Deep Neural Representation Based Face Sketch Synthesis

Sketch is defined as a kind of drawing using purely line stroke without any color.It is an ancient artistic form and be carried forward by famous artists such as Da Vinci and Michelangelo during the Renaissance.Famous sketch drawings include ”Self Portrait” by Da Vinci and ”Liberia Sibyl”.Face sketch synthesis shows great applications in a lot of fields such criminal investigation.In the process of criminal investigation,public security organization needs to draw a sketch of suspects according to the description of witnesses.If a sketch dataset is built the suspect could be quickly identified by matching those sketches with dataset.In addition to this,more and more people wants to show their personality in social network and a sketch head portrait would be quite good choice.This thesis aims to synthesize a face sketch given a face photo.Existing face sketch synthesis methods learn the patch-wise sketch style from the training dataset containing photo-sketch pairs.These methods manipulate the whole process directly in the field of RGB space,which unavoidably result in unsmooth noise at patch boundaries.If denoising methods are used,the sketch edges would be blurred and face structures couldn't be restored.Recent researches of feature maps,which are the outputs of a certain neural network layer,have achieved great success in texture synthesis and artistic image generation.The main contribution of this thesis lies in that,for the first time solving style transfer task using neural network feature map synthesis.This thesis propose a neural network feature map based face sketch synthesis approach,reformulate the face sketch synthesis problem into a neural network feature map based optimization task.Our results accurately capture sketch drawing style and make full use of the whole stylistic information hidden in training dataset.Unlike former feature map based methods,we utilize the Enhanced 3D PatchMatch and cross-layer cost aggregation algorithms to obtain the target feature maps for final results.Experiments show that our results can precisely restore face structure and preserve the hand-drawn sketch style.In the experiment design part,we design qualitative experiments and quantitative experiments to evaluate our results.In qualitative evaluation part our results are clearer and more natural than results by previous methods in CUFS and AR dataset.In quantitative evaluation our results are superior to previous methods in SSIM and FSIM metric.Our results also get higher scores in the user study part.