SEMI OpenIR  > 高速电路与神经网络实验室
基于人脸图像的机器肖像艺术技术研究
董肖莉
学位类型博士
导师李卫军
2018-05-23
学位授予单位中国科学院大学
学位授予地点北京
学位专业电路与系统
关键词线条增强 三角形坐标系 个性特征夸张 可辨识表情 轨迹规划
其他摘要

蓬勃发展的人工智能技术不断刷新着人类对未知世界和极限领域的认知,改变着人类的生活、生产方式,并随之涌现出各式各样的智能机器人,如智能工业制造机器人、智能护理机器人以及救援机器人等,这些机器人在人类的生产和生活中发挥了不可替代的作用。通过机器人来绘制人脸肖像画也逐渐成为人工智能的一个重要研究方向,代表着人工智能在艺术世界的初步探索。肖像绘制机器人的出现,使得对于艺术家简单而普通人无法完成的肖像绘制工作变得简单、高效与生动有趣。

本文以人脸肖像为研究对象,首先研究了可感知人脸特征的线条肖像生成方法,并在此基础上实现了肖像漫画、表情肖像以及表情肖像漫画等多种艺术肖像生成,最后通过规划机器人肖像绘图的轨迹,实现了机器人的拟人肖像绘制,完成可以“感知美”、“再现美”、“加工美”、“刻画美”的艺术肖像绘制机器人系统。该系统的研究具有重要的学术意义和科学意义,其相关研究成果可实现在多个领域的广泛应用。本文从以下几个方面进行了研究,并取得了创新性成果:

(1)针对低分辨率以及清晰度不高的人脸图像所生成的肖像中线条不够突出、不利于真实反映人脸特征以及不利于表达肖像的可辨识性等问题,首先分析了人类视觉特性中的视网膜神经元感受野的非经典侧抑制原理,并在此研究基础上提出了一种基于三高斯模型的线条肖像生成方法。实验结果表明,所提出的方法不仅能够增强肖像的线条效果,很好地抑制噪声,同时还具有保持原始图像的结构与方向信息等优良特性。

(2)针对现有的肖像漫画生成方法中所存在的依赖训练样本、算法复杂度高、夸张形式与夸张程度单一等问题,首先研究了基于三角形坐标系的图像拓扑变形方法,利用该方法计算简单、变形灵活等优势,设计了一种基于三角形坐标系的夸张程度可参数调节的肖像漫画生成方法。该方法通过将实验人脸与参考人脸特征进行比较以计算实验人脸的个性特征,并通过设置不同的夸张程度对个性特征进行多种形式的夸张,以生成多种不同的肖像漫画。针对夸张个性特征之后应用三角形坐标系进行图像变形所出现的肖像变形异常问题,提出了肖像变形约束准则,该准则可约束人脸不同部分特征的变形范围。实验结果表明,在肖像变形约束准则的约束下,所提出的肖像漫画生成方法可快速生成个性特征突出、生动形象的肖像漫画。

(3)针对现有的表情合成方法在合成不同表情时容易丢失人脸的可辨识信息造成其难以识别的问题,首先通过分析同一个人的中性表情与其他表情之间的差异特征,提出了表情变形约束准则,该准则可在表情合成时约束表情变形,以保留肖像的可辨识信息。在此研究基础上,提出了一种基于三角形坐标系的可辨识表情肖像生成方法,该方法通过比较实验人脸与参考表情得到表情特征,且在表情变形约束准则约束下,通过三角形坐标系的图像变形实现表情合成,生成表情肖像。实验结果表明,提出的方法可以生成带有参考表情的表情肖像,所合成的表情不仅自然生动,还能保持原有的身份可辨识。

(4)为了将人脸的个性特征与表情特征进行融合以生成表情肖像漫画,首先研究了个性特征与表情特征之间的变形方向问题,针对该问题提出了个性特征与表情特征之间的融合准则,在准则基础上提出了一种基于三角形坐标系的表情肖像漫画生成方法,该方法通过判定个性特征与表情特征的变形方向,可自动调整个性特征与表情特征的融合权重,以实现夸张个性特征同时反映表情特征而又不会导致肖像过度变形的表情肖像漫画。实验结果表明,所生成的表情肖像漫画既可以表达肖像自身的个性特征,同时又可以表达表情特征,且肖像效果仍然可辨识。

(5)为使艺术肖像绘制机器人能够像真实画家一样绘制艺术肖像,首先研究了真实画家绘制肖像的规律与技巧等,并在此研究基础上提出了机器人的肖像绘制轨迹规划方法,在肖像绘制顺序、肖像绘图方向以及粗细线条绘制三个方面规划了机器人的肖像绘制轨迹。在所构建的艺术肖像绘制机器人系统中,通过艺术肖像的实际绘图,验证了所提出方法可以实现肖像绘制机器人的拟人绘图。

本论文中所研究的相关技术方法,可在图像变形、人脸识别、异质图像合成与识别等领域中实现应用。

;
In recent years, with the vigorous development of artificial intelligence, it constantly refreshes the cognition of mankind to the unknown world and the limit fields, changing the life and the production mode of human. A variety of intelligent robots constantly emerges, such as smart industrial manufacturing robot, intelligent nursing robot and rescue robot. These robots play an irreplaceable role in human production and life. It has become a research direction of artificial intelligence by using a robot to draw human face portrait, which represents the preliminary exploration of artificial intelligence in the art world. The appearance of the portrait rendering robot makes it simple, efficient and lively to draw portraits that are simple for the artist but impossible for the ordinary to complete.
With the human face portrait as this paper’s research object, the method of generating line portraits which can perceive face feature is first studied, and on the basis of this, the generation of artistic portraits is realized, such as caricature portrait, expression portrait and expression caricature portrait. Then, by planning the trajectory of portrait rendering robot, the robot can complete drawing the portrait like a real artist. Finally, a system of intelligent artistic portrait rendering robot is established which can realize perception, show, process and rendering of real face feature. The research of this system has important academic and scientific significance, and its related research results can be widely used in many fields. In view of this, this article has carried on the research from the following several aspects, and has obtained the innovative results:
In order to solve the problem that the lines generated from face images of low resolution are so unclear that they can’t reflect the true characteristics of human face, the non-classical disinhibitory properties of concerntric receptive field are studied. Then based on this research, a line portrait generation based on Tri-Gaussian model is proposed. The experimental results show that the proposed method can not only enhance the image's line effect, but also suppress the noise, and maintain the structure and directional information of the original images.
Aiming at the problem of the dependence on training samples, high complexity and the single way and degree to exaggerating the portrait of the existing caricature portrait generation algorithms, the method of image topological deformation based on triangle coordinate system is studied. Based on the advantages of simple calculation and flexible deformation, a method of caricature portrait generation based on triangle coordinate system which can control the exaggerating degrees is designed. By comparing the experimental face with the reference face, the method can compute the personalized facial features of the experimental face and generate a variety of different caricature portraits by setting different exaggeration degrees. In allusion to the problem of deformation anomaly after exaggerating the personalized facial features, the constraint criterion of portrait deformation is proposed, which can constrain the deformation range of different parts of the face. The experimental results show that, under the constraint of the portrait deformation constraint, the proposed method of caricature portrait generation can quickly generate vivid caricature portraits with prominent personalized facial characteristics, and this method has a certain robustness to the expression and gesture of portrait.
In view of the problem that the existing expression synthesis method can easily lose the recognition information of human face when synthesizing different expressions, firstly, by analyzing the difference between the same person's neutral expression and other expressions, the constraint criterion of expression deformation is put forward, which can restrain the expression distortion in the expression synthesis to preserve the recognizable information of the portrait. On the basis of this research, an identifiable expression portrait generation method based on triangle coordinate system is proposed, which can be used to compare experimental human faces and reference expressions to get the expression features, and the expression portrait is synthesized by the image deformation of triangle coordinate system under the constraint of expression constraint criterion. The experimental results show that the proposed method can be used to synthesize a facial expression with a reference expression, which is lively and vivid, and the generated expression can be identified.
In order to combine the personalized features with the expression features to produce the expression caricature portraits, firstly, the paper studies the problem of the deformation direction between the personalized and expression features. In allusion to this problem, the fusion criterion between the personalized and expression features is put forward. Then a method based on the triangle coordinate system to generate the expression caricature portraits is proposed. This method can automatically adjust the fusion weights of personalized and expression features by judging the direction of the deformation between the above two features, so as to realize the exaggerated personalized characteristics and reflect the expression features without causing the portrait to be excessively deformed. The experimental results show that the generated expression caricature portraits can not only express the personalized features, but also express the expression, and the image can still be identified.
In order to make the artistic portrait rendering robot draw the artistic portrait like a real painter, the workflow and the techniques of real painters to draw portraits are studied. Based on that, a robot trajectory planning method is put forward, which can plan the trajectory of the robot by the drawing order, the drawing direction and drawing thick and thin lines. In the established artistic portrait rendering robot system, through a large number of actual drawing of artistic portraits, experiment results can show that the proposed method can be make the robot draw a portrait like a real artist. 
The related technical methods studied in this paper can be extended to other fields like image distortion, face recognition, heterogeneous image synthesis.
学科领域人工智能
语种英语
公开日期2018-05-29
文献类型学位论文
条目标识符http://ir.semi.ac.cn/handle/172111/28392
专题高速电路与神经网络实验室
推荐引用方式
GB/T 7714
董肖莉. 基于人脸图像的机器肖像艺术技术研究[D]. 北京. 中国科学院大学,2018.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
基于人脸图像的机器肖像艺术技术研究.pd(2678KB) 限制开放使用许可请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[董肖莉]的文章
百度学术
百度学术中相似的文章
[董肖莉]的文章
必应学术
必应学术中相似的文章
[董肖莉]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。