基于学习的人脸表情动画生成方法研究

发布时间：2018-03-18 12:30

本文选题：主动表观模型　切入点：人脸卡通化　出处：《电子科技大学》2013年硕士论文　论文类型：学位论文

【摘要】：人脸的多样性和独特性一直是计算机视觉、图形学和模式识别等领域研究的热点问题；卡通化艺术以其特有的表现形式和手法引领着艺术发展的新潮流；两者相结合的人脸卡通化通过线条的描绘和颜色的渲染，以其特有的夸张模式，形象和逼真的再现人脸图像，在网络游戏、互动论坛、社交软件以及动漫等领域应用广泛。现有的基于学习的人脸卡通化方法通常将图像分成若干小块，，并通过对图像块的匹配及合成，来实现卡通图像的合成。然而受到块效应的影响，图像中的人脸特征细节描述不详细，人脸线条效果欠佳。而目前流行的人脸定位方法，通过对人脸特征的准确定位，能够得到更好的细节描述及线条效果。本文主要研究对于一幅给定的人脸图像计算机如何自动地生成具艺术家绘画作品特定风格的卡通人脸图像，以及对卡通人脸进行表情动画变换。主要的研究内容如下：一、提出了两种基于学习的卡通人脸图像生成算法：基于参数模型的卡通化方法和基于特征点的卡通化方法。基于参数模型的卡通化方法是在主动表观模型的基础上，通过学习人脸匹配的过程，利用参数估计的方法生成卡通人脸图像；基于特征点的卡通化方法则从高层语义学的角度出发，将人脸特征分类处理，通过分别合成卡通人脸头发、轮廓和五官，实现卡通人脸图像的生成。二、为了得到较好的彩色卡通效果，采用三种颜色渲染方法对卡通图像上色，分别为基于颜色空间转换的方法和基于图像分割的方法。基于颜色空间转换的着色方法得到的卡通图像色调与输入图像相似，而基于图像分割的着色方法得到卡通图像与艺术家绘画的色彩风格更相近。三、在生成的卡通图像的基础上进行面部表情的变换。利用图像变形算法，通过控制主要面部器官特征点的位置，使无表情的卡通人脸变换出微笑、悲伤等形象、生动和俏皮的表情。实验证明，本文提出的人脸卡通化方法能够获得较理想的卡通效果，并且卡通人脸的表情动画生动、形象，时效性高。
[Abstract]:The diversity and uniqueness of human face has always been a hot topic in computer vision, graphics and pattern recognition, and the art of Katonghua leads the development of art with its unique forms and techniques. The combination of the two face cards through the description of lines and color rendering, with its unique exaggeration mode, image and lifelike reproduction of face images, in online games, interactive forums, social software and animation and other fields widely used. The existing learning-based face card generalization methods usually divide the image into several small blocks and realize the cartoon image synthesis by matching and synthesizing the image blocks. However, due to the influence of block effect, The detail description of face features in image is not detailed, and the effect of face line is not good. However, the popular face localization method can get better detail description and line effect through accurate location of face features. This paper mainly studies how a given face image computer can automatically generate a cartoon face image with a specific style of the artist's painting work, and how to make a cartoon facial expression animation transformation. The main research contents are as follows:. Firstly, two learning based cartoon face image generation algorithms are proposed: one is based on parametric model and the other is based on feature point, and the other is based on active apparent model. By learning the process of face matching, we use the method of parameter estimation to generate cartoon face image. Based on feature point, we classify the face features from the perspective of high-level semantics, and synthesize the hair of cartoon face separately. Contour and facial features to achieve cartoon face image generation. Secondly, in order to get a better color cartoon effect, three color rendering methods are used to color the cartoon image. The color space conversion method and the image segmentation method are used respectively. The color color method based on color space conversion is similar to the input image. The coloring method based on image segmentation shows that the color style of cartoon image is more similar to that of artist painting. Third, on the basis of the generated cartoon image, the facial expression is transformed. By controlling the location of the feature points of the main facial organs by using the image deformation algorithm, the expressionless cartoon faces are transformed into images of smile, sadness, etc. A vivid and playful expression. Experimental results show that the proposed method can obtain ideal cartoon effect, and the facial expression of cartoon face is vivid, vivid and time-efficient.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】