基于改进的遗传和Pareto优化算法的人脸表情识别

发布时间：2018-01-20 04:02

本文关键词： 遗传算法特征表情识别 Pareto优化随机森林 uniform LGBP 三维人脸形变模型 ISOMAP算法　出处：《东华大学》2017年硕士论文　论文类型：学位论文

【摘要】：人的大脑具备天生的人脸识别能力,可以轻易地分辨出不同的人。但是机器自动识别人脸技术的发展却面临巨大的挑战。在实际环境中,二维人脸识别不可避免地受到光照、姿态和表情的影响,这些因素已成为了二维人脸识别技术向前发展的最大障碍。同时由于三维人脸形变模型的提出使得三维人脸表情识别变得非常热门,如何建立鲁棒快速的三维人脸表情识别模型成为了人脸识别中一大挑战。本文基于改进的遗传和Pareto优化算法对人脸表情进行识别以提高人脸表情识别的准确率。首先使用Haar-like特征表示方法和双边滤波器对人脸图片进行预处理。其次,利用uniform LGBP方法对人脸的特征进行提取,降低特征维数。再次,改进GA适应度函数,提出新的Pareto目标函数并采用改进的GA和Pareto优化算法对最优的显著特征进行挑选。最后,使用随机森林分类器对人脸特征进行分类。在三维人脸表情识别中,通过将三维人脸顶点映射到二维空间中,然后使用二维人脸表情识别的方法进行识别。最后与现有的算法进行比较试验,实验表明,本文所提出的人脸表情识别算法在识别精度和计算时间上都优于现有文献中提出的算法。本文主要的创新点有:1)提出uniform LGBP对人脸特征进行提取的方法,有效地降低了提取特征的维数。2)提出将GA和Pareto算法结合,对最优显著特征进行挑选,为了提升人脸表情识别的精度,改进GA的适应度进化函数,并提出了两个新的Pareto优化算法的目标函数来刻画最小化类内变化和最大化类间变化。3)基于鲁棒的三维人脸形变模型,提出加入正则化项的三维人脸形状拟合目标函数,使三维人脸形状不会产生过拟合问题。4)提出使用ISOMAP算法将三维人脸的顶点映射到二维空间中,可以有效地对三维顶点内在的几何结构进行学习,使映射得到的二维人脸不存在形变问题。
[Abstract]:The human brain has the natural ability of face recognition, it can easily distinguish different people, but the development of the machine automatic face recognition technology is facing a huge challenge. In the actual environment. Two-dimensional face recognition is inevitably influenced by illumination, pose and expression. These factors have become the biggest obstacle to the development of two-dimensional face recognition technology. At the same time, 3D facial expression recognition has become very popular due to the proposed three-dimensional face deformation model. How to establish a robust and fast 3D facial expression recognition model has become a major challenge in face recognition. This paper presents an improved genetic and Pareto algorithm for facial expression recognition in order to improve facial expression recognition. First, the Haar-like feature representation method and bilateral filter are used to preprocess the face image. Uniform LGBP method is used to extract face features to reduce the feature dimension. Thirdly, the GA fitness function is improved. A new Pareto objective function is proposed and the improved GA and Pareto optimization algorithms are used to select the salient features of the optimization. Finally. A random forest classifier is used to classify facial features. In 3D facial expression recognition, 3D face vertices are mapped to two-dimensional space. Then the two-dimensional facial expression recognition method is used to recognize. Finally, compared with the existing algorithms, the experiment shows that. The facial expression recognition algorithm proposed in this paper is superior to the one proposed in the literature in terms of recognition accuracy and computation time. The main innovation of this paper is: 1). A method of facial feature extraction based on uniform LGBP is proposed. In order to improve the accuracy of facial expression recognition, GA and Pareto algorithms are combined to select the optimal salient features. The fitness evolution function of GA is improved. The objective functions of two new Pareto optimization algorithms are proposed to characterize the 3D face deformation model based on robust algorithm. The objective function of 3D face shape fitting with regularization term is proposed, so that the 3D face shape will not be overfitted. (4) A ISOMAP algorithm is proposed to map the vertex of 3D face to two-dimensional space. The inherent geometric structure of 3D vertices can be effectively studied, so that there is no deformation problem in the two-dimensional face generated by mapping.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】