Protecting image features in stylized representations of a source image
Systems and methods herein describe an image stylization system. The image stylization system accesses a set of images corresponding to a target domain style, generates a set of paired images using a first machine learning model, analyze the generated set of paired images using a second machine learning model trained to analyze the generated set of paired images based on a plurality of protected feature criteria, determines a set of image transformations for the generated set of pairs, generates a transformed set of paired images by performing the set of image transformations on the set of paired images, and generates stylized images corresponding to the target domain style using a supervised image translation model trained on the transformed set of paired images.
1. A method comprising:
generating a set of paired images using a first machine learning model, the set of paired images comprising a first image corresponding to a source image and a second image corresponding to a target domain style;
analyzing the generated set of paired images using a second machine learning model trained to analyze the generated set of paired images based on a plurality of protected feature criteria;
based on the analyzing, determining a set of image transformations for the generated set of paired images;
generating a transformed set of paired images, the generating comprising performing the set of image transformations on the generated set of paired images; and
generating one or more stylized images corresponding to the target domain style, the one or more stylized images generated using a supervised image translation model trained on the transformed set of paired images.
2. The method of claim 1 , wherein the source image comprises a human face.
3. The method of claim 1 , wherein the set of image transformations comprises: image processing and image-to-image translation.
4. The method of claim 1 , wherein each protected feature criterion of the plurality of protected feature criteria is an unbiased classifier for a protected feature.
5. The method of claim 4 , wherein the protected feature is a feature of the first image that is preserved during transformation of the first image to the second image.
6. The method of claim 1 , wherein the target domain style is an anime image style.
7. The method of claim 1 , wherein analyzing the generated set of paired images further comprises:
accessing a set of images corresponding to the target domain style, the accessed set of images different from the set of paired images; and
generating an adjusted set of images, the generating comprising, augmenting the accessed set of images corresponding to the target domain style with a set of new images corresponding to the target domain style.
8. The method of claim 7 , wherein generating the adjusted set of images further comprises:
removing paired images from the accessed set of images, wherein the removed paired images fail to meet at least one protected feature criteria of the plurality of protected feature criteria.
9. The method of claim 7 , wherein the first machine learning model is retrained on the adjusted set of images.
10. A system comprising:
a processor; and
a memory storing instructions that, when executed by the processor, configure the system to perform operations comprising:
generating a set of paired images using a first machine learning model, the set of paired images comprising a first image corresponding to a source image and a second image corresponding to a target domain style;
analyzing the generated set of paired images using a second machine learning model trained to analyze the generated set of paired images based on a plurality of protected feature criteria;
based on the analyzing, determining a set of image transformations for the generated set of paired images;
generating a transformed set of paired images, the generating comprising performing the set of image transformations on the generated set of paired images; and
generating one or more stylized images corresponding to the target domain style, the one or more stylized images generated using a supervised image translation model trained on the transformed set of paired images.
11. The system of claim 10 , wherein the first image is a source image comprising a human face.
12. The system of claim 10 , wherein the set of image transformations comprises: image processing and image-to-image translation.
13. The system of claim 10 , wherein each protected feature criterion of the plurality of protected feature criteria is an unbiased classifier for a protected feature.
14. The system of claim 13 , wherein the protected feature is a feature of the first image that is preserved during transformation of the first image to the second image.
15. The system of claim 13 , wherein the protected feature is provided as input to the first machine learning model.
16. The system of claim 10 , wherein the target domain style is an anime image style.
17. The system of claim 10 , wherein analyzing the generated set of paired images further comprises:
accessing a set of images corresponding to the target domain style, the accessed set of images different from the set of paired images; and
generating an adjusted set of images, the generating comprising, augmenting the accessed set of images corresponding to the target domain style with a set of new images corresponding to the target domain style.
18. The system of claim 17 , wherein generating the adjusted set of paired images further comprises:
removing paired images from the accessed set of images, wherein the removed paired images fail to meet at least one protected feature criteria of the plurality of protected feature criteria.
19. The system of claim 17 , wherein the first machine learning model is retrained on the adjusted set of images.
20. A non-transitory processor-readable storage medium storing processor executable instructions that, when executed by one or more processors of a machine, cause the machine to perform operations comprising:
generating a set of paired images using a first machine learning model, the set of paired images comprising a first image corresponding to a source image and a second image corresponding to a target domain style;
analyzing the generated set of paired images using a second machine learning model trained to analyze the generated set of paired images based on a plurality of protected feature criteria;
based on the analyzing, determining a set of image transformations for the generated set of paired images;
generating a transformed set of paired images, the generating comprising performing the set of image transformations on the generated set of paired images; and
generating one or more stylized images corresponding to the target domain style, the one or more stylized images generated using a supervised image translation model trained on the transformed set of paired images.