基于协同表示的人脸识别研究

发布时间：2018-06-13 16:59

本文选题：人脸识别 + 协同表示　；参考：《山东大学》2017年硕士论文

【摘要】：随着社会的快速发展,信息安全问题逐渐得到人们的重视,如何进行准确、快速的身份识别已成为一项亟待解决的重要课题。人脸识别技术利用每个人独有的面部特征进行身份识别,与基于其他生物特征的身份识别技术相比,具有采集简单方便、非侵入性、符合人类认知习惯等优点。人脸识别技术是模式识别和人工智能等领域的研究热点,并且在公共安全及经济金融方面显示出了巨大的应用前景。人脸识别技术虽然取得了许多突破,但在不受控条件下的识别效果并不理想,仍然具有高度的挑战性。本文首先介绍了人脸识别研究的背景、意义、研究现状和一般的算法流程,接着主要研究了人脸识别中的特征提取和分类方法。其中重点研究了稀疏表示(SRC)理论,针对其最小化l,范数的计算复杂度过高的问题,利用l2范数代替l,范数,引入协同表示(CRC)算法,大大降低了计算成本。通过对这两种算法的对比研究,发现SRC与CRC算法的目标都是利用训练样本实现对测试样本的最优重构,再由编码系数判断出测试样本的类别。由于协同表示算法的计算复杂度低,实时性强,研究协同表示的改进算法更适于实际应用。因此,本文针对协同表示方法的不足之处,进行了一系列的改进。本文的主要创新点如下:(1)由于训练样本中不可避免的存在一定的噪声误差,会在协同表示算法重构测试样本时引起差错。针对这一缺点,本文利用低秩矩阵恢复(LR)算法处理人脸图像,减弱了人脸图像中的噪声误差,突出同类人脸的共同特征,以便更好的重构出测试样本。在计算低秩逼近矩阵时,利用非精确的拉格朗日乘子法(IALM)代替精确的拉格朗日乘子法(EALM),降低了计算难度,提升了运行效率,使算法更具实用性。(2)协同表示算法忽略了近邻样本在重构测试样本时的重要作用。针对这一缺点,将局域一致性引入编码方案中,使得相似的样本(近邻样本)的编码系数尽可能的相似。低秩恢复可以有效地去除原始图片中的噪声和误差,突出同类的人脸图片的共同特征,使得同类人脸图片的距离尽可能的相近。将这两种方法相结合,提出了基于低秩恢复和局域约束协同表示(LR-LCCR)的人脸识别算法,能够使得同类人脸图像的编码系数相似,增强鲁棒性。在AR和Extend Yale B人脸库上的实验证明了算法的良好性能。(3)稀疏表示和协同表示一般利用图像的全局特征构成训练库字典。由于全局特征对人脸图像的描述并不充分,所以针对协同表示的字典对其进行了改进。详细阐述了一种提取人脸局部特征的有效算法,方向边缘幅值模式(POEM)。利用人脸图像的POEM特征构成协同表示的字典,提出了基于POEM特征的协同表示(POEM-CRC)算法。在AR和FERET人脸库上的实验证明了 POEM-CRC算法的性能优于其他同类算法。
[Abstract]:With the rapid development of society, people pay more and more attention to the problem of information security. How to identify accurately and quickly has become an important task to be solved urgently. Compared with other biometric identification techniques, face recognition technology has the advantages of simple and convenient collection, non-invasive, consistent with human cognitive habits and so on. Face recognition is a hot topic in pattern recognition and artificial intelligence, and it has shown great application prospect in public security and economic and financial fields. Although many breakthroughs have been made in face recognition technology, the recognition effect under uncontrolled conditions is not satisfactory, and it is still highly challenging. This paper first introduces the background, significance, research status and general algorithm flow of face recognition, and then mainly studies the feature extraction and classification methods in face recognition. The theory of sparse representation (SRC) is mainly studied in this paper. Aiming at the problem of high computational complexity of minimization of l and norm, we use l 2 norm instead of l norm and introduce cooperative representation (CRC) algorithm, which greatly reduces the computation cost. Through the comparative study of the two algorithms, it is found that both SRC and CRC algorithms aim at the optimal reconstruction of test samples using training samples, and then determine the types of test samples from the coding coefficients. Because of its low computational complexity and high real-time performance, the improved collaborative representation algorithm is more suitable for practical application. Therefore, this paper makes a series of improvements to the shortcomings of collaborative representation. The main innovation of this paper is as follows: (1) due to the inevitable noise error in the training samples, it will cause errors in the reconstruction of test samples by the cooperative representation algorithm. Aiming at this shortcoming, this paper uses the low rank matrix restoration algorithm to process face image, reduces the noise error in the face image, highlights the common features of the same kind of face, so as to better reconstruct the test sample. In the calculation of low rank approximation matrix, the inexact Lagrange multiplier method (IALM) is used instead of the exact Lagrange multiplier method to reduce the calculation difficulty and improve the running efficiency. In order to make the algorithm more practical, the cooperative representation algorithm ignores the important role of neighbor samples in reconstructing test samples. To solve this problem, local consistency is introduced into the coding scheme to make the coding coefficients of similar samples (nearest neighbor samples) as similar as possible. Low-rank restoration can effectively remove the noise and error in the original image, highlight the common features of the same kind of face image, and make the distance of the same face image as close as possible. By combining these two methods, a face recognition algorithm based on low rank restoration and local constraint cooperative representation (LR-LCCR) is proposed, which can make the coding coefficients of similar face images similar and enhance the robustness. Experiments on AR and extend Yale B face database show that the algorithm has good performance. The sparse representation and cooperative representation generally use the global features of images to form a training library dictionary. Because the global feature is not sufficient to describe the face image, the dictionary for cooperative representation is improved. In this paper, an effective algorithm for extracting face local features, directional edge amplitude mode, is described in detail. Based on Pom features of face images, a dictionary of cooperative representation is proposed, and a Poem-based collaborative representation algorithm is proposed. Experiments on AR and Feret face database show that the performance of POEM-CRC algorithm is better than that of other similar algorithms.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】