Harmonizing composite images utilizing a semantic-guided transformer neural network
The present disclosure relates to systems, non-transitory computer-readable media, and methods that implement a multi-branch harmonization neural network architecture to harmonize composite images. For example, in one or more implementations, the semantic-guided transformer-based harmonization system uses a convolutional branch, a transformer branch, and a semantic branch to generate a harmonized composite image based on an input composite image and a corresponding segmentation mask. More particularly, the convolutional branch comprises a series of convolutional neural network layers followed by a style normalization layer to extract localized information from the input composite image. Further, the transformer branch comprises a series of transformer neural network layers to extract global information based on different resolutions of the input composite image. The semantic branch includes a visual neural network that generates semantic features that inform the harmonization of the composite images.
1. A computer-implemented method comprising:
receiving a composite image and a segmentation mask of a foreground object portrayed against a background of the composite image;
utilizing a global neural network branch to extract global information of the background;
utilizing a semantic neural network branch to extract semantic information from the composite image; and
generating a harmonized composite image comprising the foreground object harmonized with the background by decoding the global information and the semantic information.
2. The computer-implemented method of claim 1 , wherein utilizing the global neural network branch to extract the global information of the background comprises extracting the global information utilizing a transformer neural network.
3. The computer-implemented method of claim 1 , wherein utilizing the semantic neural network branch to extract semantic information from the composite image comprises extracting semantic features utilizing a visual neural network.
4. The computer-implemented method of claim 1 , further comprising utilizing a local neural network branch to extract local information from the composite image.
5. The computer-implemented method of claim 4 , wherein generating the harmonized composite image comprises decoding the local information, the global information, and the semantic information.
6. The computer-implemented method of claim 5 , wherein utilizing the local neural network branch to extract local information comprises extracting local information of a background portion adjacent to the foreground object utilizing a convolutional neural network.
7. A non-transitory computer-readable medium comprising instructions that, when executed by at least one processor, cause the at least one processor to perform operations comprising:
receiving a composite image and a segmentation mask of a foreground object portrayed against a background of the composite image;
utilizing a local neural network branch to extract local information from the composite image;
utilizing a global neural network branch to extract global information of the background;
utilizing a semantic neural network branch to extract semantic information from the composite image; and
generating a harmonized composite image comprising the foreground object harmonized with the background by decoding the local information, global information, and the semantic information.
8. The non-transitory computer-readable medium of claim 7 , wherein the operations further comprise:
generating a semantic aware image crop from the composite image;
generating a semantic aware segmentation mask crop from the segmentation mask; and
concatenating the semantic aware image crop and the semantic aware segmentation mask crop to generate a cropped four-channel semantic-aware input.
9. The non-transitory computer-readable medium of claim 8 , wherein utilizing the semantic neural network branch comprises extracting semantic features from the cropped four-channel semantic-aware input utilizing a visual neural network.
10. The non-transitory computer-readable medium of claim 9 , wherein extracting semantic features from the cropped four-channel semantic-aware input utilizing the visual neural network comprises extracting the semantic features from a convolutional layer just prior to average pooling and a classification layer of the visual neural network.
11. The non-transitory computer-readable medium of claim 9 , wherein utilizing the global neural network branch comprises utilizing one or more transformer neural network layers to extract the global information of the background from the cropped four-channel semantic-aware input.
12. The non-transitory computer-readable medium of claim 11 , wherein utilizing the one or more transformer neural network layers to extract the global information of the background comprises utilizing one or more layers comprising a self-attention neural network layer, a mix feed-forward neural network layer, and an overlap patch merging layer.
13. The non-transitory computer-readable medium of claim 9 , wherein utilizing the local neural network branch comprises utilizing one or more convolutional neural network layers to extract the local information of a background portion adjacent to the foreground object from the cropped four-channel semantic-aware input.
14. The non-transitory computer-readable medium of claim 13 , wherein utilizing the local neural network branch comprises utilizing one or more style normalization layers to extract style information from the background portion adjacent to the foreground object by determining a mean and standard deviation of pixel color values between pixels corresponding to the background portion adjacent to the foreground object.
15. The non-transitory computer-readable medium of claim 14 , wherein the operations further comprise utilizing the one or more style normalization layers to apply the style information to the foreground object.
16. The non-transitory computer-readable medium of claim 7 , wherein receiving the composite image and the segmentation mask comprises:
identifying, via a graphical user interface, a user selection of a background image for combining with the foreground object as the composite image;
identifying, via the graphical user interface, an additional user selection corresponding to a harmonization filter for harmonizing the foreground object and the background image; and
utilizing a segmentation neural network to generate the segmentation mask of the foreground object portrayed against the background of the composite image.
17. A system comprising:
a composite image comprising a foreground object portrayed against a background of the composite image;
a segmentation mask of the foreground object;
a local neural network branch to extract local information of a background portion adjacent to the foreground object;
a global neural network branch to extract global information of the background;
a semantic neural network branch to extract semantic information from the composite image;
a decoder to decode the local information, the global information, and the semantic information; and
one or more processors configured to cause the system to generate a harmonized composite image comprising the foreground object with modified pixel color values based on decoding of the local information, the global information, and the semantic information.
18. The system of claim 17 , wherein:
the local neural network branch comprises one or more convolutional neural network layers;
the global neural network branch comprises a transformer neural network; and
the semantic neural network branch comprises a visual neural network.
19. The system of claim 18 , wherein the global neural network branch comprises a series of transformer neural networks composed of at least one of a self-attention neural network layer, a mix feed-forward neural network layer, or an overlap patch merging layer.
20. The system of claim 18 , wherein the local neural network branch comprises a style normalization layer to generate style-normalized foreground feature vectors for the foreground object based on style information for the background portion adjacent to the foreground object.