Deep feature generative adversarial neural networks
A mobile device can implement a neural network-based domain transfer scheme to modify an image in a first domain appearance to a second domain appearance. The domain transfer scheme can be configured to detect an object in the image, apply an effect to the image, and blend the image using color space adjustments and blending schemes to generate a realistic result image. The domain transfer scheme can further be configured to efficiently execute on the constrained device by removing operational layers based on resources available on the mobile device.
1. A method comprising:
identifying an image on a device;
generating, by a processor of the device, a modified image by applying an adversarial neural network to the image, the adversarial neural network comprising an adversarial transformation network within a generator network, and the generator network further comprising an encoder network and a decoder network, wherein the adversarial neural network is trained in stages comprising a first stage, in which at least one of the encoder network or the decoder network are trained, and a second stage, in which the adversarial transformation network is trained, the adversarial transformation network not being trained in the first stage; and
publishing, by the device, the modified image to a network site as an ephemeral message.
2. The method of claim 1 , further comprising:
identifying an instruction to apply an image effect to the image; and
storing the modified image.
3. The method of claim 2 , wherein the adversarial transformation network is between the encoder network and the decoder network.
4. The method of claim 3 , wherein the adversarial transformation network is adversarially trained using a discrimination network that receives output data from the decoder network.
5. The method of claim 3 , wherein
the adversarial transformation network is trained using a discrimination network.
6. The method of claim 5 , wherein in the first stage the encoder network and the decoder network are jointly trained using feature loss.
7. The method of claim 5 , wherein the encoder network and the decoder network are not trained in the second stage.
8. The method of claim 5 , wherein the adversarial transformation network is trained using Generative Adversarial Loss (GAN) loss.
9. The method of claim 2 , further comprising:
receiving, from a server, the adversarial transformation network pre-trained to apply the image effect.
10. The method of claim 2 , further comprising:
receiving, from an input interface of the device, the instruction to apply the image effect on the image.
11. The method of claim 10 , further comprising:
in response to receiving the instruction, selecting, from a plurality of neural networks stored on the device, the adversarial neural network based on the adversarial neural network being trained for the image effect specified by the instruction, wherein each of the plurality of neural networks is trained for different image effects using adversarial transformation networks within respective generator networks.
12. The method of claim 1 , wherein the adversarial transformation network is trained to transfer an image between different domains.
13. The method of claim 12 , wherein the domains includes one or more of the following: an old person domain, a young person domain, a masculine person domain, a feminine person domain, a painted domain, a photorealistic domain.
14. The method of claim 1 , wherein the adversarial neural network is a convolutional neural network.
15. The method of claim 1 , further comprising:
generating the image using an image sensor of the device.
16. The method of claim 1 , wherein, in the second stage, the adversarial transformation network is trained to apply a transformation effect to input data.
17. A device comprising:
one or more processors;
an image sensor; and
a memory storing instructions that, when executed by the one or more processors, cause the device to perform operations comprising:
identifying an image on the device;
generating, by a processor of the device, a modified image by applying an adversarial neural network to the image, the adversarial neural network comprising an adversarial transformation network within a generator network, and the generator network further comprising an encoder network and a decoder network, wherein the adversarial neural network is trained in stages comprising a first stage, in which at least one of the encoder network or the decoder network are trained, and a second stage, in which the adversarial transformation network is trained, the adversarial transformation network not being trained in the first stage; and
publishing, by the device, the modified image to a network site as an ephemeral message.
18. The device of claim 17 , wherein the operations further comprise:
identifying an instruction to apply an image effect to the image; and
storing the modified image.
19. The device of claim 18 , wherein the adversarial transformation network is between the encoder network and the decoder network.
20. A non-transitory machine-readable storage device embodying instructions that, when executed by a device, cause the device to perform operations comprising:
identifying an image on the device;
generating, by a processor of the device, a modified image by applying an adversarial neural network to the image, the adversarial neural network comprising an adversarial transformation network within a generator network, and the generator network further comprising an encoder network and a decoder network, wherein the adversarial neural network is trained in stages comprising a first stage, in which at least one of the encoder network or the decoder network are trained, and a second stage, in which the adversarial transformation network is trained, the adversarial transformation network not being trained in the first stage; and
publishing, by the device, the modified image to a network site as an ephemeral message.