KR20200126327A - Normalization method for machine-learning and apparatus thereof - Google Patents

Abstract

Provided are a normalization method for machine learning and a device therefor. According to some embodiments of the present disclosure, the normalization method can calculate the normalization parameter value of an input image through a normalization model before inputting the input image to a target model, and normalize the input image by using the calculated normalization parameter value, wherein the input image can be normalized to an image suitable for a target task since the normalization model is updated based on a predicted error of the target model. Therefore, the stability of training and the performance of the target model can be improved.

Description

Normalization method for machine learning and its device {NORMALIZATION METHOD FOR MACHINE-LEARNING AND APPARATUS THEREOF}

The present disclosure relates to a normalization method and apparatus for machine learning. In more detail, the performance of the target model can be improved by removing the latent bias existing in the input image and normalizing the input image to an image suitable for the target task. And a device supporting the method.

The convolutional neural network is a machine learning model specialized for image recognition tasks. Convolutional neural networks are attracting attention in various domains dealing with images because they can automatically extract features from input images and perform tasks such as object classification and object recognition with high accuracy.

The convolutional neural network is a model that mimics the human visual perception process, but the way that the actual convolutional neural network recognizes images is completely different from the way that visual cells recognize images. This is because the convolutional neural network recognizes the input image by analyzing pixel information (e.g. brightness, brightness, saturation, color, etc.) of the image in multiple ways. Due to the way the convolutional neural network operates, a latent bias in the image can act as a factor that hinders the learning of the convolutional neural network. In addition, the potential bias of the image may be caused by various causes, such as differences in image processing techniques, parameters of photographing equipment and/or photographing environments.

For the same reason as above, the pixel information of the image that can maximize the recognition performance of the convolutional neural network may vary depending on the target task of the neural network. For example, a first convolutional neural network performing a first task may exhibit the best recognition performance when learning an image having a first range of intensity, but a second convolutional neural network performing a second task The neural network may exhibit better recognition performance in the second range of brightness. However, a methodology for notifying pixel information of an image that is most appropriate for an objective task has not yet been proposed.

Korean Patent Publication No. 10-2016-0000271 (published on January 4, 2016)

The technical problem to be solved through some embodiments of the present disclosure is to provide a method of normalizing an input image to an image optimized for a target model (or target task) and an apparatus supporting the method in order to improve the performance of the target model. Is to do.

Another technical problem to be solved through some embodiments of the present disclosure is to provide a method capable of removing a latent bias existing in an input image and an apparatus supporting the method to improve stability of learning. .

Another technical problem to be solved through some embodiments of the present disclosure is to provide a method of automatically calculating a parameter value capable of converting an original image into a form of a target image, and an apparatus supporting the method.

The technical problems of the present disclosure are not limited to the technical problems mentioned above, and other technical problems that are not mentioned will be clearly understood by those skilled in the art from the following description.

In order to solve the above technical problem, a normalization method for machine learning according to some embodiments of the present disclosure is a normalization method for machine learning performed by a computing device, wherein the input image is Calculating a value of a normalization parameter, normalizing the input image using the calculated value of the normalization parameter, obtaining a prediction label for the normalized input image through a target model, and the prediction It may include the step of updating the normalization model based on the error of the label.

In order to solve the above technical problem, an image transformation method according to some embodiments of the present disclosure includes, in an image transformation method performed by a computing device, a value of a transformation parameter for a first image through an image transformation model. Calculating, converting the first image using the calculated conversion parameter value, training the image conversion model based on an error between the converted first image and a target image, and the learned It may include calculating a value of a conversion parameter for the second image by using the image conversion model.

A normalization apparatus according to some embodiments of the present disclosure for solving the above-described technical problem includes a memory storing one or more instructions and a normalization of an input image through a normalization model by executing the stored one or more instructions. Calculate a parameter value, normalize the input image using the calculated normalization parameter value, obtain a prediction label for the normalized input image through a target model, and calculate the error of the prediction label. It may include a processor that updates the normalization model based on.

1 and 2 are diagrams for describing a machine learning apparatus and a learning environment according to some embodiments of the present disclosure.
3 is an exemplary flowchart illustrating a normalization method in a learning procedure according to some embodiments of the present disclosure.
4 is a diagram for further explaining a normalization method according to some embodiments of the present disclosure.
5 is an exemplary diagram illustrating a structure of a normalization model and a process of calculating a normalization parameter according to some embodiments of the present disclosure.
6 is an exemplary diagram illustrating a structure of a normalization model and a process of calculating a normalization parameter according to some other embodiments of the present disclosure.
7 is an exemplary flowchart showing a normalization process according to the first embodiment of the present disclosure.
8 and 9 are exemplary diagrams for explaining a normalization process according to the first embodiment of the present disclosure.
10 is an exemplary flow diagram illustrating a normalization process according to a second embodiment of the present disclosure.
11 is an exemplary diagram for describing a normalization process according to a second embodiment of the present disclosure.
12 is an exemplary diagram for describing a normalization process according to a third embodiment of the present disclosure.
13 is an exemplary diagram for explaining a normalization process according to a fourth embodiment of the present disclosure.
14 is an exemplary flowchart illustrating a normalization method in a prediction procedure according to some embodiments of the present disclosure.
15 is a diagram for explaining an example in which the technical idea of the present disclosure is utilized in a medical domain.
16 is an exemplary flowchart illustrating a method of converting an image according to some embodiments of the present disclosure.
17 is an exemplary diagram for further explaining the image transformation model training step S1100 shown in FIG. 16.
18 is an exemplary diagram for describing an image conversion process according to the first embodiment of the present disclosure.
19 and 20 are exemplary diagrams for explaining an image conversion process according to the second embodiment of the present disclosure.
21 to 23 are exemplary diagrams for explaining an image conversion process according to a third embodiment of the present disclosure.
24 illustrates an example computing device that may implement a device according to various embodiments of the present disclosure.

Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. Advantages and features of the present disclosure, and a method of achieving them will be apparent with reference to the embodiments described below in detail together with the accompanying drawings. However, the technical idea of the present disclosure is not limited to the following embodiments, but may be implemented in various different forms, and only the following embodiments complete the technical idea of the present disclosure, and in the technical field to which the present disclosure belongs. It is provided to completely inform the scope of the present disclosure to those of ordinary skill in the art, and the technical idea of the present disclosure is only defined by the scope of the claims.

In adding reference numerals to elements of each drawing, it should be noted that the same elements are assigned the same numerals as possible even if they are indicated on different drawings. In addition, in describing the present disclosure, when it is determined that a detailed description of a related known configuration or function may obscure the subject matter of the present disclosure, a detailed description thereof will be omitted.

Unless otherwise defined, all terms (including technical and scientific terms) used in the present specification may be used as meanings that can be commonly understood by those of ordinary skill in the art to which this disclosure belongs. In addition, terms defined in a commonly used dictionary are not interpreted ideally or excessively unless explicitly defined specifically. The terms used in the present specification are for describing exemplary embodiments and are not intended to limit the present disclosure. In this specification, the singular form also includes the plural form unless specifically stated in the phrase.

In addition, in describing the constituent elements of the present disclosure, terms such as first, second, A, B, (a) and (b) may be used. These terms are only used to distinguish the component from other components, and the nature, order, or order of the component is not limited by the term. When a component is described as being "connected", "coupled" or "connected" to another component, the component may be directly connected or connected to that other component, but another component between each component It will be understood that elements may be "connected", "coupled" or "connected".

As used in the specification, "comprises" and/or "comprising" refers to the presence of one or more other components, steps, actions and/or elements, and/or elements, steps, actions and/or elements mentioned. Or does not exclude additions.

Prior to the description of the present specification, some terms used in the present specification will be clarified.

In the present specification, a task refers to a task to be solved through machine learning or a task to be performed through machine learning. For example, when performing facial recognition, facial expression recognition, gender classification, pose classification, and the like from facial data, each of face recognition, expression recognition, gender classification, and pose classification may correspond to individual tasks. As another example, when recognizing, classifying, and predicting anomaly from a medical image, each of anomaly recognition, anomaly classification, and anomaly prediction may correspond to individual tasks. In addition, the task may be referred to as a target task.

In the present specification, a target model may mean a model that performs a target task and a model to be built through machine learning. Since the target model may be implemented based on any machine learning model including a neural network, the technical scope of the present disclosure is not limited by the implementation method of the target model.

In the present specification, a normalization model may mean a model capable of calculating a value of a normalization parameter for a given image. Since the normalization model may be implemented in various ways, the technical scope of the present disclosure is not limited by the implementation method of the normalization model. For some examples of the normalization model, refer to FIGS. 5 and 6.

In the present specification, an image transformation model may mean a model capable of calculating a value of a transformation parameter for transforming a given image into a shape of a target image. Since the image conversion model may be implemented in various ways, the technical scope of the present disclosure is not limited by the implementation method of the image conversion model.

In this specification, a neural network is a term encompassing all kinds of machine learning models designed by mimicking neural structures. For example, the neural network may include all types of neural network-based models such as an artificial neural network (ANN), a convolutional neural network (CNN), and the like.

In this specification, an instruction refers to a series of computer-readable instructions grouped on a function basis, which is a component of a computer program and executed by a processor.

Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

1 illustrates a machine learning apparatus 10 and a learning environment according to some embodiments of the present disclosure.

As shown in FIG. 1, the machine learning device 10 is a computing device that performs machine learning on a given image set 11 in order to perform a target task. More specifically, the machine learning apparatus 10 may build a target model by machine learning the image set 11 given a label. In addition, the machine learning apparatus 10 may predict the label 15 of the prediction image 13 by performing the target task through the target model. Hereinafter, for convenience of description, the machine learning device 10 will be abbreviated as the learning device 10.

The computing device may be a notebook, a desktop, a laptop, a server, etc., but is not limited thereto, and may include all types of devices equipped with a computing function. Refer to FIG. 15 for an example of the computing device.

1 illustrates that the learning device 10 is implemented as one computing device as an example, but in an actual physical environment, the first function of the learning device 10 is implemented in the first computing device, and the learning device 10 The second function of may be implemented in the second computing device. That is, the function of the learning device 10 may be implemented by a plurality of computing devices. Also, a plurality of computing devices may be implemented by dividing the first function and the second function.

According to various embodiments of the present disclosure, the learning device 10 normalizes an input image using a parameter value calculated by a predetermined machine learning model, or converts the input image into a target image (eg, a tone-mapped image). Can be converted to Hereinafter, in order to clearly distinguish the uses of the machine learning model, the machine learning model is referred to as a normalization model for an embodiment related to normalization, and the machine learning model is an image conversion model for an embodiment related to image conversion. It should be referred to as.

In some embodiments, the training device 10 may perform normalization on the input image using the normalization model before each image is input as a target model. In addition, the learning device 10 may train a target model or predict a label of the image using the normalized image. The normalization may be performed based on the value of the normalization parameter output from the normalization model, and the normalization model may calculate a normalization parameter value appropriate for the target model (or target task) by learning a prediction error of the target model. A more detailed description of the configuration and operation of the normalization model will be described later with reference to FIGS. 3 to 15. According to the present embodiment, a potential bias existing in the input image can be removed through normalization before the input image is input to the target model, and thus stability of learning can be greatly improved. In addition, through the normalization, the pixel information of the input image can be corrected to fit the target model (or the target task) (eg, the values of the pixels in a specific range important for task execution are amplified), which results in the performance of the target model. This can be greatly improved.

In addition, in some embodiments, the training device 10 may perform multi-domain training to further improve the performance of the target model. For example, as illustrated in FIG. 2, multi-domain learning may be performed using image sets 21 and 23 belonging to different domains A and B. At this time, a potential bias may exist in the image sets 21 and 23 of each domain due to various causes, such as differences in photographing environment, parameters of photographing equipment, and/or image processing techniques. In this case, the learning device 10 may remove potential biases existing in the image sets 21 and 23 through a normalization process, and normalize each image to an image suitable for the target model. By doing so, the effect of improving performance according to multi-domain learning can be guaranteed.

As described above, in various embodiments of the present disclosure related to normalization, the learning device 10 may be referred to as the normalization device 10.

In addition, in some embodiments, the learning device 10 may build an image conversion model capable of converting an original image into a target image (e.g. tone-mapped image) through machine learning. In addition, the learning device 10 may provide a value of a conversion parameter calculated by an image conversion model to a user, or convert an original image into a shape of a target image by using the value of the conversion parameter. In this embodiment only, the learning device 10 may be referred to as an image conversion device 10. Hereinafter, for convenience of description, the image conversion model will be abbreviated as a conversion model. A detailed description of the present embodiment will be described later with reference to FIGS. 16 to 23.

So far, the learning device 10 and the learning environment according to some embodiments of the present disclosure have been described with reference to FIGS. 1 and 2. Hereinafter, a normalization method according to various embodiments of the present disclosure will be described in detail with reference to FIGS. 3 to 14.

Each step of the method to be described below may be performed by a computing device. In other words, each step of the method may be implemented with one or more instructions executed by the processor of the computing device. All of the steps included in the method may be executed by one physical computing device, but may be distributed and executed by a plurality of physical computing devices. For example, first steps of the method may be performed by a first computing device, and second steps of the method may be performed by a second computing device. In the following, description will be continued on the assumption that each step of the method is performed by the apparatus 10 illustrated in FIG. 1 or 2. Therefore, when the subject of each operation is omitted in the description of the present embodiment, it may be understood that it can be performed by the illustrated apparatus 10. In addition, it goes without saying that in the method according to the present embodiment, the execution order of each operation may be changed within a range in which the execution order can be logically changed as necessary.

3 is an exemplary flowchart illustrating a normalization method according to some embodiments of the present disclosure. In particular, FIG. 3 illustrates a process in which the normalization method is performed in the learning procedure of the target model. However, this is only a preferred embodiment for achieving the object of the present disclosure, and of course, some steps may be added or deleted as necessary.

As shown in Fig. 3, the normalization method starts in step S100 of obtaining an image set given a label. The image set may mean a training dataset including a plurality of images.

In some embodiments, the image set may include a plurality of image sets associated with different domains. For example, the image set may include a first image set generated by a first photographing device and a second image set generated by a second photographing device. As another example, the image set may include a first image set generated by a first image processing technique and a second image set generated by a second image processing technique. Here, the image processing technique may include all types of processing techniques such as image conversion, image filtering (e.g. bayer filter, etc.), image correction (e.g. white balancing, crop, auto-focusing, etc.). In this case, through normalization, potential biases caused by different photographing devices or different image processing techniques may be removed. Accordingly, the effect of improving the performance of the target model by multi-domain learning can be further improved.

Steps S200 and S300 represent a normalization process for the input image. The input image may mean an image input to a target model. As illustrated in FIG. 4, before the input image 31 is input to the target model 35, the input image 31 may be converted into a normalized image 34 based on the normalized model 33. Also, the normalized image 34 may be input to the target model 35. The reason for normalizing the input image 31 before the input image 31 is input to the target model 35 is to induce stable learning and prevent overfitting by removing potential biases. Another reason is to adjust the pixel information of the input image 31 to suit the desired task. Hereinafter, steps S200 and S300 will be described in detail.

In step S200, a value of a normalization parameter for the input image is calculated through the normalization model. The normalized model includes a learnable parameter, which may be trained (ie, trained to minimize a prediction error) based on a prediction error of a target model. Therefore, the normalization model can accurately calculate a normalization parameter value for converting an input image into an image suitable for the target task.

The normalization model may be designed and selected in various ways according to embodiments. Also, a specific method of calculating the value of the normalization parameter may vary according to embodiments.

In some embodiments, the normalization model may be implemented based on a convolutional neural network. For example, as shown in FIG. 5, the normalization model may be implemented as a convolutional neural network including a feature extraction layer 43 and an output layer 45. The feature extraction layer 43 may include a convolutional layer, and may optionally further include various layers such as a pooling layer. In this case, the normalization model extracts feature data (eg feature map) from the input image 41 through the feature extraction layer 43, and extracts at least one prediction value (eg feature map) based on the feature data through the output layer 45 47-1 to 47-n) can be calculated. In addition, at least one predicted value 47-1 to 47-n may be used as a value of the normalization parameter. In some embodiments, a specific value derived from the predicted values 47-1 to 47-n may be used as the value of the normalization parameter. Since the convolutional neural network is a neural network specialized in image recognition, it can most accurately understand and grasp the relationship between the input image and the normalization parameter. Therefore, according to the present embodiment, the effect of normalization may be further improved by utilizing the characteristics of the image-specific convolutional neural network. Meanwhile, the normalization model may be implemented based on various machine learning models other than the above-described convolutional neural network.

In some other embodiments, a predefined feature is extracted from an input image, and a normalization model may calculate a value of a normalization parameter based on the preset feature. That is, in this embodiment, the normalization model does not automatically extract features from the input image, but uses predefined features. Here, the predefined features may include image style information (e.g. various statistical information such as average and standard deviation), pixel value patterns, and statistical information of pixel values. In addition, features widely known in the art, such as Scale Invariant Feature Transform (SIFT), Histogram of Oriented Gradient (HOG), Haar, and Local Binary Pattern (LBP), may be further included. A specific example of this embodiment is shown in FIG. 6.

6, the feature extraction module 53 extracts at least one of the exemplified features (eg 55-1 to 55-3) from the input image 51, and extracts the extracted features 55- 1 to 55-3) may be input as the normalization model 57. Then, the normalization model 57 may output values of the normalization parameters (e.g. 59-1, 59-2) based on the input features 55-1 to 55-3. 6 shows as an example that the normalization model 57 is implemented as an artificial neural network, the normalization model 57 may be implemented as another type of machine learning model. For example, the normalization model 57 may be implemented based on a traditional machine learning model such as a Support Vector Machine (SVM). According to the present embodiment, an appropriate normalization parameter value may be calculated based on an important characteristic designated by a user. For example, when the target task is a task close to style information, since the value of the normalization parameter is calculated based on the style information of the input image, normalization may be performed appropriately for the target task.

It will be described again with reference to FIG. 3.

In step S300, the input image is normalized using the value of the normalization parameter. The detailed process of this step may vary according to embodiments. Hereinafter, various embodiments related to this step will be described in detail with reference to FIGS. 7 to 13.

First, a normalization process according to a first embodiment of the present disclosure will be described with reference to FIGS. 7 to 9.

In the first embodiment, a normalization parameter is defined based on a range of pixel values, and normalization may be independently performed for each pixel value range. For example, assume that a first normalization parameter is defined for a first pixel value range, and a second normalization parameter is defined for a second pixel value range. In this case, as shown in FIG. 7, a first pixel group consisting of pixels belonging to the first pixel value range is normalized based on a value of a first normalization parameter, and pixels belonging to the second pixel value range The second pixel group consisting of the elements may be normalized based on the second normalization parameter (S320 and S340). In order to provide a more convenient understanding, description will be made with reference to the examples shown in FIGS. 8 and 9.

8 shows that the normalization model 63 calculates the values of five normalization parameters 65-1 to 65-5, and that each normalization parameter 65-1 to 65-5 corresponds to a specific range of pixel values. It is illustrated. For example, in FIG. 8, a first normalization parameter 65-1 corresponds to a pixel value range from 0 to v ₁ , and a second normalization parameter 65-2 corresponds to a pixel value from v ₁ to v ₂ May correspond to a range. The range of pixel values corresponding to each of the normalization parameters 65-1 to 65-5 may be equal or different from each other. In some examples, a range of pixel values in which a plurality of pixels are distributed is subdivided into a plurality of sections, and different normalization parameters may be defined for each section.

9 illustrates that normalization is performed for each pixel of the input image 61. In particular, FIG. 9 shows that the value of the first pixel 61-1 included in the image 61 belongs to the first pixel value range, and the value of the second pixel 61-2 belongs to the second pixel value range. , It is exemplified that the value of the third pixel 61-3 falls within the range of the third pixel value.

As illustrated in FIG. 9, normalization may be performed independently for each pixel value range. Specifically, the value of the first pixel 61-1 is adjusted by the first normalization parameter 65-1, and the value of the second pixel 61-2 is adjusted by the second normalization parameter 65-2. It is adjusted and the value of the third pixel 61-3 can be adjusted by the third normalization parameter 65-3. When all the pixel values of the input image 61 are adjusted, the input image 61 may be converted into a normalized image 67.

Meanwhile, FIGS. 8 and 9 illustrate that the normalization operation is performed by multiplying the pixel values of the input image 61 by the values of the normalization parameters 65-1 to 75-5. However, the normalization operation may be performed in various ways, such as various arithmetic operations such as summation, linear transformation, and nonlinear transformation in addition to multiplication, which may be variously designed and selected according to embodiments.

In addition, in some embodiments, when the input image is composed of multiple channels (e.g. color channels, brightness channels), a normalization parameter may be defined for each channel. For example, a first normalization parameter may be defined for a first channel, and a second normalization parameter may be defined for a second channel. In this case, pixel values of the first channel may be normalized based on the first normalization parameter, and pixel values of the second channel may be normalized based on the second normalization parameter. By doing so, the precision of normalization and the performance of the target model can be improved.

In addition, in some embodiments, a normalization parameter is independently defined for each specific region of the image, and normalization may be independently performed for each specific region of the image based on the normalization parameter. In this case, the precision and accuracy of normalization can be further improved. For a detailed description of the present embodiment, refer to FIGS. 18 to 23.

So far, the normalization process according to the first embodiment of the present disclosure has been described with reference to FIGS. 7 to 9. As described above, normalization may be performed based on a range of pixel values. Accordingly, normalization can be performed more precisely, and performance of the target model can be further improved.

Hereinafter, a normalization process according to a second embodiment of the present disclosure will be described with reference to FIGS. 10 and 11.

In the second embodiment, the input image may be normalized using a sigmoid function. The sigmoid function has the characteristic of converting an input value into an output value within a certain range. Due to this characteristic, when the sigmoid function is applied, the input image can be converted to an appropriately normalized image, and potential bias can be effectively removed.

A detailed process of the normalization process according to the second embodiment is shown in FIG. 10. As shown in FIG. 10, a parameter of the sigmoid function may be set using the value of the normalization parameter (S310). In addition, the input image may be normalized using the set sigmoid function (S330). That is, each pixel value of the input image may be adjusted to a value within an appropriate range through a sigmoid function.

The parameters of the sigmoid function include, for example, a parameter that adjusts the displacement of the sigmoid function (eg, a shift parameter related to the x-axis or y-axis movement), a parameter that adjusts the size of the output value (eg, adjusts the size of the function value). Scale parameter) and the like. However, in addition to the various parameters may be defined, the technical scope of the present disclosure is not limited to the examples listed above.

11 illustrates a case where there are four sigmoid function parameters. As illustrated in FIG. 11, the normalization model 73 calculates values of four normalization parameters 75-1 to 75-4 based on the input image 71, and each normalization parameter 75-1 to 75 Each parameter of the sigmoid function 77 may be set by a value of -4).

In some embodiments, sigmoid function parameters are independently set for each pixel value range and/or channel, and normalization may be performed independently for each pixel value range and/or channel. In this case, the precision and accuracy of normalization can be further improved.

So far, the normalization process according to the second embodiment of the present disclosure has been described with reference to FIGS. 10 and 11. As described above, normalization focusing on normalization may be performed using the characteristics of the sigmoid function. Therefore, the above-described normalization method can effectively remove potential biases included in an image, and can be usefully used to improve the performance of a target model in a multi-domain environment in which various biases exist.

Hereinafter, a normalization process according to a third embodiment of the present disclosure will be described.

In the third embodiment, parameters of the linear transformation model (e.g. scale and shift parameters) may be set based on the value of the normalization parameter calculated by the normalization model. In addition, the input image may be normalized through the linear transformation model.

The technical idea inherent in the third embodiment can be utilized for windowing processing of an image (e.g. a DICOM format medical image). The windowing refers to an operation of extracting pixel values of a specific range from a given image and converting them into a specific shape (e.g. 8-bit grayscale) so that the region of interest can be well viewed. In general, values of windowing parameters (e.g. center, width) are defined in the DICOM header, but even if they are not, the values of the windowing parameters can be predicted through the normalization model. It will be further described with reference to FIG. 12.

As shown in FIG. 12, the normalization model 83 may receive an image 81 and calculate a value of the windowing parameter 85 for the input image 81. In addition, the windowing process is performed on the image 81 according to the value of the windowing parameter 85, and the windowed image 87 may be input as a target model. Since the normalized model 83 learns the prediction error of the target model, the normalized model 83 can calculate the value of the windowing parameter 85 so that the region of interest of the target task is well displayed.

For reference, according to the DICOM protocol, the window area [center-width/2, center+width/2] of the X-ray image defined by the windowing parameters center and width is mapped to [0, 1] of the converted image. (Eg when center = 1000, width = 400, the [800, 1200] area of the actual X-ray image is mapped to [0, 1] of the converted image), so the windowing parameters center and width are linear transformation models. Will have an association of parameters of. In other words, there is always a parameter value of the linear transformation model that allows the same operation as the windowing parameter center and width (eg y = ax + b in the model a = 1/400, b = -2, x-ray Area [800, 1200] of the image is mapped to [0, 1]).

In some other embodiments, a parameter of a nonlinear transformation model (e.g. a sigmoid function) is set based on a value of a normalization parameter calculated by the normalization model, and the input image may be normalized through the nonlinear transformation model.

Hereinafter, a normalization process according to a fourth embodiment of the present disclosure will be described with reference to FIG. 13.

In the fourth embodiment, normalization may be performed based on a combination of the first to third embodiments described above. For example, as shown in FIG. 13, the first normalization process according to the first embodiment is performed on the input image 91, and then the second normalization process according to the second embodiment is further performed. I can. The input image 91 is converted to an image 95 through the first normalization process, and the image 95 is converted to a normalized image 99 through a second normalization process based on the sigmoid function 97. I can.

In some other embodiments, the second normalization process may be performed first, followed by the first normalization process.

So far, the normalization process according to the fourth embodiment of the present disclosure has been described with reference to FIG. 13. As described above, not only the pixel value of the input image is adjusted to a value suitable for the target model through the first normalization process, but also the potential bias existing in the input image can be effectively removed through the second normalization process. Through this, the performance of the target model can be further improved.

So far, a normalization process according to various embodiments of the present disclosure has been described with reference to FIGS. 7 to 13. With reference to FIG. 3 again, a description of the subsequent steps will be continued.

In step S400, a predictive label for the input image normalized through the target model is obtained. That is, a normalized input image may be input to the target model, and a label of the normalized input image may be predicted through the target model.

In step S500, the target model and the normalization model are updated based on the error of the prediction label. As shown in FIG. 4, the error 37 of the prediction label may be calculated based on a difference between the prediction label 36 output from the target model 35 and the correct answer label 32. The error may be calculated by various loss functions. Since the type of the loss function may vary depending on the target task, the technical scope of the present disclosure is not limited by the type of the loss function.

In addition, the target model 35 and the normalized model 33 may be updated in a direction that minimizes the error 37. The update may be understood as adjusting a value of a learnable parameter (e.g. a weight parameter of a neural network) included in each model. By adjusting the value of the learnable parameter of each model (33, 35), the target model (35) can be trained to perform the target task more accurately, and the normalization model (33) fits the input image to the target model. It can be learned to be able to normalize to an image. That is, as the update proceeds, the normalization model 33 can more accurately calculate the value of the normalization parameter.

In step S600, it is determined whether or not the learning has ended. Whether or not to end learning may be determined based on a preset end condition. In addition, the termination condition may be defined and set based on various criteria such as an epoch, whether there is unlearned data, and a performance of a target model. Therefore, the technical scope of the present disclosure is not limited to a specific termination condition.

In response to a determination that the termination condition is not satisfied, steps S200 to S600 described above may be performed again. In response to determining that the termination condition is satisfied, the learning of the target model (or the normalized model) may be terminated.

So far, the normalization method performed in the learning procedure of the target model has been described with reference to FIGS. 3 and 13. Hereinafter, a normalization method performed in the prediction procedure will be described with reference to FIG. 14.

When the target model is trained, a prediction procedure may be performed using the target model. The prediction procedure may mean performing a target task on an image for prediction by using the learned target model. In addition, the prediction image may mean an image to which no label is given. The detailed process of the prediction procedure is shown in FIG. 14.

As shown in FIG. 14, the prediction procedure starts in step S700 of obtaining an image for prediction.

In steps S800 and S900, a normalization process is performed on the prediction image. Since the normalization process is performed based on the learned normalization model, the prediction image may be converted into an image suitable for the target model (ie, an image that the target model can easily analyze).

In step S1000, a label for the normalized prediction image is predicted through the target model. Since the normalized prediction image is an image converted into a state suitable for a target model, the target model can more accurately predict the label of the prediction image. That is, since a potential bias is removed from the prediction image through the normalization process and pixel values of the prediction image are adjusted to values optimized for the target model, the performance of the target model can be improved.

So far, a method of performing normalization in the prediction procedure has been described with reference to FIG. 14.

The technical idea of the present disclosure described so far and various embodiments thereof may be used to improve the performance of the target model in various fields. As an example of this, the technical idea of the present disclosure will be briefly introduced with reference to FIG. 15 that is utilized in the medical field (or domain).

As shown in FIG. 15, the technical idea of the present disclosure may be used to build a diagnostic model 107 in the medical field. The diagnostic model 107 may be, for example, a model that detects a lesion, classifies a lesion, and recognizes a location of a lesion, but the technical scope of the present disclosure is not limited thereto. In particular, FIG. 15 shows as an example that the diagnosis model 107 is implemented based on a convolutional neural network and a diagnosis related to breast cancer is performed in the medical image 101.

The normalization model 103 according to various embodiments of the present disclosure may normalize the input image 101 to an image 105 suitable for the diagnostic model 107 before the input image 101 is input to the diagnostic model 107. I can. In addition, the diagnostic model 107 may be trained using the normalized image 105, and the diagnostic model 107 may perform diagnosis using the normalized image 105. By doing so, the accuracy of the diagnosis result can be improved.

Pixel information of the medical image requested by the diagnostic model 107 may be different depending on the type of medical image, the type of diagnostic task, and the like. In this case, the normalization model 103 may convert the input image into an image optimized for the diagnostic model 107 by learning the prediction error of the diagnostic model 107. Accordingly, the accuracy of the diagnosis model 107 is improved, and a diagnosis auxiliary service with high satisfaction may be provided based on the diagnosis model 107.

Hereinafter, an image conversion method according to some embodiments of the present disclosure will be described with reference to FIGS. 16 to 23.

16 is an exemplary flowchart illustrating a method of converting an image according to some embodiments of the present disclosure. However, this is only a preferred embodiment for achieving the object of the present disclosure, and of course, some steps may be added or deleted as necessary.

As shown in Fig. 16, the image conversion method starts in step S1100 of training a conversion model. More specifically, the transformation model may be trained using a first original image and a target image (ie, a correct answer image). Here, the target image may mean an image for which an original image is desired to be converted through a conversion model. For example, if an original image is to be converted to a tone-mapped image by a specific filter (or tone mapping function), the target image is toned by applying the specific filter (or tone mapping function) to the first original image. It can be a mapped image. For another example, when an original image is to be converted into an image of a specific style, the target image may be an image obtained by converting the first original image to the specific style. The target image may be obtained by various image conversion/correction programs, but the method of obtaining the target image may be any method.

The detailed process of building the transformation model is illustrated in FIG. 17. As illustrated in FIG. 17, the transformation model 112 may receive an original image 111 and calculate a value of the transformation parameter 113 for the original image 111. In addition, the image conversion process is performed on the original image 111 based on the value of the conversion parameter 113, and as a result, the converted image 114 is obtained, and the converted image 114 and the target image 115 The transformation model 112 may be trained based on the error 116 between ). When learning is performed in this way, the transformation model 112 calculates a value of a parameter capable of converting the input original image into a target image (e.g. 115). The transformation model 112 may be implemented as an image-specific neural network such as a CNN, but the technical scope of the present disclosure is not limited thereto.

The conversion parameter 113 may include at least one of, for example, color temperature, luminance, and saturation of an image. However, the type of the conversion parameter 113 is not limited to the examples listed above. For example, the conversion parameter 113 may include parameters of various conversion functions (e.g. tone mapping function, style conversion function) capable of converting an original image into a shape of a target image. This will be described with reference to FIG. 16 again.

In step S1200, a value of a transformation parameter for the second original image is predicted using the learned transformation model. The second original image may be an image different from the aforementioned first original image, but may be the same image. In this step, when the second original image is input to the learned transformation model, a value of a transformation parameter for the second original image may be calculated through the learned transformation model.

In step S1300, an image conversion process for converting the second original image into a shape of a target image is performed based on the calculated conversion parameter value. For example, the color temperature of the second original image may be adjusted according to the value of the color temperature parameter.

In some embodiments, the value of the calculated conversion parameter may be provided to the user. Then, the user may directly perform image conversion using the provided parameter value as an initial value. For example, the user may set an initial value of a conversion parameter (e.g. color temperature) as the supplied parameter value through an image conversion program, finely adjust the value of the conversion parameter, and perform image conversion. According to the present embodiment, since a user can perform image conversion (e.g. tone mapping) in which his or her subjectivity is mixed, the usability of the conversion model can be further increased.

Further, in some embodiments, the calculated conversion parameter value may be applied to a specific mode or a specific setting function and provided to the user. In addition, the user may select a specific mode and a specific setting function provided to perform image conversion using the calculated conversion parameter value.

So far, an image conversion method according to some embodiments of the present disclosure has been described with reference to FIG. 17. Hereinafter, embodiments of the present disclosure for performing more precise and advanced image conversion will be described with reference to FIGS. 18 to 23. Embodiments to be described below may also be applied to the normalization method described with reference to FIGS. 3 to 15.

18 is an exemplary diagram for describing an image conversion process according to the first embodiment of the present disclosure.

As shown in FIG. 18, in the first embodiment, conversion is performed for each area (123-1 to 123-3) of the original image 121, so that the original image 121 is converted to the image 129. I can. For example, the first area 123-1 of the original image 121 is converted based on the value of the first conversion parameter 127-1, and the second area 123-2 of the original image 121 Is converted based on the value of the second conversion parameter 127-2, and the third area 123-3 of the original image 121 may be converted based on the value of the third conversion parameter 127-3. have. In this case, different transformation methods may be applied for each region of the original image 121 or the same transformation method may be applied. Values of the transformation parameters 127-1 to 127-3 applied to each of the image regions 123-1 to 123-3 may be calculated by the transformation model 125. The transformation model 125 may be trained in the above-described manner to calculate values of the transformation parameters 127-1 to 127-3.

In some embodiments, a plurality of transformation models are used, and each transformation model may correspond to each region of the original image 121. In this case, the first region 123-1 of the original image 121 is transformed based on the value of the first transformation parameter calculated by the first transformation model, and the second region 123 of the original image 121 -2) is converted based on the value of the second conversion parameter calculated by the second conversion model, and the third area 123-3 of the original image 121 is a third conversion calculated by the third conversion model. It can be converted based on the value of the parameter.

18 shows by way of example that each area 123-1 to 123-3 of the original image 121 is defined as a set of consecutive pixels. However, this is only to provide convenience of understanding, and the area on the original image may be defined in any number of different ways (e.g. discrete form). For example, the first region of the original image may be defined as a set of pixels belonging to the first pixel value range, and the second region may be defined as a set of pixels belonging to the second pixel value range.

19 and 20 are exemplary diagrams for explaining an image conversion process according to a second embodiment of the present disclosure

The second embodiment is a more advanced embodiment of the above-described first embodiment, and performs image conversion for each area of the original image, but in order to prevent blurring of edges, a three-dimensional grid (eg bilateral grid).

As shown in FIG. 19, the three-dimensional grid 134 has two axes (x-axis and y-axis) representing the position of pixels on the original image 131 and one axis representing the size of a pixel value (eg intensity). It can be defined by (z-axis). When the 3D grid 134 shown in FIG. 19 is used, even if the pixels are adjacent to each other on the original image 131, they may be mapped to different grids according to a difference in pixel values. For example, two adjacent pixels 132 and 133 based on an edge may be mapped to different grids 135 and 136 on the 3D grid 134. Accordingly, when image transformation is performed for each grid, image transformation may be performed in consideration of a positional relationship and a size relationship of pixels, and edges may be preserved even after image transformation. Here, performing image transformation on the grid means performing image transformation on a set of pixels mapped to the grid (ie, a specific area on the original image). It will be further described with reference to FIG. 20.

As shown in FIG. 20, the transformation model 142 may calculate values of a plurality of transformation parameters 143-1 to 143-n for the original image 141. Each of the conversion parameters 143-1 to 143-n may correspond to one or more grids. Also, image conversion may be performed for each grid by using values of the conversion parameters 143-1 to 143-n. For example, image conversion is performed on the first grid 144-1 using the value of the first conversion parameter 143-1, and the second conversion parameter 143-2 is used. Image conversion is performed on the grid 144-2, and image conversion on the n-th grid 144-n may be performed using the value of the n-th conversion parameter 143-n. When image conversion is performed for each grid, a more natural and elaborately converted image 145 may be generated while preserving the edges of the original image 141.

Hereinafter, an image conversion process according to a third embodiment of the present disclosure will be described.

The third embodiment is also similar to the second embodiment described above in that image conversion is performed for each area defined on the original image. However, if the second embodiment described above defines each region using a three-dimensional grid, the third embodiment is different from the second embodiment in that each region is defined using a super-pixel algorithm. There is. The superpixel algorithm refers to an algorithm that groups one or more pixels with similar characteristics into superpixels. Representative examples of superpixel algorithms include Simple Linear Iterative Clustering (SLIC) and SEEDS algorithms. However, the technical scope of the present disclosure is not limited to the examples listed above, and other types of superpixel algorithms may be used. Those skilled in the art will be able to clearly understand the superpixel algorithm, and a detailed description thereof will be omitted. Hereinafter, the third embodiment will be described in detail with reference to the exemplary diagrams shown in FIGS. 21 to 23.

21 illustrates that pixels having similar characteristics are grouped into 12 superpixels (e.g. 154, 155) by applying a superpixel algorithm to the original image 151. As illustrated in FIG. 21, when superpixels (e.g. 154, 155) are formed, image conversion may be performed for each superpixel. Of course, image conversion may be performed for each area including one or more superpixels.

For the image transformation, the transformation model may receive the original image 151 and superpixel information, and may calculate a value of a transformation parameter for transforming each superpixel (e.g. 154, 155). In this case, a specific method of inputting superpixel information into the transformation model may vary according to embodiments.

In some embodiments, as illustrated in FIG. 22, a map image 156 including a plurality of superpixel information may be input as a transform model 157 together with the original image 151. For example, the map image 156 may be stacked on the original image 151 to form a multi-channel image, and the multi-channel image may be input as the transformation model 157. Then, the transformation model 157 may calculate values of transformation parameters 158-1 to 158-12 corresponding to each superpixel (e.g. 154, 155). The map image 156 may be configured to have information (e.g. an identifier) of a superpixel to which a corresponding pixel belongs for each pixel, but the technical scope of the present disclosure is not limited thereto.

In some other embodiments, as shown in FIG. 23, a plurality of map images 161-1 to 161-12 including individual superpixel information are input to the transform model 163 together with the original image 151. I can. For example, a multi-channel image is formed by stacking a plurality of map images 161-1 to 161-12 on the original image 151, and the multi-channel image may be input as the transformation model 163. have. Then, the transformation model 163 may calculate values of the transformation parameters 165-1 to 165-12 corresponding to each superpixel (e.g. 154, 155). Here, the first map image 161-1 is constructed by marking a specific value (eg 1) on a pixel belonging to the first superpixel, and marking a different value (eg 0) on a pixel not belonging to the first superpixel. Can be. In addition, the twelfth map image 161-12 is constructed by marking a specific value (eg 1) on a pixel belonging to the twelfth superpixel, and marking a different value (eg 0) on a pixel not belonging to the twelfth superpixel. Can be. However, the technical scope of the present disclosure is not limited thereto, and the map image (e.g. 161-1) may be configured in any manner as long as it can represent information of individual superpixels.

According to the third embodiment, since image conversion is performed for each superpixel having similar characteristics, sophisticated image conversion can be performed while maintaining the characteristics of the pixels.

Up to now, an image conversion process according to various embodiments of the present disclosure has been described with reference to FIGS. 18 to 23. As described above, since different conversion methods may be applied for each of a plurality of regions included in the original image, more precise and highly sophisticated image conversion can be performed. For example, by performing various types of tone mapping for each area, a more natural and elaborately converted tone mapping image may be generated.

Hereinafter, an exemplary computing device 170 capable of implementing a device (e.g. a learning device, a normalization device, and an image conversion device 10 of FIG. 1) according to various embodiments of the present disclosure will be described with reference to FIG. 24.

24 is a hardware configuration diagram illustrating the computing device 170.

As shown in FIG. 24, the computing device 170 is a memory for loading a computer program executed by one or more processors 171, a bus 173, a communication interface 174, and a processor 171 ( 172) and a storage 175 for storing a computer program 176. However, only the components related to the embodiment of the present disclosure are shown in FIG. 24. Accordingly, those of ordinary skill in the art to which the present disclosure belongs may recognize that other general-purpose components may be further included in addition to the components illustrated in FIG. 24.

The processor 171 controls the overall operation of each component of the computing device 170. Processor 171 is at least one of a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), or any type of processor well known in the technical field of the present disclosure. It can be configured to include. In addition, the processor 171 may perform an operation on at least one application or program for executing a method/operation according to various embodiments of the present disclosure. The computing device 170 may include one or more processors.

The memory 172 stores various types of data, commands and/or information. The memory 172 may load one or more programs 176 from the storage 175 to perform various methods/operations according to embodiments of the present disclosure. The memory 172 may be implemented as a volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

The bus 173 provides communication functions between components of the computing device 170. The bus 173 may be implemented as various types of buses such as an address bus, a data bus, and a control bus.

The communication interface 174 supports wired/wireless Internet communication of the computing device 170. In addition, the communication interface 174 may support various communication methods other than Internet communication. To this end, the communication interface 174 may be configured to include a communication module well known in the art. In some embodiments, the communication interface 174 may be omitted.

The storage 175 may non-temporarily store the one or more programs 176. The storage 175 is a nonvolatile memory such as a Read Only Memory (ROM), an Erasable Programmable ROM (EPROM), an Electrically Erasable Programmable ROM (EEPROM), a flash memory, or the like, a hard disk, a removable disk, or well It may be configured to include any known computer-readable recording medium.

Computer program 176 may include one or more instructions that when loaded into memory 172 cause processor 171 to perform a method/operation in accordance with various embodiments of the present disclosure. That is, the processor 171 may perform methods/operations according to various embodiments of the present disclosure by executing the one or more instructions.

For example, the computer program 176 may perform an operation of calculating a value of a normalization parameter for an input image through a normalization model, an operation of normalizing the input image using the calculated value of the normalization parameter, and a target model. ) To obtain a prediction label for the normalized input image and an operation of updating the normalization model based on an error of the prediction label. In this case, the learning device 10 or the normalization device 10 according to some embodiments of the present disclosure may be implemented through the computing device 170.

For another example, the computer program 176 is an operation of calculating a value of a conversion parameter for a first image through an image conversion model, an operation of converting the first image using the calculated value of the conversion parameter, Instructions for performing an operation of training the image conversion model based on an error between the converted first image and a target image and calculating a value of a conversion parameter for a second image using the learned image conversion model. Can include. In this case, the learning device 10 or the image conversion device 10 according to some embodiments of the present disclosure may be implemented through the computing device 170.

So far, various embodiments of the present disclosure and effects according to the embodiments have been mentioned with reference to FIGS. 1 to 24. The effects according to the technical idea of the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

The technical idea of the present disclosure described so far with reference to FIGS. 1 to 24 may be implemented as computer-readable codes on a computer-readable medium. The computer-readable recording medium is, for example, a removable recording medium (CD, DVD, Blu-ray disk, USB storage device, removable hard disk) or a fixed recording medium (ROM, RAM, computer-equipped hard disk). I can. The computer program recorded in the computer-readable recording medium may be transmitted to another computing device through a network such as the Internet and installed in the other computing device, thereby being used in the other computing device.

In the above, even if all the constituent elements constituting the embodiments of the present disclosure have been described as being combined into one or operating in combination, the technical idea of the present disclosure is not necessarily limited to these embodiments. That is, as long as it is within the scope of the object of the present disclosure, one or more of the components may be selectively combined and operated.

Although the operations are illustrated in a specific order in the drawings, it should not be understood that the operations must be executed in the specific order shown or in a sequential order, or all illustrated operations must be executed to obtain a desired result. In certain situations, multitasking and parallel processing may be advantageous. Moreover, the separation of the various components in the above-described embodiments should not be understood as necessitating such separation, and the program components and systems described may generally be integrated together into a single software product or packaged into multiple software products. It should be understood that there is.

Although the embodiments of the present disclosure have been described with reference to the accompanying drawings, the present disclosure may be implemented in other specific forms without changing the technical spirit or essential features of those of ordinary skill in the art. I can understand that there is. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not limiting. The scope of protection of the present disclosure should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the technical ideas defined by the present disclosure.

Claims (20)

In the image transformation (transformation) method performed by the computing device,
Learning an image conversion model for converting an image using a first image that is a target image and a target image obtained by converting the first image; And
Comprising the step of transforming a second image using the learned image transformation model
Image conversion method. The method of claim 1,
The step of training the image transformation model
Calculating a conversion parameter value for the first image,
Converting the first image based on the calculated conversion parameter value,
And training the image conversion model based on an error between the converted first image and the target image.
Image conversion method. The method of claim 2,
The conversion parameter value includes a first parameter value corresponding to a first area of the first image and a second parameter value corresponding to a second area of the first image,
Converting the first image,
Transforming the first region of the first image by using the first parameter value,
Transforming the second region of the first image using the first parameter value,
Image conversion method. The method of claim 3,
The plurality of pixels included in the first image are mapped to a three-dimensional grid defined by two axes representing a position of a pixel and one axis representing a size of a pixel value,
The first area is determined by a pixel mapped to a first grid among a plurality of grids included in the 3D grid,
The second area is determined by a pixel mapped to a second grid among the plurality of grids,
Image conversion method. The method of claim 3,
The first region corresponds to a first superpixel among a plurality of superpixels obtained by applying a super-pixel algorithm to the first image,
The second region corresponds to a second superpixel among the plurality of superpixels,
The step of calculating the conversion parameter value,
Comprising the step of calculating the first parameter value and the second parameter value by inputting information on the first image, the first superpixel, and the second superpixel into the image conversion model,
Image conversion method. The method of claim 2,
The conversion parameter value is
Including a value for at least one of color temperature, luminance, and saturation associated with the image,
Image conversion method The method of claim 1,
The target image above is
An image generated by applying at least one filter to the first image
Image conversion method. The method of claim 1,
The image conversion model is
Convolutional Neural Network
Image conversion method. The method of claim 1,
Converting the second image
Calculating a conversion parameter value for the second image by using the learned image conversion model,
Converting the second image based on the calculated conversion parameter value for the second image
Image conversion method. The method of claim 9,
Further comprising the step of receiving an initial value related to the conversion parameter value,
Converting the second image
Converting the second image based on a conversion parameter value reflecting the input initial value
Image conversion method. At least one processor; And
Contains a memory for storing one or more instructions,
The processor executes the stored one or more instructions,
Using a first image that is a target image and a target image obtained by converting the first image, an image conversion model for converting the image is trained,
Converting the second image using the learned image conversion model
Electronic device. The method of claim 11,
The processor is
Calculate a conversion parameter value for the first image,
Transforming the first image based on the calculated conversion parameter value,
Training the image conversion model based on an error between the converted first image and the target image
Electronic device. The method of claim 12,
The conversion parameter value includes a first parameter value corresponding to a first area of the first image and a second parameter value corresponding to a second area of the first image,
The processor is
Transform the first area of the first image using the first parameter value,
Transforming the second area of the first image by using the first parameter value
Electronic device The method of claim 13,
The plurality of pixels included in the first image are mapped to a three-dimensional grid defined by two axes representing a position of a pixel and one axis representing a size of a pixel value,
The first area is determined by a pixel mapped to a first grid among a plurality of grids included in the 3D grid,
The second area is determined by a pixel mapped to a second grid among the plurality of grids,
Electronic device. The method of claim 13,
The first region corresponds to a first superpixel among a plurality of superpixels obtained by applying a super-pixel algorithm to the first image,
The second region corresponds to a second superpixel among the plurality of superpixels,
The processor is
Inputting information on the first image, the first superpixel, and the second superpixel into the image conversion model to calculate the first parameter value and the second parameter value
Electronic device. The method of claim 12,
The conversion parameter value is
Including a value for at least one of color temperature, luminance, and saturation associated with the image,
Electronic device. The method of claim 11,
The target image above is
An image generated by applying at least one filter to the first image
Electronic device. The method of claim 11,
The image conversion model is
Convolutional Neural Network
Electronic device. The method of claim 11,
The processor is
Using the learned image transformation model, calculate a transformation parameter value for the second image,
Converting the second image based on the calculated conversion parameter value for the second image
Electronic device. The method of claim 19,
The processor is
Receiving an initial value related to the conversion parameter value,
Converting the second image based on a conversion parameter value reflecting the input initial value
Electronic device.

Images