Device and method to generate image using image learning model

1. A method of generating a synthetic image, the method comprising:

receiving input images;

extracting feature values corresponding to the input images using a first image learning model, the first image learning model permitting an input and an output to be identical; and

generating a synthetic image based on the feature values corresponding to the input images using the first image learning model, the generating the synthetic image includes,

combining the feature values, wherein each of the feature values are extracted from a different corresponding input image among the input images, and

generating the synthetic image from the combined feature values using the first image learning model.

2. The method of claim 1 , wherein the first image learning model comprises a neural network, the neural network including at least one layer connected to another layer through a connection line, the connection line having a connection weight.

3. The method of claim 1 , wherein the combining the feature values comprises:

combining the feature values based on a set ratio.

4. The method of claim 2 , wherein the feature values correspond to an uppermost layer of the neural network from each input image.

5. The method of claim 2 , wherein the connection weight reduces an error between a reference input image and an output image corresponding to the reference input image.

6. The method of claim 1 , wherein the input images comprise face images of different individuals.

7. The method of claim 1 , further comprising:

receiving the synthetic image and angle information, the synthetic image including an object; and

generating an output image based on the synthetic image and the angle information using a second image learning model, the second image learning model being trained in pairs of references images corresponding to different angles, the output image including the object.

8. The method of claim 7 , wherein the generating the output image comprises:

generating the output image by converting the synthetic image to correspond to the angle information using the second image learning model.

9. The method of claim 7 , wherein

the pairs of reference images comprise a first reference angle image corresponding to a first angle and a second reference angle image corresponding to a second angle, and

wherein the first reference angle image and the second reference angle image include an identical reference object.

10. The method of claim 9 , wherein the second image learning model comprises a neural network, the neural network including at least one layer connected to another layer through a connection line, the connection line having a connection weight.

11. The method of claim 10 , wherein the connection weight of the neural network allows the second reference angle image to be output based on the second angle and the first reference angle image.

12. The method of claim 7 , wherein the angle information comprises information associated with an angle for the object.

13. The method of claim 7 , wherein the object comprises a human face.

14. The method of claim 7 , further comprising:

removing noise from at least one patch in the output image using at least one third image learning model.

15. The method of claim 14 , wherein the third image learning model comprises a neural network, the neural network including at least one layer connected to another layer through a connection line, the connection line having a connection weight.

16. The method of claim 15 , wherein the connection weight is learned to allow a reference output image to be generated from a patch of a reference input image.

17. The method of claim 14 , wherein

the third image learning model permits a reference output image from which noise is eliminated to be generated from a reference input image including noise, and

the removing the noise includes,

eliminating the noise from the output image using the third image learning model.

18. The method of claim 14 , wherein the third image learning model comprises:

a fourth learning model to allow a first size patch of a reference output image from which noise is eliminated to be generated from the first sized patch of a reference input image; and

a fifth learning model to allow a second size patch of the reference output image from which noise is eliminated to be generated from the second size patch of the reference input image.

19. The method of claim 18 , wherein the removing the noise comprises:

eliminating noise from the output image in a unit of the first size patch using the fourth learning model; and

eliminating noise in a unit of the second size patch, using the fifth learning model, from the output image from which the noise is eliminated in the unit of the first size patch.