三维人脸表情识别中特征提取算法研究

发布时间：2018-05-14 09:53

本文选题：三维人脸表情识别 + 特征提取　；参考：《北京交通大学》2016年博士论文

【摘要】：表情是人类情感理解和分析的主要载体之一。人脸表情识别的目的是让计算机学习人类的情感表达,使其能像人类一样具有识别、理解和表达情感的能力。随着人们对计算机智能性要求的日益增强,该课题已受到国内外研究机构的广泛关注。传统基于二维图像的人脸表情识别已取得较好的识别结果，，但仍存在一些问题没有解决,如光照及姿态变化等,而这是由二维图像的固有属性决定的,因此,基于二维图像的人脸表情识别很难突破该类问题。三维人脸表情是对表情的三维形状的表达,其获取不受光照等外部环境变化的影响。因此,三维人脸表情能够有效避免这些外部因素的影响,从而获得较好的识别结果。本文在对三维人脸表情深入分析的基础上,针对特征提取算法进行研究。主要的研究内容和创新点包括以下四个方面：1.引入夹角特征,并提出其与距离特征融合的人脸表情特征描述形式基于FACS (Facial Action Coding System)对不同人脸表情定义的对比,本文首先提出夹角特征用于表征表情变化引起的人脸形变信息。而传统距离特征可用于描述人脸上发生的相对位移,这两组特征分别对人脸表情变化的不同属性进行描述,使得它们的特征分布的相关性较小。因此,本文进一步提出采用这两组特征融合的方法获取更多的人脸表情变化信息量,从而实现对人脸表情更精准的描述。实验表明,本文提出的夹角特征对人脸表情的表征具有一定的适用性,且它与距离特征的融合能够有效地改善三维人脸表情的识别结果。2.提出了基于PCA相对熵的特征鉴别力判断准则根据贝叶斯理论分析可得,特征的鉴别力依赖其异构的条件概率分布,相对熵可用于评估两组概率分布之间的差异。本文基于算术平均法改进了相对熵使其具有对称性,以满足度量要求。同时,融入PCA算法,使得PCA相对熵具有更好的普适性。该准则计算复杂度小,可用于有效地选取鉴别力强的表情特征。实验中对三维人脸表情提取多组特征,并采用多种分类器进行表情识别。结果表明,三维人脸表情的识别结果与其所采用特征的PCA相对熵的相对变化趋势保持一致,从而验证了该准则的有效性。3.提出了基于平均中性脸生成人脸表情深度差分图的算法本文首先提出采用三维栅格化的方法将三维人脸的存储形式转换为方格结构,并统计获得三维人脸的平均中性脸,进而去除原三维人脸表情的中性脸成分,并生成其对应的深度图。由于该类人脸表情深度图去除了中性脸成分,仅由表情的残留成分表征人脸表情的变化,使得所获得的人脸表情特征具有更好的表情可分性和个体无关性。实验表明,基于平均中性脸生成的人脸表情深度差分图保留了人脸表情变化的主要信息,并有效地保持了人脸表情变化的相对强度及其主要集中于人脸局部区域的特性。该类深度图能够有效地表征人脸表情变化的本质特性,提高其识别结果,这进一步验证了利用深度图分析三维人脸表情的有效性。4.提出了一种有效的三维人脸表情特征：IreEnLBP本文首先提出“类图像方格结构”,并基于该结构实现三维人脸表情的预处理,使得传统的特征可直接用于表征三维人脸的表情变化。进而提出了基于人脸特征分布的“不均匀子块划分法”和“熵加权算法”,并应用于LBP特征,最终提出一种有效的三维人脸表情特征：IreEnLBP.该特征不仅具有LBP特征的局部描述特性,而且“不均匀子块划分法”依照人脸主要器官的分布情况划分子块,以保证人脸局部器官结构的完整性,使得不同表情对应划分出不同的人脸子块,进而提高所提取特征的类间可分性。同时,基于每一子块的熵值对局部特征赋予权重,使得所获得的IreEnLBP特征体现出不同局部区域对表情变化的影响,增强其对不同表情描述的独特性。实验表明,IreEnLBP特征能够有效地提高三维人脸表情识别结果。
[Abstract]:Facial expression is one of the main carriers of human emotion understanding and analysis. The purpose of facial expression recognition is to let the computer learn the emotional expression of human beings and make it capable of recognizing, understanding and expressing emotions like human beings. With the increasing demand for computer intelligence, the subject has been widely used by research institutions both at home and abroad. The traditional two-dimensional image based facial expression recognition has obtained better recognition results, but there are still some problems that are not solved, such as illumination and posture change, which are determined by the inherent properties of two-dimensional images. Therefore, facial expression recognition based on two-dimensional images is difficult to break through this kind of problem. The expression of three-dimensional shape is not affected by the changes of the external environment such as illumination. Therefore, the three-dimensional facial expression can effectively avoid the influence of these external factors and obtain better recognition results. Based on the in-depth analysis of the three-dimensional facial expression, this paper studies the feature extraction algorithm. The main research content and the main research content are in this paper. The innovation points include the following four aspects: 1. introducing the feature of the angle, and putting forward the representation of facial expression with the fusion of distance features based on the comparison of different facial expressions based on FACS (Facial Action Coding System). In this paper, the angle feature is first proposed to represent the human face change information caused by the change of the surface condition. The features can be used to describe the relative displacement on the face. These two sets of features respectively describe the different attributes of facial expression changes, making the correlation of their features smaller. Therefore, this paper further proposes to use these two sets of feature fusion methods to obtain more facial expression change information, thus realizing the face table. The experiment shows that the angle feature proposed in this paper has certain applicability to the representation of facial expression, and it can effectively improve the recognition results of 3D facial expression with the fusion of distance features.2. proposed the characteristic discriminability criterion based on the relative entropy of PCA based on Bias theory. The discriminability depends on its heterogeneous conditional probability distribution, and the relative entropy can be used to evaluate the differences between the two groups of probability distributions. Based on the arithmetic mean method, the relative entropy is improved to make it symmetric to meet the measurement requirements. At the same time, the PCA algorithm is incorporated into the PCA relative entropy to have a better universality. The results show that the recognition result of 3D facial expression is consistent with the relative change trend of the relative entropy of PCA, which verifies the validity of the criterion.3.. In this paper, an algorithm for generating facial expression depth differential graph with average neutral face is proposed in this paper. Firstly, a three-dimensional grid method is proposed to convert the storage form of 3D face into square structure, and the average neutral face of the 3D face is obtained, and then the neutral face component of the original 3D face is removed and its corresponding depth map is generated. The depth map of the facial expression removes the neutral face component. It only characterizes the changes of the facial expression by the residual components of the expression, making the facial expression features better expressive and individual independent. The experiment shows that the facial expression depth difference graph based on the average neutral face preserves the main letter of the facial expression change. It also effectively maintains the relative intensity of facial expression changes and the characteristics that mainly focus on the local area of the face. This kind of depth map can effectively characterize the essential characteristics of facial expression changes and improve the recognition results. This further verifies that the effectiveness of the 3D facial expression analysis by using the depth map is an effective.4.. Three dimensional facial expression features: IreEnLBP first proposed "class image square structure", and based on this structure, the three-dimensional facial expression was preprocessed. The traditional features can be used to represent the facial expression changes directly. Then the "uneven subdivision method" and "entropy weighting algorithm" based on the distribution of facial features were proposed. "And applied to the features of LBP, an effective three-dimensional facial expression feature is proposed. IreEnLBP. features not only the local characterization of LBP features, but the" uneven subdivision method "divides the sub blocks according to the distribution of the main organs of the face in order to ensure the integrity of the organ structure of the human face, and make the different expressions corresponding to them. Different human face blocks are divided, and then the separability between the extracted features is improved. At the same time, based on the entropy value of each sub block, the local feature is given weight, which makes the IreEnLBP features reflect the influence of different local regions on the expression change and enhance its uniqueness to different expressions. The experiment shows that the IreEnLBP features can be characterized. The results of 3D facial expression recognition are effectively improved.

【学位授予单位】：北京交通大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】