Training method and apparatus for image processing model, image processing method and apparatus for image processing model, and storage medium

1. A training method for an image processing model, the method comprising:

updating, by circuitry, parameters of an encoder in the image processing model according to a to-be-replaced face in an original image, to configure the encoder to encode the to-be-replaced face to obtain a visual feature of the to-be-replaced face;

updating parameters of a decoder in the image processing model according to the to-be-replaced face in the original image, to configure the decoder to perform decoding based on the visual feature of the to-be-replaced face; and

updating the parameters of the decoder according to a target face in a plurality of target images without changing the parameters of the encoder, to configure the decoder to perform decoding based on the visual feature of the to-be-replaced face and obtain the target face having the same visual feature as the to-be-replaced face,

wherein the decoder is trained with the target face in the plurality of target images and the encoder is not trained with the target face in the plurality of target images.

2. The method according to claim 1 , wherein the updating the parameters of the encoder comprises:

substituting a facial image of a to-be-replaced face in a distorted original image of the original image into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determining the parameters corresponding to the encoder based on the loss function satisfying a convergence condition.

3. The method according to claim 1 , wherein the updating the parameters of the decoder comprises:

substituting a facial image of a to-be-replaced face in a distorted original image of the original image into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determining the parameters of the decoder based on the loss function satisfying a convergence condition.

4. The method according to claim 1 , wherein the updating the parameters of the decoder comprises:

substituting a facial image of the target face in a distorted target image of the plurality of target images into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determining, without changing the parameters of the encoder in the loss function, the parameters of the decoder based on the loss function satisfying a convergence condition.

5. The method according to claim 1 , wherein before the updating the parameters of the encoder, the method further comprises:

capturing, from sample images of a sample image set, the original image comprising the to-be-replaced face; and

aligning a plurality of captured original images based on feature points of the to-be-replaced face.

6. The method according to claim 1 , wherein before the updating the parameters of the encoder, the method further comprises:

performing random augmentation on the original image;

performing random distortion on the original image on which the random augmentation has been performed, to obtain a distorted original image;

capturing a facial image of a to-be-replaced face in the distorted original image; and

scaling the captured facial image according to an input resolution supported by the encoder during the updating of the parameters of the encoder.

7. The method according to claim 6 , wherein the performing the random augmentation comprises:

performing random rotation based on a center of the original image;

performing random scaling on the original image obtained after the random rotation; and

performing random translation on the original image obtained after the random scaling.

8. The method according to claim 6 , wherein the performing the random distortion comprises:

inserting noise into the original image on which the random augmentation has been performed, the noise comprising at least one of fixed noise and dynamic noise.

9. The method according to claim 8 , wherein the inserting the noise into the original image comprises:

generating a grid on the original image on which the random augmentation has been performed; and

adding noise corresponding to coordinates of at least one node in the grid.

10. A non-transitory computer-readable storage medium storing instructions which when executed by a processor cause the processor to perform the training method according to claim 1 .

11. An image processing method for an image processing model, the method comprising:

encoding, by an encoder of the image processing model, an original image to obtain a visual feature of a to-be-replaced face in the original image; and

performing, by circuitry of a decoder of the image processing model, decoding based on the visual feature of the to-be-replaced face to obtain a target face having the same visual feature as the to-be-replaced face, wherein

parameters of the decoder are obtained by performing an update according to a plurality of target images comprising the target face without changing parameters of the encoder, and

the decoder is trained with the target face in the plurality of target images and the encoder is not trained with the target face in the plurality of target images.

12. A training apparatus for an image processing model, comprising:

circuitry configured to

update parameters of an encoder in the image processing model according to a to-be-replaced face in an original image, to configure the encoder to encode the to-be-replaced face to obtain a visual feature of the to-be-replaced face;

update parameters of a decoder in the image processing model according to the to-be-replaced face in the original image, to configure the decoder to perform decoding based on the visual feature of the to-be-replaced face; and

update the parameters of the decoder according to a target face in a plurality of target images without changing the parameters of the encoder, to configure the decoder to perform decoding based on the visual feature of the to-be-replaced face and obtain the target face having the same visual feature as the to-be-replaced face,

wherein the decoder is trained with the target face in the plurality of target images and the encoder is not trained with the target face in the plurality of target images.

13. The apparatus according to claim 12 , wherein the circuitry is configured to:

substitute a facial image of a to-be-replaced face in a distorted original image of the original image into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determine the parameters corresponding to the encoder based on the loss function satisfying a convergence condition.

14. The apparatus according to claim 12 , wherein the circuitry is configured to:

substitute a facial image of a to-be-replaced face in a distorted original image of the original image into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determine the parameters of the decoder based on the loss function satisfying a convergence condition.

15. The apparatus according to claim 12 , wherein the circuitry is configured to:

substitute a facial image of the target face in a distorted target image of the plurality of target images into a loss function corresponding to a self-encoding network formed by the encoder and the decoder; and

determine, without changing the parameters of the encoder in the loss function, the parameters of the decoder based on the loss function satisfying a convergence condition.

16. The apparatus according to claim 12 , wherein the circuitry is configured to:

capture, from sample images of a sample image set, the original image comprising the to-be-replaced face; and

align plurality of captured original images based on feature points of the to-be-replaced face.

17. The apparatus according to claim 12 , wherein the circuitry is configured to:

perform random augmentation on the original image;

perform random distortion on the original image on which the random augmentation has been performed, to obtain a distorted original image;

capture a facial image of a to-be-replaced face in the distorted original image; and

scale the captured facial image according to an input resolution supported by the encoder during the updating of the parameters of the encoder.

18. The apparatus according to claim 17 , wherein the circuitry is configured to:

perform random rotation based on a center of the original image;

perform random scaling on the original image obtained after the random rotation; and

perform random translation on the original image obtained after the random scaling.

19. The apparatus according to claim 17 , wherein the circuitry is configured to:

insert noise into the original image on which the random augmentation has been performed, the noise comprising at least one of fixed noise and dynamic noise.

20. The apparatus according to claim 19 , wherein the circuitry is configured to:

generate a grid on the original image on which the random augmentation has been performed; and

add noise corresponding to coordinates of at least one node in the grid.