Image correction system and image correcting method thereof

1. An image correcting method comprising:

a step of performing a preprocessing process on an original image to generate a mask image including only an erased area of the original image;

a step of predicting, by using generative adversarial networks, an image which is to be synthesized with the erased area in the mask image; and

a step of synthesizing the predicted image with the erased area of the original image to generate a new image,

wherein the method further comprises, before the step of generating the mask image, a step of allowing a generative neural network to learn by using big training data,

wherein the big training data comprises training the original image, training the mask image including only the erased area, training sketch images where various shapes are sketched in the erased area, and training color images where various colors are filled into the erased area.

2. The image correcting method of claim 1 , wherein

the generative adversarial networks comprise a generative neural network and a discriminative neural network, which have learned to maintain an adversarial relationship therebetween, and

the predicting step is a step of predicting the image to be synthesized by using the generative neural network.

3. The image correcting method of claim 1 ,

wherein

the step of allowing the generative neural network to learn comprises:

allowing the generative neural network to learn by using the big training data; and

allowing a discriminative neural network to learn to maintain an adversarial relationship with the generative neural network,

wherein the predicting step is a step of predicting the image to be synthesized by using the learning-completed generative neural network.

4. The image correcting method of claim 3 , wherein

the step of allowing the generative neural network to learn is a step of allowing the generative neural network to learn by using loss LG output from the discriminative neural network, and

the loss LG comprises a pixel difference value (Igen−Igt) between a pixel value M of an erased area of a training mask image included in the big training data, the original image, and the new image.

5. The image correcting method of claim 3 , wherein the step of allowing the generative neural network to learn is a step of allowing a gated convolution layer included in the generative neural network to learn by using a pixel difference value (Igen −Igt) between a pixel value M of an erased area of a training mask image included in the big training data, the original image, and the new image.

6. An image correction system comprising:

a preprocessor performing a preprocessing process on an original image to generate a mask image including only an area, erased based on a user input, of the original image, a sketch image including only a shape sketched in the erased area on the basis of the user input, and a color image including only a color painted in the erased area on the basis of the user input;

an image generator predicting, by using generative adversarial networks, an image to be synthesized with the erased area from the mask image, the sketch image, and the color image and synthesizing the predicted image with the erased area to generate a new image from the original image; and

a display unit displaying the new image.

7. The image correction system of claim 6 , wherein

the generative adversarial networks comprise a generative neural network and a discriminative neural network, which have learned to maintain an adversarial relationship therebetween, and

the image generator predicts the image to be synthesized by using the generative neural network.

8. The image correction system of claim 6 , further comprising:

a learning unit allowing the generative neural network to learn by using big training data; and

a storage unit storing the generative neural network where learning is completed by the learning unit,

wherein the big training data comprises training original images, training mask images including only the erased area, training sketch images where various shapes are sketched in the erased area, and training color images where various colors are filled into the erased area.

9. The image correction system of claim 8 , wherein

the learning unit allows the generative adversarial networks including a generative neural network and a discriminative neural network to learn to maintain an adversarial relationship therebetween, and

the storage unit stores the learning-completed generative neural network.

10. The image correction system of claim 9 , wherein

the learning unit allows the generative neural network to learn to decrease loss LG output from the discriminative neural network, and

the loss LG comprises a pixel difference value (Igen−Igt) between a pixel value M of an erased area of the training mask image included in the big training data, the original image, and the new image.

11. The image correction system of claim 9 , wherein the learning unit allows a gated convolution layer included in the generative neural network to learn by using the pixel difference value (Igen−Igt) between the pixel value M of the erased area of the training mask image included in the big training data, the original image, and the new image.

12. An image correcting method comprising:

a step of allowing generative adversarial networks including a generative neural network and a discriminative neural network having an adversarial relationship therebetween to learn;

a step of storing the learning-completed generative neural network in a storage unit;

a step of generating a mask image including only an area, erased based on a user input, of the original image, a sketch image including only a shape sketched in the erased area on the basis of the user input, and a color image including only a color painted in the erased area on the basis of the user input, from an original image through a preprocessing process;

a step of predicting, by using the generative neural network stored in the storage unit, an image to be synthesized with the erased area from the mask image, the sketch image, and the color image; and

a step of synthesizing the predicted image with the erased area to generate a new image from the original image.

13. The image correcting method of claim 12 , wherein

the learning step is a step of allowing a generative neural network and a discriminative neural network to learn to maintain an adversarial relationship therebetween, and

the step of allowing the generative neural network to learn is a step of allowing a gated convolution layer included in the generative neural network to learn by using a pixel difference value (Igen−Igt) between a pixel value M of an erased area of a training mask image included in big training data, the original image, and the new image.

14. The image correcting method of claim 12 , wherein, in the step of predicting the image to be synthesized with the erased area, the image to be synthesized is a realistic image predicted based on the shape sketched in the erased area.

15. The image correcting method of claim 12 , wherein the original image is a face image.