【作者】 张铮；

【导师】 赵政；

【作者基本信息】 天津大学 ， 计算机应用技术， 2010， 博士

【摘要】 目前表情识别已经成为模式识别和机器视觉研究领域的一个活跃分支,越来越多的研究者开始关注这个领域。很多相关的实际应用都要求能够在单幅图像中完成表情识别任务,由于静态图像中所含有的表情信息量相对较少,在其上的表情识别较之基于视频序列的表情识别更为困难。为了在独立于个体身份,基于单幅面部图像的表情识别中取得更加理想的效果,本文探索几种不同的特征在表情识别中的效果,并尝试通过特征加权改善分类性能;同时就变化光照下的表情识别这一颇具挑战性的问题进行了研究,提出了一种能够在表情图像中检测光照条件的方法。论文的主要工作和创新点如下:(1)在特征提取环节,将Gabor多分辨率分析的思想同MB-LBP编码的概念相结合,从而得到了对于图像局部和整体信息都有较好描绘能力的多尺度局部Gabor二进制模式(Multi-block Local Gabor Binary Patterns, MB-LGBP)复合特征。(2)考虑到区分不同的表情需要对图像局部纹理比较精确的描述,而LBP特征是对图像纹理的一阶描述,不能反映其空间结构信息,单纯通过对图像分区的方法无法从根本上解决这个问题。因此,本文以灰度共生矩阵(GLCM)替代传统的直方图,提出了一种能够表征空间特定结构信息的二维LBP特征,进而又根据特征融合的思路得到了2维MB-LGBP复合特征作为分类特征。(3)在识别中,选用C-SVM作为分类器,并提出了一种适用于两两分类、依赖于二分器的特征加权机制,对于每次两两分类都使用不同的权重分布。同其他在表情识别中被广泛采用的分类器相比,本文的方法取得了更加理想的识别率。(4)为解决在表情识别中的光照干扰问题,提出了一种基于三维表征脸(Representative Face, RF)和聚类的方法来识别某幅二维人脸表情图像中的光照条件,从而为进一步消除光照干扰做好了准备。该方法利用K-Means技术对三维模型中的面部主要器官位置进行聚类,在此基础上生成表征脸模型,使得光照条件的匹配仅需基于表征脸进行。同传统基于三维重建的光照检测技术相比,该方法具有匹配复杂度上的优势。更多还原

【Abstract】 Nowadays, expression recognition has become an active topic in pattern recognition and machine vision community. More and more researchers have focused on this domain. In many applications it is necessary to recognize expression in a single static image. However, due to the fact that less expression information is available in static images, expression recognition from static images is more difficult than that from image sequences. Inorder to achieve better person-independent expression recognition accuracy in static images, two kinds of feature and dichotomy-dependent weights are presented in this paper. And to eliminate the negative effect caused by variant illumination in expression recognition, a 3D representative face and clustering based framework is presented in this paper. The main work and innovations of this paper are as follows:(1)In feature selection, we combine the concept of Multi-scale Gabor analysis with MB-LBP encoding to achieve the so-called MB-LGBP features which is both locally and globally informative.(2)The discrimination of different expressions needs more precise description of local textures. As 1st-order description, LBP can not encode spatial structure information. And this problem can not be solved in nature only by partitioning. So we utilize Gray Level Co-occurrence Matrix instead of the traditional statistical histogram and present the so-called 2-dimensional LBP features. By following the fashion of feature fusion, finally we get the 2D MB-LGBP composite features for classification.(3)In classification, dichotomy-dependent weights for SVM is introduced and its performance is compared with the traditional k-nearest neighbor paradigm based on weighted Chi Square distance. The result we get is promising, which proves the superiority of the 2-dimensional MB-LGBP features to some other popular features in expression recognition.(4)To eliminate the negative effect caused by variant illumination in expression recognition. A 3D representative face (RF) and clustering based framework is presented in this paper, which can estimate 13 illumination conditions under certain poses accurately. By adaptively clustering 3D faces into a number of facial types, subjects with similar facial appearance are clustered together. Then the RF of each cluster is generated, which provides our system the generalization ability to do subject-dependent illumination estimation. Compared with other works which rely on 3D reconstruction, our method has less computation complexity.更多还原