Train a neural network to learn multiple style-transfer.
The network architecture is the same with feed forward neural style transfer with two modifications:
- Change zero-padding to mirror-padding.
- Change transposed convolution to up-sampling with convolution.
Main Idea: To model a style, it is sufficient to specialize scaling and shifting parameters after normalization to each specific style.
Instance Normalization: Instance normalization normalizes feature maps after each convolution layer as: Xnorm = (X - μ) / σ. Where X is the feature map, μ, σ are mean and covariance of X.
Conditional Instance Normalization: Introduce two NxC matrixes parameters: γ, β (N is the number of styles, C is the channel of feature map X) to normalize the feature map X as: Xnorm = γs((X - μ) / σ) + βs. γs and βs are the sth row of γ and β, specialized to scaling and shifting feature map X to match the sth style. Show as below:
Dumoulin, Vincent, Jonathon Shlens, and Manjunath Kudlur. "A learned representation for artistic style." (2017).