CASIA OpenIR  > 毕业生  > 博士学位论文
基于统计学习的实时表情识别研究
其他题名Real-time Facial Expression Recognition Based on Statistical Learning
周晓旭
学位类型工学博士
导师王阳生
2005-05-10
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词实时表情识别 嵌入式 Hmm Aml Adaehmm Boosting Real-time Facial Expression Recognition Embedded Hmm
摘要本文研究基于统计学习的实时表情识别,在理论研究和实际应用方面都取得了很好的成果。主要贡献包括:第一,在表情识别的预处理阶段,本文创造了一种全新的特征空间:区域磁力线(AML,Area Magnetic Line)特征空间,并将其用于人脸检测方面。我们将 AML 与当前最流行的 Haar 特征空间做了对比实验。AML不但可以描述 Haar 特征所能描述的情况,并且更优于 Haar 特征的描述能力。利用 AML,我们不单单只是去描述相邻区域的相互关系,还可以描述不相邻区域的相互关系,这是 Haar 特征无法做到的。通过实验可以看出,使用 AML 特征空间提升了人脸检测的性能。 第二,在人脸定位部分,我们对局部搜索策略进行了改进,提出了新颖的发散 ASM 方法。此方法解决了传统的 ASM 方法中很容易将搜索点定位到局部极小值的情况。当输入图像的质量较差而难以对齐的时候,我们将可能趋于局部极小值的点发散到该组观察点最外侧的点的位置上,然后通过判断最终得到的模型挑选出对齐正确的图像。通过发散 ASM,我们可以对人脸定位的结果进行评估,进而提高整个表情识别系统的性能。第三,提出了一种新型的基于嵌入式隐马尔可夫模型(EHMM,Embedded HMM)的实时表情识别系统,首次将嵌入式隐马尔可夫模型应用到表情识别中,并且取得了不错的效果。我们的算法使用观察序列窗口的二维离散余弦变换(2D-DCT)系数作为观察序列,从而减少了观察向量的个数,大大降低了训练和识别系统的复杂程度。从试验结果可以看出,该方法对于实时的面部表情识别是一种有效的方法。此外,我们还将我们的实时表情识别系统与实时的人机交互网络游戏相结合,通过摄像头采集玩家的脸部视频流,并将截获到的各帧图像送到表情识别系统中进行表情识别。游戏角色表情判别模块通过表情识别的结果更改游戏窗口中虚拟角色的表情,并保持与现实中玩家的表情同步变化,从而大大提高了用户与计算机之间的互动,增加了在网络游戏环境下人与人之间的交互性。
其他摘要The main contributions are as follows:Firstly, we proposed a novel character space in facial expression recognition called Area Magnetic Line and applied it in face detection. We implemented a comparative experiment to analyze the difference between AML and Harr-like. AML can describe all the circumstances that Haar can describe with a better feature description capability. Using AML, we can not only describe the relationship between adjacent areas, but also those isolated areas, as is not attained by Haar.Secondly, we improved the local detecting strategy and proposed a creative scattered ASM. This approach solves the problem that traditional ASM tend to anchor the detecting point at the local minimum value. Once the input image quality is low and not easy to be aligned, we scatter the points that may go towards the local minimum point to the most outboard points, and then selected the rightly aligned image by judging the final models. With scattered ASM, we can evaluate the face alignment results and enhance the performance of the entire facial expression system.Thirdly, we introduced embedded HMM based real time expression recognition system as the debut of embedded HMM in the field of facial expression recognition research. The present algorithm uses the 2D-DCT of observed sequence windows as the observed sequence. It largely reduces the number of observed vectors and thus reduces the complexity of the training and recognition system. The results showed that this approach is an effective way for real time face recognition. Furthermore, we also combined the real time expression system with real-time human-computer interactive network game. During this course, face vision flows of the players were captured by webcams and all the image frames were delivered to the expression recognition system to be checked. The role-play expression judgment module will change the visual roles’ expression according to the expression recognition results to be synchronous with the true players’ expression. Finally, based on the studying of chapter 3, we proposed the statistical learning approach, Boosting, which owns self studying ability, into EHMM real time facial expression recognition system. Then we construct an embedded HMM classifier: AdaEHMM classifier, by which we endow self learning ability to parameter training and structure selecting of EHMM. By adjusting weight of the samples, it enhanced the classifying capability of the classifier and finally got the combination of AdaEHMM classifiers with strong classifying capability. Not only optimal EHMM structured can be attained, but the expression recognition rate can be enhanced as well.
馆藏号XWLW944
其他标识符200318014600958
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5868
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
周晓旭. 基于统计学习的实时表情识别研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2005.
条目包含的文件
条目无相关文件。
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
[周晓旭]的文章
百度学术
百度学术中相似的文章
[周晓旭]的文章
必应学术
必应学术中相似的文章
[周晓旭]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。