人脸识别中若干特征优化方法研究

发布时间：2018-10-17 19:45

【摘要】：随着信息技术的快速发展,生物特征识别技术正在被大范围地应用到金融、安防等领域并受到了社会各界的广泛认可。与其它生物特征识别技术相比,人脸识别技术具有友好、直观、可靠等特点,所以人脸识别技术已成为生物特征识别技术中比较具有代表性的一项技术。人脸图像易受光照、表情等因素的影响,因此在实际应用中仍然存在很多挑战。在人脸识别模型中涉及到图像预处理、维数约简以及分类判决三个主要组成部分,如何有效利用特征与分类器之间的关系,从而增强特征判别力并且提高特征简洁度成为了当前研究的热点问题。本文从基于表示的分类器(Representation based Classifier,RC)角度出发,在有监督信号的情况下分别通过特征增强和特征提取方法实现了特征与分类器间的相互影响和相互制约。另外,本文以特征自表示模型为基础,通过特征间的线性表示及内积约束等实现了无监督的特征选择。有关这三种特征优化方法的具体工作总结如下:1.提出了一种新的滤波器学习方法,即基于表示的有监督滤波器学习方法。该方法通过有针对性地学习得到滤波器,并使滤波后图像的局部特征判别力增强,从而实现减小同一个人不同图像间差异,并且增大不同人间图像差异的目的。该方法的特点有:(1)在监督信号下,从局部二值模式(Local Binary Pattern,LBP)角度出发设计滤波器,从而使滤波后图像的LBP特征具有判别能力;(2)利用线性回归方法刻画图像像素点间的类内和类间表示误差,并在线性判别分析的约束下得到滤波器,从而使滤波后特征在稀疏表示分类器和线性回归分类器下得到更好的识别结果;(3)与采用固定模式的传统滤波器(如均值滤波器)不同,该方法是在数据驱动情况下有针对性地学习滤波器;(4)在单模态和多模态人脸数据库上均验证了该方法的有效性。通过大量的实验可以看出,该方法可以有效提高特征的判别力,并且在RC下可以得到更好的分类结果。2.结合字典学习提出了一种新的特征提取方法,即基于判别字典与投影联合学习的稀疏表示分类方法。该方法通过同时学习带有约束的字典和投影矩阵,不仅得到了更具表示力和判别力的字典,还得到了维数更低且更具判别力的特征,从而提升了人脸识别模型的分类性能。该方法的特点有:(1)通过对稀疏表示系数矩阵加入线性判别分析约束得到了具有判别能力的字典,并且通过对降维后样本加入线性判别分析约束得到了具有判别能力的投影矩阵;(2)通过联合学习使得字典和投影矩阵能够更好地相互配合,进而得到更好的识别结果;(3)提出了一种有效的迭代优化求解算法,并分别从理论分析和数值实验两方面验证了算法的收敛性;(4)在人脸图像和视频数据库上均验证了该方法的有效性。通过大量的实验可以看出,该方法可以有效提高特征简洁度并增强特征的判别力,即使在训练样本数较少的情况下仍然可以取得较好的识别性能。3.提出了一种新的无监督特征选择方法,即基于内积正则化非负自表示模型的无监督特征选择方法。该方法通过特征自表示模型和内积约束等去除了不相关特征及冗余特征,从而使特征子集具有较高的稀疏性和较低的冗余性。该方法的特点有:(1)利用特征的自表示模型来描述特征的显著程度,从而获得特征的权重矩阵;(2)采用内积正则化对特征权重矩阵进行约束,由此可以获得具有较高稀疏性和较低冗余性特点的特征子集;(3)采用非负约束对特征权重矩阵进行约束,从而保证所选特征的实际意义;(4)提出了一种有效的迭代优化求解算法,并分别从理论分析和数值实验两方面验证了算法的收敛性。实验结果表明该方法不仅可以有效提高特征的简洁度,而且可以得到更好的分类和聚类结果。综上所述,本文主要围绕人脸识别模型中特征优化问题展开了广泛而深入地研究,针对如何增强局部特征(LBP)的判别力、如何通过学习投影矩阵提高特征的简洁度和判别力以及如何提高特征子集有效性的问题,分别提出了三种特征优化方法。从实验结果可以看出,本文提出的方法对人脸识别研究有一定的推动作用并具有较好的应用前景。
[Abstract]:With the rapid development of information technology, biometric identification technology is being widely applied to finance, security and other fields and is widely recognized by all circles of society. Compared with other biometric identification technologies, face recognition technology has the characteristics of friendship, intuition, reliability and so on, so the face recognition technology has become a representative technique in biological feature recognition technology. Face images are easy to be influenced by light, expression and other factors, so there are still many challenges in practical application. In the face recognition model, three main components of image preprocessing, dimension reduction and classification decision are involved, how to effectively utilize the relationship between feature and classifier, so that the characteristic discrimination force is enhanced and the characteristic simplicity is improved to become the hot issue of the current research. Based on the representation based classifier (RC), the interaction and mutual restriction between the feature and the classifier are realized by feature enhancement and feature extraction. In addition, based on the feature self-representation model, the feature-free feature selection is realized through linear representation and inner product constraint among features. Specific work on these three feature optimization methods is summarized as follows: 1. A new method of filter learning is proposed, which is based on the representation of supervised filter learning method. the method achieves the purpose of reducing the difference between different images of the same person and increasing the difference between different human images by carrying out targeted learning to obtain a filter and enhancing the local characteristic discrimination force of the filtered image. The method is characterized in that: (1) under the supervision signal, the filter is designed from the local binary pattern (LBP) angle, so that the LBP characteristic of the filtered image has the discrimination ability; (2) using the linear regression method to depict the intra-class and inter-class representation errors among image pixel points, and obtaining a filter under the constraint of linear discriminant analysis, so that the filtered features obtain better recognition results under sparse representation classifier and linear regression classifier; (3) Different from the traditional filter with fixed mode (such as the mean filter), the method is to study the filter with pertinence under the condition of data driving, and (4) the validity of the method is verified on the single mode and the multi-modal face database. As can be seen from a large number of experiments, the method can effectively improve the distinguishing force of the characteristic, and can obtain better classification result under the RC. A new feature extraction method is proposed in combination with dictionary learning, which is based on the sparse representation classification method of discrimination dictionary and projection joint learning. By studying the dictionary and the projection matrix with constraints at the same time, the method not only obtains the dictionary which is more representative of the force and the discrimination force, but also obtains the feature of lower dimension number and more discriminating force, thereby improving the classification performance of the face recognition model. The method is characterized in that: (1) a dictionary with discrimination capability is obtained by adding a linear discriminant analysis constraint to a sparse representation coefficient matrix, and a projection matrix with discrimination capability is obtained by adding linear discriminant analysis constraint to the reduced-dimensional post-sample; (2) through joint learning, the dictionary and the projection matrix can be better matched with each other so as to obtain better recognition results; (3) an effective iterative optimization solution algorithm is proposed, and the convergence of the algorithm is verified from two aspects of theoretical analysis and numerical experiment, respectively; (4) The validity of the method is verified on the face image and the video database. As can be seen from a large number of experiments, the method can effectively improve the characteristic simplicity and enhance the distinguishing force of the characteristic, and even if the number of the training samples is small, better identification performance can be obtained. A new non-supervised feature selection method is proposed, i.e., the non-supervised feature selection method based on inner product regularization non-negative self-representation model. According to the method, the feature self-representation model and the inner product constraint are used to remove irrelevant features and redundant characteristics, so that the feature subsets have higher sparsity and lower redundancy. The method is characterized in that: (1) a characteristic self-representation model is utilized to describe the salient extent of the feature so as to obtain a weight matrix of the feature; and (2) the feature weight matrix is constrained by using the inner product regularization, therefore, a feature subset with higher sparsity and lower redundancy characteristics can be obtained; (3) a feature weight matrix is constrained by adopting a non-negative constraint, so that the practical significance of the selected feature is ensured; and (4) an efficient iterative optimization solution algorithm is proposed, The convergence of the algorithm is verified from both theoretical and numerical experiments. The experimental results show that the method not only can effectively improve the simplicity of the feature, but also can obtain better classification and clustering results. To sum up, this paper mainly studies the feature optimization problem in face recognition model, and aims at how to enhance the distinguishing force of local feature (LBP). How to improve the simplicity and discrimination of the feature by studying the projection matrix and how to improve the effectiveness of the feature subset are presented. It can be seen from the experimental results that the method proposed in this paper has a certain promoting effect on face recognition research and has a good application prospect.
【学位授予单位】：东北师范大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】