A Neural Algorithm of Artistic Style - Gatys et al. - CVPR 2016

Info

• Title: A Neural Algorithm of Artistic Style
• Task: Style Transfer
• Author: L. A. Gatys, A. S. Ecker, and M. Bethge
• Arxiv: 1508.06576
• Date: Aug. 2015
• Published: CVPR 2016

Motivation & Design

To obtain a representation of the style of an input image, we use a feature space originally designed to capture texture information.

Key Finding the representations of content and style in the Convolutional Neural Network are separable.

Network Architecture

Content Match

To visualise the image information that is encoded at different layers of the hierarchy (Fig 1, content reconstructions) we perform gradient descent on a white noise image to find another image that matches the feature responses of the original image.

$$L\_\text{content}(\overrightarrow{p},\overrightarrow{x},l)=\frac{1}{2}\sum _{i,j}{(F_{ij}^l-P_{ij}^l)}^2$$

Style Match

These feature correlations are given by the Gram matrix $G^l\in R_N^l\times N^l$, where $G_{ij}^l$ is the inner product between the vectorised feature map of $i$ and $j$ in layer $l$:

$G_{ij}^l={\sum }_k{F}_{ik}^l{F}_{jk}^l$ To generate a texture that matches the style of a given image, we use gradient descent from a white noise image to find another image that matches the style representation of the original image. This is done by minimising the mean-squared distance between the entries of the Gram matrix from the original image and the Gram matrix of the image to be generated.

$$E_l=\frac{1}{4N_l^2{M}_l^2}\sum _{i,j}{(G_{ij}^l-A_{ij}^l)}^2$$ $$L\_style(\overrightarrow{a},\overrightarrow{x})=\sum _{l=0}^L{w}_l{E}_l$$

Total Loss

$$L\_\text{total}(\overrightarrow{p},\overrightarrow{a},\overrightarrow{x})=\alpha L\_\text{content}(\overrightarrow{p},\overrightarrow{x})+\beta L_{\text{style}}(\overrightarrow{a},\overrightarrow{x})$$

Code

• TensorFlow
• PyTorch