基于稀疏表示的人脸识别研究

发布时间：2018-06-03 01:33

本文选题：人脸识别 + 稀疏表示　；参考：《安徽大学》2016年硕士论文

【摘要】：人脸识别是生物特征识别中的一项重要技术,也是图像处理,机器学习等领域的热点研究课题,在公安系统,保险,银行,海关,身份证系统等领域具有广阔的应用前景。不同于传统的Eigenface, fisherface算法,2009年Wright等人提出了基于稀疏表示的人脸识别(Sparse Representation-based Classification, SRC),由于其对噪声有较好的鲁棒性,在人脸识别上取得巨大的成功,并将人脸识别引入了一个新的发展方向。本文主要对稀疏表示的人脸识别进行了研究,论文的主要工作如下：(1)加权稀疏近邻表示的人脸识别(WSNRC)。在每一类训练样本中寻找与测试样本最近的k个样本构成此类新的训练字典,然后在求解l1范数最小化的稀疏系数时,为每一个新的训练样本对应的稀疏系数赋上一个权值；最后在新的字典下,根据最小的重构误差来完成识别任务。在Yale B数据库和ORL数据库上的大量实验结果表明,WSNRC与NN算法和稀疏近邻表示(SNRC)算法相比,取得了较高的识别率,证明了该方法的有效性。(2)正则化Fisher分析和稀疏表示的人脸识别。首先使用正则化Fisher分析算法从训练样本中提取出最优投影矩阵,然后将训练样本和测试样本在投影矩阵下投影获得其低维表示,最后使用稀疏表示分类器进行人脸识别,在AR数据库和扩展的YaleB数据库上的大量实验结果表明,正则化Fisher分析和稀疏表示结合的方法取得较好的效果。(3)结合Gabor特征和对称脸的稀疏表示人脸识别。首先根据原始训练样本获得其对应的虚拟对称脸,然后将原训练样本和对称脸结合起来构成新的训练样本,最后提取训练样本和测试样本的Gabor特征并使用SRC分类器进行人脸识别。在ORL数据库,Yale数据库,FERET数据库上的实验表明了GMSRC的有效性。
[Abstract]:Face recognition is an important technology in biometrics. It is also a hot research topic in image processing, machine learning and other fields. It has broad application prospects in public security system, insurance, bank, customs, identity card system and other fields. It is different from the traditional Eigenface, Fisherface algorithm, and in 2009, Wright and others proposed a sparse table. Sparse Representation-based Classification (SRC), because of its good robustness to noise, has achieved great success in face recognition, and introduces face recognition to a new direction. This paper mainly studies the face recognition of sparse representation. The main work of this paper is as follows: (1) weighting Face recognition (WSNRC) of sparse near neighbor representation. In each class of training samples, searching for the nearest K sample with the test sample constitutes such a new training dictionary, and then a weight is assigned to the sparse coefficient corresponding to each new training sample when the sparse coefficient of the L1 norm minimized, and finally under the new dictionary, according to the minimum. A large number of experimental results on the Yale B database and the ORL database show that WSNRC has a higher recognition rate compared with the NN algorithm and the sparse nearest neighbor representation (SNRC) algorithm. (2) the regularized Fisher analysis and the sparse representation of the face recognition. First, the regularized Fisher is used. The algorithm extracts the optimal projection matrix from the training sample. Then the training sample and the test sample are projected under the projection matrix to obtain their low dimensional representation. Finally, the sparse representation classifier is used for face recognition. A large number of actual results on the AR database and the extended YaleB database show that the regularized Fisher analysis and sparse representation are regularized. The combined method has good results. (3) face recognition based on the sparse representation of Gabor features and symmetrical faces. First, the corresponding virtual symmetry faces are obtained according to the original training samples. Then the original training samples and symmetrical faces are combined to form a new training sample. Finally, the Gabor features of the training samples and the test samples are extracted and the SRC is used. The classifier is used for face recognition. Experiments on ORL database, Yale database and FERET database show the effectiveness of GMSRC.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】