Research On Stylized Facial Animation Generation Method Based On Sample Learning

Stylized rendering technology derived from the historical painting art is very important to stylized animation,games and cultural heritage.The stylized animation,which has a close connection with stylized rendering,is more expressive than static stylized rendering since it has an additional dimension of time,which has attracted lots of attention in the research area of computer graphics.Example-based facial stylized-animation is one of the important research topics in this area,and is a key to the generation of stylized character animation.However,in order to learn a variety of styles from examples and automatically generate stylized facial animation,many problems remain to be solved.In this thesis,we combine technologies from computer graphics,computer vision,and machine learning to propose novel example-based methods for rendering and animation generation,which employ case-based reasoning(CBR)framework.The proposed methods can improve the production efficiency of stylized facial animation with different styles.More concretely,our work contains:1.We propose methods for ca.se representation and candidate case retrieval for stylized facial animation generation.Case representation is the basis for designing a CBR method.In order to mimic the artistic processing of stylized rendering,we use paired facial photo and stylized portrait as an example of stylized rendering,from which stylized rendering case is created.In inbetweening process,2D keyframes lack 3D information,making it difficult to handle occlusion.To solve this problem,we introduce a method which infers 3D information from master frames.Candidate case retrieval requires a method to evaluate fitness of each case.As individual artist has unique evaluation method,we propose a learning method for fitness evaluation.Case representation and candidate case retrieval methods play important roles in following methods.2.We introduce an inbetweening method by learning from master frames.The automation of inbetweening is a key problem in traditional keyframing technology.The main challenge is adding 3D information to 2D keyframes,with as-least-as-possible user interaction.We propose a method which can specify correspondences between keyframes automatically by learning from master frames,which are drawn in an early stage of traditional animation production pipeline.We further extend view-morphing method to interpolate keyframes at four different views,generating inbetweens with correct 3D view change effect and hand-drawn style.3.We present a stylized rendering method employing CBR framework.The variety of rendering style is a challenge of stylized rendering.In order to learn artist’ s unique style from minimum examples,we adopt CBR technology and propose a novel facial stylized rendering method.In this method,we use generation-and-evaluate mechanism to predict fitness of each candidate case and retrieve the best-fit case.We use learning-based parameter estimation model to achieve automatic case adaptation.Our method is able to learn several different facial portrait rendering styles.4.We design a character animation generation method blends 2D and 3D assets.Animation is popular as a kind of art.Dunhuang murals are one of the most important cultural heritage of the Chinese nation.Therefore,To use animation as a carrier for presenting Dunhuang mural is very conducive to attract public attentions,which is significant to the protection and inheritance of such cultural heritage.Animating mural requires the preservation of 2D hand-drawn styles.Although our example-based stylized rendering and main-frame-based inbetweening techniques can mimic artist’ s individual styles in generated stylized animation,current 2D animation assets and talents are relatively more rarer than 3D animation.Meanwhil,2D and 3D animation techniques have their own advantages.Therefore,we present a new character animation generation method which blends 2D and 3D assets.This method use 3D assets to animate 2D assets,preserving 2D hand-drawn style in resultant animation and allowing user smoothly adjust the style between 2D and 3D.We have studied some key problems of example-based stylized facial animation techniques and proposed several new effective methods,which enriches the research in the field and advances the cost reducing of stylized animation production.