Feature Fusion from Multiple Paintings for Generalized Artistic Style Transfer
6 Pages Posted: 22 Aug 2019
Date Written: March 14, 2019
The paper presents a deep CNN based artistic style transfer approach which automatically superimposes the artistic style on any input photograph. The generated image has a similar visual appearance to that of the artist’s painting(s). Existing approaches to artistic style transfer use a single painting of an artist to generate his/her painting style. We propose a more general framework in which a generalized style feature is extracted by fusing style feature from multiple paintings of an artist. A pre-trained VGG model is employed which extracts the generalized style feature from the artist’s paintings and the content feature from the input photograph. In this work, we study the different techniques for style feature fusion and present a new technique to evaluate the potency of any style transfer technique. Experimental evaluation indicates that the proposed style feature fusion technique captures the artistic style at a higher resolution than existing approaches which uses only a single painting from each artist for style transfer.
Keywords: Neural Style Transfer, Feature Fusion, VGG Network, L-BFGS Optimizer
Suggested Citation: Suggested Citation