基于回归算法的人脸识别分类器设计
发布时间：2018-11-09 09:56
【摘要】：目前,人脸识别技术已经被应用于我们的日常生活当中的某些领域,但是该技术在手机端身份验证和支付这些场景中的应用还没普及,一方面是由于人脸识别准确度面部姿势、光照、表情变化影响较大,另一方面是目前已有的对复杂环境下的人脸分类效果好的分类算法计算量太大,不适合应用在内存较小的手机端。因此,本文的研究目的是改进计算量小的分类器设计,提升人脸在不稳定条件下的人脸识别率。在模式识别领域中,基于最近邻分类器提出的最近特征面分类器(Nearest Feature Plane,NFP)和线性回归分类器(Linear Regression Classifier,LRC)是两种计算量小且分类识别较好的分类器,具备被应用于手机端的条件。因此,本课题对线性和非线性回归分类器的算法思想进行了详细的分析和研究,基于线性与非线性回归分类器的算法思想提出了几种分类器的设计方法。本文的研究方法和成果包括以下几个方面,针对分类过程中易错分点的分类不精确性问题,基于线性回归分类器的算法思想,提出了三种改进的分类器,分别为基于点线距离分类器、伪线空间分类器和距离受限分类器。三种新方法分别利用了样本点与回归直线之间的距离信息及样本点间的空间特性对线性回归算法进行改进,大量的对比实验表明新方法可以有效提升分类器在光照、角度和表情变化下易错分人脸图像的识别率。同时,基于核函数和最近特征面分类器,提出了一种加核最近最远分类器和基于中心受角度限制的最近特征面分类器。其中,加核最近最远分类器利用核函数能有效地解决样本非线性可分问题的优势,将核函数与最近最远线性回归分类器结合,有效地提升了原样本空间中非线性可分样本的分类准确性。而基于最近特征面的通过增加角度来改进由于特征面的无限延长而引起的交叉样本错分问题,通过将两种分类器在多个标准人脸库上与其他的改进方法进行对比实验,验证了两种新方法对易错分交叉样本的分类准确性。
[Abstract]:At present, face recognition technology has been used in some areas of our daily life, but the application of this technology in mobile phone authentication and payment of these scenes has not been widely used, on the one hand, because of the face recognition accuracy of facial posture, The change of illumination and expression has a great influence on the face classification. On the other hand the existing classification algorithms which have good effect on face classification in complex environment have too much computation and are not suitable for the mobile phone with small memory. Therefore, the purpose of this paper is to improve the classifier design with small computational complexity, and to improve the face recognition rate under unstable conditions. In the field of pattern recognition, the nearest feature surface classifier (Nearest Feature Plane,NFP) and the linear regression classifier (Linear Regression Classifier,LRC) proposed by the nearest neighbor classifier are two kinds of classifiers with small computation and good classification recognition. Has the condition to be applied to the mobile phone. Therefore, the algorithm of linear and nonlinear regression classifier is analyzed and studied in detail. Based on the algorithm of linear and nonlinear regression classifier, several design methods of classifier are proposed. The research methods and achievements of this paper include the following aspects. Aiming at the problem of the inaccuracy of the classification of error-prone points in the classification process, three improved classifiers are proposed based on the idea of linear regression classifier. It is based on point line distance classifier, pseudo line space classifier and distance limited classifier respectively. The three new methods make use of the distance information between the sample points and the regression lines and the spatial characteristics of the sample points respectively to improve the linear regression algorithm. A large number of comparative experiments show that the new method can effectively improve the illumination of the classifier. The recognition rate of error-prone face image under the change of angle and expression. At the same time, based on kernel function and nearest feature surface classifier, a kernel nearest furthest classifier and a nearest feature surface classifier based on center angle are proposed. The kernel nearest farthest classifier combines the kernel function with the nearest farthest linear regression classifier by utilizing the advantage of kernel function to solve the nonlinear separable problem of samples effectively. The classification accuracy of nonlinear separable samples in the original sample space is improved effectively. Based on the most recent feature surface by adding angle to improve the cross-sample misclassification problem caused by the infinite extension of the feature surface, two classifiers are compared with other improved methods on multiple standard face databases. The accuracy of the two new methods for the classification of error-prone cross samples is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2320076
[Abstract]:At present, face recognition technology has been used in some areas of our daily life, but the application of this technology in mobile phone authentication and payment of these scenes has not been widely used, on the one hand, because of the face recognition accuracy of facial posture, The change of illumination and expression has a great influence on the face classification. On the other hand the existing classification algorithms which have good effect on face classification in complex environment have too much computation and are not suitable for the mobile phone with small memory. Therefore, the purpose of this paper is to improve the classifier design with small computational complexity, and to improve the face recognition rate under unstable conditions. In the field of pattern recognition, the nearest feature surface classifier (Nearest Feature Plane,NFP) and the linear regression classifier (Linear Regression Classifier,LRC) proposed by the nearest neighbor classifier are two kinds of classifiers with small computation and good classification recognition. Has the condition to be applied to the mobile phone. Therefore, the algorithm of linear and nonlinear regression classifier is analyzed and studied in detail. Based on the algorithm of linear and nonlinear regression classifier, several design methods of classifier are proposed. The research methods and achievements of this paper include the following aspects. Aiming at the problem of the inaccuracy of the classification of error-prone points in the classification process, three improved classifiers are proposed based on the idea of linear regression classifier. It is based on point line distance classifier, pseudo line space classifier and distance limited classifier respectively. The three new methods make use of the distance information between the sample points and the regression lines and the spatial characteristics of the sample points respectively to improve the linear regression algorithm. A large number of comparative experiments show that the new method can effectively improve the illumination of the classifier. The recognition rate of error-prone face image under the change of angle and expression. At the same time, based on kernel function and nearest feature surface classifier, a kernel nearest furthest classifier and a nearest feature surface classifier based on center angle are proposed. The kernel nearest farthest classifier combines the kernel function with the nearest farthest linear regression classifier by utilizing the advantage of kernel function to solve the nonlinear separable problem of samples effectively. The classification accuracy of nonlinear separable samples in the original sample space is improved effectively. Based on the most recent feature surface by adding angle to improve the cross-sample misclassification problem caused by the infinite extension of the feature surface, two classifiers are compared with other improved methods on multiple standard face databases. The accuracy of the two new methods for the classification of error-prone cross samples is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 甘俊英;何思斌;;基于2DLDA与SVM的人脸识别算法[J];计算机应用;2009年07期
相关博士学位论文 前1条
1 钱建军;人脸图像特征提取和分类的若干方法研究[D];南京理工大学;2014年
相关硕士学位论文 前1条
1 侯义福;基于低秩与特征脸的稀疏表示人脸识别研究[D];安徽大学;2015年
，本文编号：2320076
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2320076.html
