Method and device for image synthesis
Computer-implemented method for transferring style features from at least one source image to a target image, comprising the steps of generating a result image, based on the source and the target image, wherein one or more spatially-variant features of the result image correspond to one or more spatially variant features of the target image; and wherein a texture of the result image corresponds to a texture of the source image; and outputting the result image, and a corresponding device. According to the invention, the texture corresponds to a summary statistic of spatially variant features of the source image.
1. A computer-implemented method for transferring style features from at least one source image to a target image, the method comprising:
(A) generating a result image, based on the source image and the target image, by constructing a style representation of the source image based on filter responses of a convolutional neural network to the source image in a number of layers of the network, and by calculating a correlation between the different filter responses,
wherein an expectation is taken over the spatial extent of the source image, and
wherein semantic content of the result image corresponds to semantic content of the target image, and
wherein a texture of the result image corresponds to a texture of the source image; and
(B) outputting the result image.
2. The method of claim 1 , wherein the correlation corresponds to a Gram matrix.
3. The method of claim 1 , wherein the semantic content of an image corresponds to a result of a non-linear transformation of that image.
4. The method of claim 3 , wherein the non-linear transformation corresponds to one or more convolutions of the image.
5. The method according to claim 1 , characterized in that the result image is made available in a social network.
6. The method according to claim 1 , wherein the target image is received from a user or wherein the result image is sent to a user over a telecommunications network.
7. A non-transitory computer program product comprising software comprising instructions for performing a method according to claim 1 on a computer.
8. A device for transferring style features from at least one source image to a target image, the device comprising:
(A) a generating section for generating a result image based on the target image and the source image,
by constructing a style representation of the source image based on filter responses of a convolutional neural network to the source image in a number of layers of the network, by calculating a correlation between the different filter responses,
wherein an expectation is taken over the spatial extent of the source image,
wherein semantic content of the result image corresponds to semantic content of the target image, and
wherein a texture of the result image corresponds to a texture of the source image; and
(B) an output unit for outputting the result image.
9. The device of claim 8 , further comprising a digital camera for capturing one or more source images.