Artistic Style TransferNicholas Tan, Elias Wang

Department of Electrical Engineering, Stanford University

Motivation Algorithm Overview

Related Work

Experimental ResultsUsing a biologically inspiredvision model, called “DeepNeural Networks,” anartificial system can betrained to learn differentstyles of painting andsubsequently recreateimages rendered in thatstyle.

The model is successfulbecause it is able todistinguish between “style”vs. “content”, and mergethe two.

Using a modification ofTexture-Synthesis it ispossible to get fast andcomparable style-transferresults. This method focuseson separating information-poor areas to hallucinate thestyle and high content areaswhere it keeps the originalimage

Content Image Style Image

Color transfer Add noise

Resize Estimate(finer scale)

Input:

SegmentationInitialization:

Loop over scales:Loop over patch sizes:

Patch Matching Style Synthesis

Content Fusion

Color Transfer

Denoise

Resize Patch

Hallucination

Chosen parameters: 3 scales, 2 patch sizes (except smallest scale uses 3 patch sizes), 10% NN noise, 25%+75% hallucination averaging

0% 10% 25% 50%

Hallucination percentage influence :

Content

Style

Source: L.A. Gatys, A.S. Ecker, and M. Bethge, Image Style Transfer Using Convolutional Neural Networks, CVPR, 2016.

Source: M. Elad, P. Milanfar, Style-Transfer via Texture-Synthesis, Google Research, September 21, 2016

Source: V. Kwatra, I. Essa, A. Bobick, and N. Kwatra, Texture Optimization for Example-Based Synthesis,ACM ToG, Vol. 24, No. 3, pp. 795-802, 2005

Style Fusion