The Algorithm Of Automatic Face Caricatures Synthesis And It's Application

With the development of the multimedia technology and popularization of the SNS,people often change their portraits into caricatures for the purpose of entertainment and social requirement.However there are still some problems exist in the related work of face caricature automatic synthesis: 1.the face caricature synthesis is not completely automatically;2.the style of the synthesized face caricature is limited;3.the exaggeration of face is not stable;4.the computation speed is too slow.To solve these problems,this thesis focuses on the automatic face caricature synthesis algorithm on this thesis and divide the work into two parts---face style transfer and face shape exaggeration.The contributions of the thesis mainly include:Face style transfer.This thesis proposes a face style transfer algorithm based on multiscale style transfer and semantic segmentation.A multi-scale encoder-decoder network is designed and the transferred feature is extracted through reassembling the matching result of the style feature and content feature via patch-based matching in the feature space.The reassembling feature is both statistically aligned the style feature and spatial structure of the content feature.Then a multi-scale style fusion is applied to finally synthesize the face caricature in the decoder network.The multi-scale fusion helps strengthen the edge and color information of the result.For the specified image like face portrait,the face portrait is first transferred with the segmented style image as reference.Then the corresponding transferred face regions is segmented and reassembled.The experiment results show that our method has better performance both on effect and speed,which proves the validity of our algorithm.Face shape exaggeration.This thesis proposes a face shape exaggeration method based on the statistical information of face feature.A rule is designed to exaggerate the face parts.Then we analyze the difference between each part's of face shape and mean face shape,and calculate the significance of each face parts.By this way,the whole face shape is exaggerated through the different exaggeration level of each part.Our method can learn the artistic exaggeration without a big face caricature dataset.The thesis shows the experiment result of face exaggerated shape and compares the final exaggerated face caricature with other works,which show our result is better than others.Finally a face caricature automatic synthesis system is designed according to our work on face style transfer and face shape exaggeration.As show in the thesis,our software system has beautiful UI and is easy to use.