the image conversion device using the mixed images is the same as that described in FIG. 2, a detailed description thereof will be omitted.

Meanwhile, the image conversion apparatus according to an embodiment of the present invention uses relative entropy between a style image and an input image as an additional loss function to reduce image conversion time. Relative entropy refers to the similarity of the pixel distributions of two images.

Specifically, the calculation unit 110 calculates relative entropy between a plurality of feature maps for a style image and a plurality of feature maps for an input image among a plurality of feature maps extracted from the output convolution layer of the trained neural network 100, , The conversion unit 120 converts the style and content of the input image additionally based on the relative entropy. When the output convolution layer is the farthest layer from the input image input layer, the feature map extracted therefrom includes detailed information of the image. Accordingly, the image conversion device may more quickly reflect the style characteristics of the style image to the input image by further using the relative entropy calculated from the output convolution layer.

The relative entropy of the style image and the content image can be expressed as Equation 5.

은 쿨백-라이블러 발산(kullback-leibler divergence),

는 입력 이미지에 대해 출력단 레이어로부터 추출된 복수의 특징맵,

는 스타일 이미지에 대해 출력단 레이어로부터 추출된 복수의 특징맵을 의미한다. 쿨백-라이블러 발산을 계산하는 과정은 해당 기술분야에서 통상의 지식을 가진 자에게 자명하므로 이에 대한 설명은 생략한다.in Equation 5

is the kullback-leibler divergence,

Is a plurality of feature maps extracted from the output stage layer for the input image,

Means a plurality of feature maps extracted from the output stage layer for the style image. Since the process of calculating the Kullback-Leibler divergence is obvious to those skilled in the art, a description thereof will be omitted.

)를 결정한다. 총 손실 함수는 스타일 손실 함수에 가중치가 곱해진 값, 컨텐트 손실 함수에 가중치가 곱해진 값, 및 상대적 엔트로피에 가중치가 곱해진 값의 합일 수 있다. 세 가중치의 비율은 적절하게 조정될 수 있다.The calculation unit 110 calculates a total loss function (based on the style loss function, content loss function, and relative entropy).

) to determine The total loss function may be the sum of a value obtained by multiplying a style loss function with a weight, a value obtained by multiplying a content loss function with a weight, and a value obtained by multiplying a relative entropy with a weight. The ratio of the three weights can be adjusted appropriately.

The conversion unit 120 converts the input image into a converted input image including the style of the style image and the content of the content image by applying the total loss function to the input image.

4 is a flowchart illustrating an operating method of an image conversion device according to an embodiment of the present invention.

Hereinafter, steps S400 to S404 may be performed in each time order, but are not limited thereto, and the three steps may be performed in a different order or in parallel. Also, steps S406 and S408 may be executed out of order or in parallel.

The image conversion device inputs a style image to the trained neural network and acquires a plurality of feature maps for the style image (S400). The trained neural network includes a plurality of convolutional layers and is trained to extract a plurality of feature maps of an image for each convolutional layer.

The image conversion device inputs an input image to the trained neural network and obtains a plurality of feature maps for the input image (S402).

An image conversion device according to an embodiment of the present invention may use a mixed image obtained by mixing pixel values of an input image and pixel values of a style image corresponding to the pixel values as an input image.

The image conversion device inputs the content image to the trained neural network and obtains a plurality of feature maps of the content image (S404).

The image conversion apparatus calculates a style loss function based on a plurality of feature maps for a style image and a plurality of feature maps for an input image among a plurality of feature maps extracted from a plurality of convolution layers (S406). Here, the style loss function is the sum of the mean square errors of the correlation for the style image and the correlation for the input image, calculated for each convolution layer.

The image conversion apparatus calculates a content loss function based on a plurality of feature maps for a content image and a plurality of feature maps for an input image among a plurality of feature maps extracted from one convolution layer (S408). Here, the content loss function is the sum of mean square errors of a plurality of feature maps for the content image and a plurality of feature maps for the input image among a plurality of feature maps extracted from one convolution layer.

The image conversion device converts the style and content of the input image using the style loss function and the content loss function (S410).

The image conversion apparatus according to an embodiment of the present invention further includes a process of calculating relative entropy between a plurality of feature maps for a style image and a plurality of feature maps for an input image among a plurality of feature maps extracted from an output convolution layer. include At this time, the process of converting the style and content of the input image by the image conversion device is a process of converting the style and content of the input image based on the style loss function, the content loss function, and relative entropy.

The image conversion device inputs the converted input image to the trained neural network (S412). The number of times the converted input image is used again as an input image is determined according to a preset number of times, a preset time period, and a preset loss function value. The image conversion device outputs the finally converted image as an output image.

Although it is described in FIG. 4 that steps S400 to S412 are sequentially executed, this is merely an example of the technical idea of an embodiment of the present invention. In other words, those skilled in the art to which an embodiment of the present invention pertains may change and execute the sequence described in FIG. 4 without departing from the essential characteristics of the embodiment of the present invention, or one of steps S400 to S412. Since it will be possible to apply various modifications and variations by executing the above process in parallel, FIG. 4 is not limited to a time-series sequence.

Meanwhile, the processes shown in FIG. 4 can be implemented as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all types of recording devices in which data that can be read by a computer system is stored. That is, such a computer-readable recording medium may be a non-transitory medium such as ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc., and may also be a carrier wave (e.g. , transmission over the Internet) and a transitory medium such as a data transmission medium. In addition, the computer-readable recording medium may be distributed to computer systems connected through a network to store and execute computer-readable codes in a distributed manner.

The above description is merely an example of the technical idea of the present embodiment, and various modifications and variations can be made to those skilled in the art without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of the present embodiment, but to explain, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The scope of protection of this embodiment should be construed according to the claims below, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of rights of this embodiment.

100: trained neural network 110: calculation unit
120: conversion unit

Claims (12)

An image conversion device for applying the style of a style image and the content of a content image to an input image,
A neural network that includes a plurality of convolution layers and is trained to extract a plurality of feature maps for an image for each convolution layer, wherein the trained neural network includes pixel values of an initial image and the pixel values A plurality of feature maps for the input image are extracted in response to input of an input image, which is an image in which pixel values of style images corresponding to values are mixed, and a plurality of feature maps for the style image are extracted in response to input of the style image. Extracting a map and extracting a plurality of feature maps for the content image in response to an input of the content image -;
A style loss function is calculated based on the plurality of feature maps of the style image and the plurality of feature maps of the input image extracted from the plurality of convolution layers, and one selected from the plurality of convolution layers is selected. a calculator for calculating a content loss function based on a plurality of feature maps of the content image and a plurality of feature maps of the input image extracted from a convolution layer; and
a conversion unit adjusting pixel values of the input image using the style loss function and the content loss function;
Image conversion device comprising a. According to claim 1,
The calculation unit further calculates a correlation between a plurality of feature maps output from one convolution layer,
The style loss function is
It is the sum of the mean square error (MSE) of the correlation for the style image and the correlation for the input image, calculated for each convolution layer,
The content loss function,
The image conversion device is the sum of mean square errors between a plurality of feature maps for the content image and a plurality of feature maps for the input image among the plurality of feature maps extracted from the one convolution layer. According to claim 1,
The operation unit further calculates relative entropy between a plurality of feature maps for the style image and a plurality of feature maps for the input image among the plurality of feature maps extracted from the output convolution layer,
The conversion unit converts the style and content of the input image additionally based on the relative entropy. According to claim 3,
The relative entropy is,
An image conversion device that is a Kullback Leibler divergence value between a plurality of feature maps for the style image and a plurality of feature maps for the input image among the plurality of feature maps extracted from the output convolution layer. delete A method of operating an image conversion device that applies the style of a style image and the content of a content image to an input image,
generating, as an input image, a mixed image obtained by mixing pixel values of an initial image and pixel values of a style image corresponding to the pixel values;
inputting a style image to a neural network trained to extract a plurality of feature maps of an image for each convolution layer including a plurality of convolution layers, and acquiring a plurality of feature maps of the style image;
inputting the input image to the trained neural network and acquiring a plurality of feature maps of the input image;
inputting the content image to the trained neural network and acquiring a plurality of feature maps of the content image;
calculating a style loss function based on a plurality of feature maps for the style image and a plurality of feature maps for the input image among a plurality of feature maps extracted from the plurality of convolution layers;
calculating a content loss function based on a plurality of feature maps of the content image and a plurality of feature maps of the input image among a plurality of feature maps extracted from one convolution layer selected from the plurality of convolution layers; ; and
Adjusting pixel values of the input image using the style loss function and the content loss function
Method of operating an image conversion device comprising a. According to claim 6,
The style loss function is
It is the sum of the mean square errors of the correlation for the style image and the correlation for the input image, calculated for each convolution layer,
The content loss function,
The method of operating the image conversion device, which is the sum of mean square errors of a plurality of feature maps for the content image and a plurality of feature maps for the input image among a plurality of feature maps extracted from the one convolution layer. According to claim 6,
Further comprising calculating relative entropy between a plurality of feature maps for the style image and a plurality of feature maps for the input image among the plurality of feature maps extracted from the output convolution layer,
The converting process is a process of converting the style and content of the input image additionally based on the relative entropy. According to claim 8,
The relative entropy is,
A Kullback-Leibler divergence value between a plurality of feature maps for the style image and a plurality of feature maps for the input image among a plurality of feature maps extracted from the output convolution layer. delete According to claim 6,
inputting an input image whose style and content are converted into the trained neural network when a condition set by at least one of a preset number of times, a preset time, and a preset loss function is not satisfied;
Method of operating an image conversion device further comprising a. A computer-readable recording medium recording a program for executing the method of any one of claims 6 to 9 and 11 on a computer.

Images