人脸关键点定位及其应用研究
文献类型：学位论文
| 作者 | 王书昌 |
| 学位类别 | 工学硕士 |
| 答辩日期 | 2007-06-20 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 王阳生 |
| 关键词 | 人机交互 表情交互 人脸关键点定位 主动形状模型 主动外观模型 Human-Computer Interaction Interaction Based on Facial Expression Face Alignment Active Shape Models Active Appearance Models |
| 其他题名 | Face Alignment and its Applications |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 人脸关键点定位研究有着重要的学术价值和广泛的应用前景。本文以设计一个自动的人脸表情交互系统为目标，针对此系统涉及的人脸检测，人脸关键点定位，特征表达和表情识别及表情驱动进入了深入的研究。研究中，我们着重考虑了人脸关键点定位的改进和提高。鉴于人脸模式的特点及特征的特性，我们将融合策略、层次策略和加权策略引入到学习框架中，提高了人脸关键点定位的精度和鲁棒性。本文实现了鲁棒实用的人脸关键点定位并应用于人机交互，对若干核心技术和关键问题做出了一定的贡献，论文的主要贡献如下： １.针对基于主动形状模型的人脸关键点定位算法，提出了一种基于灰度和边强度的融合特征－－我们提出的融合特征，强调了边强度信息对局部搜索的启发性，减少了搜索发散的可能性。 ２.针对基于主动外观模型的人脸关键点定位算法，我们从五个方面予以改进： 提出了以灰度和边强度为综合纹理的主动外观模型； 将综合成纹理的AAMs推广到更为一般的情况，并提出了由粗到细的迭代策略；考查了纹理中不同象素的重要性差别，提出了一种基于加权图的动外观模型；针对人工标定样本的误差会在纹理放大的问题，我们提出了一种半自动的多次反复的样本标定方法； 提出了对纹理进行区域分块和多次平滑的方法，有效地克服了光线的影响，拓展了动外观模型的使用环境。 ３.利用基于权值图的主动外观模型为核心技术我们开发了一个人脸表情驱动系统，从工程角度探讨了各个子功能最优的实现方法。 |
| 英文摘要 | The research on Face Alignment has both significant academic importance and wide applications. With the aim of designing an automatic Human-Computer interaction system based on face expression, we focus our research on the face detection, face alignment, feature representation, facial expression analysis and avatar control by facial expression. During the process, an in-depth study is made on face alignment. With consideration of face model and feature speciality, we incorporate feature combining and transforming strategy into the learning framework and searching process which dramatically improves the robustness and accuracy of alignment result. And the main contribution of this thesis includes: 1. With regard to the face alignment based on Active Shape Models, we proposed a combination feature of gray and edgeness which balances the global and local information. 2. With regard to the face alignment based on Active Appearance Models, we made improvement from five aspects. We proposed a combined Active Appearance Models whose texture feature is the combination of gray and edge information. We extended the combined Active Appearance Models to a weighted form, which enables us to employ a coarse-to-fine search strategy. Considering the face that different pixels in face area take different role of importance in the texture match process, we incorporated a weight-map into the Active Appearance Models. In order to minimize the error introduced during labeling training samples by hand, which dramatically affects the robustness of the texture model, we suggest a semi-automatic and multi-pass labeling strategy be employed to build a good texture model. In order to build a texture model less sensitive to light variation, we proposed a texture pre-process method which is subdividing the face area and then smoothing the sub-areas several times respectively. 3. Using the improved Active Appearance Models, we developed an avatar control system and in every aspect of the function of which, we proposed the best ways of implementation. |
| 语种 | 中文 |
| 其他标识符 | 200428014628038 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/7416]  |
| 专题 | 毕业生_硕士学位论文 |
| 推荐引用方式 GB/T 7714 | 王书昌. 人脸关键点定位及其应用研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2007. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
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。