基于稀疏表示的人脸特征提取与识别算法研究

发布时间：2018-01-22 11:43

本文关键词： 人脸识别特征提取加权稀疏重构稀疏子空间学习图嵌入　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：人脸识别是模式识别领域中的一个涉及面非常广的重要研究方向。由于人脸图像在采集时受环境、光照、表情和姿态等多种变化的影响,使得人脸识别研究极富挑战性。如何快速准确地利用计算机进行人脸的检测与识别是当前人脸识别技术的关键所在。目前,尽管人脸识别研究已经取得了一些成绩,但是仍有许多问题和关键技术有待进一步解决和完善,其中主要包括:人脸特征提取的充分性研究,即如何充分嵌入局部和全局结构信息等;人脸特征识别的分类性能研究,即设计具有高精度识别率和快速分类的算法等。基于稀疏表示的人脸识别技术,具有简单的理论基础和较好的鲁棒性。因此本文对稀疏表示人脸识别算法进行了研究,研究重点在特征提取和分类识别上,提出了一些新的人脸特征提取和识别算法。通过在人脸基准数据库上进行的大量实验,表明本文算法在人脸识别的计算效率和识别率上获得了良好效果。本论文的主要工作和贡献如下:(1)提出一种加权主成分分析特征提取算法。新算法首先通过线性拟合标记信息与特征维来对各特征加权,并通过稀疏约束使部分特征的权值为零,然后进行主成分分析特征提取。该方法实现了特征预选择并且突出了重要特征属性。实验结果表明,新算法不仅能够降低计算复杂度,还能提高分类的精度。(2)提出两种稀疏保持投影特征提取算法。一种是加权稀疏邻域保持投影,使用一个加权的稀疏重构模型去学习重构系数,并通过限制非零重构系数的个数,降低了时间复杂度,提高了识别精度和全局鲁棒性。另一种是基于聚类的无监督判别加权稀疏保持投影,区别于传统的稀疏保持投影方法,新算法将聚类与判别加权稀疏重构结合起来,通过聚类得到每个训练样本的标记,实现了无监督的判别性能,从而在提升简单性的同时提升了识别精度。(3)提出一种图嵌入的判别协同保持投影特征提取算法。本文提出了一种对分类器适应的特征提取算法,并将该模型融合到图嵌入框架。新算法使用协同表示构建类内和类间图,不仅避免了传统流形学习算法的参数寻优困难,而且继承了协同表示的鲁棒性。通过引入标记信息增加了算法的判别性能,从而提升了算法的识别性能。(4)提出一种加权稀疏表示分类器。稀疏表示分类器SRC通过重构误差来分类测试样本,但是它同等地对待每一个训练样本,不能体现样本之间的差异性。本文提出应用预先定义的重构模型的重构误差对每个样本赋予不同的权值,然后求解加权的稀疏重构模型。实验结果表明,新算法提高了识别精度并降低了时间复杂度。
[Abstract]:Face recognition is an important research field in the field of pattern recognition. Face images are affected by environment, illumination, expression and posture. Face recognition research is very challenging. How to quickly and accurately use the computer to detect and recognize faces is the key of current face recognition technology. Although some achievements have been made in face recognition research, there are still many problems and key technologies to be further solved and improved, including: face feature extraction adequacy research. That is, how to embed local and global structure information; Research on the classification performance of face feature recognition, that is, the design of high accuracy recognition rate and fast classification algorithm, etc. Face recognition technology based on sparse representation. It has simple theoretical foundation and good robustness. Therefore, this paper studies sparse representation face recognition algorithm, focusing on feature extraction and classification recognition. Some new face feature extraction and recognition algorithms are proposed, and a large number of experiments are carried out on the face reference database. The results show that the algorithm has achieved good results in the computation efficiency and recognition rate of face recognition. The main work and contribution of this paper are as follows: 1). A new feature extraction algorithm based on weighted principal component analysis (PCA) is proposed. Firstly, each feature is weighted by linear fitting marking information and feature dimension. The weight of some features is zero by sparse constraint, and then the feature extraction of principal component analysis (PCA) is carried out. This method realizes feature pre-selection and highlights important feature attributes. The experimental results show that. The new algorithm can not only reduce the computational complexity, but also improve the accuracy of classification.) two sparse preserving projection feature extraction algorithms are proposed, one is weighted sparse neighborhood preserving projection. A weighted sparse reconstruction model is used to learn the reconstruction coefficients, and the time complexity is reduced by limiting the number of non-zero reconstruction coefficients. The recognition accuracy and global robustness are improved. The other is the unsupervised discriminant weighted sparse preserving projection based on clustering, which is different from the traditional sparse preserving projection method. The new algorithm combines clustering with discriminant weighted sparse reconstruction to get the marks of each training sample by clustering to achieve unsupervised discriminant performance. Thus, a feature extraction algorithm of discriminant co-preserving projection based on graph embedding is proposed, and a feature extraction algorithm adapted to the classifier is proposed in this paper. The new algorithm uses cooperative representation to construct intra-class and inter-class graphs, which not only avoids the difficulty of parameter optimization of traditional manifold learning algorithm. Moreover, the robustness of cooperative representation is inherited, and the discriminant performance of the algorithm is improved by introducing label information. A weighted sparse representation classifier is proposed. The sparse representation classifier (SRC) classifies test samples by refactoring errors, but it treats each training sample equally. The reconstruction error of the pre-defined reconstruction model is proposed to assign different weights to each sample, and then to solve the weighted sparse reconstruction model. The new algorithm improves the recognition accuracy and reduces the time complexity.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】