基于压缩感知的人脸识别算法研究

发布时间：2018-04-18 22:03

本文选题：人脸识别 + 稀疏表示　；参考：《大连海事大学》2014年硕士论文

【摘要】：人脸识别是计算机技术研究领域的一项热门学科,它属于生物特征识别技术,利用生物个体自身特征实现对生物体的区分与识别。人脸识别由于自身的优越性以及在多媒体、模式识别、图像处理、计算机视觉等领域的广泛运用,近些年来人脸识别技术得到了长足的进步,大量的人脸识别系统应用在身份鉴别上。压缩感知理论是由Donoho与Candes等人于2006年提出的一个新的理论框架,很快被引入到人脸识别领域并引发了一阵研究热潮,该理论应用最成功的要数基于稀疏表示的人脸识别算法(Sparse Representation-based Classification, SRC)和已经存在的大多数算法相比较,SRC方法利用了高维数据分布的稀疏性特点,能够较为有效地应付人脸图像的高维数问题。此外,SRC方法处理的过程利用的是图像的原始像素,大大减小了一些预处理操作导致的信息丢失。但是,人脸姿态及表情的变化容易引起对齐误差,而现有SRC方法要求训练图像和测试图像严格对齐,这也影响了SRC方法的性能,成为了阻碍SRC方法走向实用的主要因素。本论文首先研究了基于稀疏表示的人脸识别算法,介绍了传统人脸识别中常用的PCA、LDA方法并进行了一些仿真试验,然后通过引入压缩感知理论实现了基于稀疏表示的人脸识别算法并取得了不错的识别效果。论文给出了上述稀疏表示方法的基本原理、实现方法,并且用该方法及传统方法对ORL人脸数据库进行了仿真,分别计算出了识别率,比较和揭示了这些方法之间的区别和联系。在上述研究的基础上,针对SRC方法的固有缺点实现了基于两阶段稀疏表示的人脸识别,通过在实验证明两阶段稀疏表示方法相对于传统SRC方法的优势,然后针对两阶段稀疏表示方法应用过程中常常训练样本不足的问题实现了基于改进人脸库的两阶段稀疏表示方法。本文通过在YALE、FERET和AR库上的仿真实验证明了的本方法的有效性,同时该方法对最邻近样本数的依赖有所减弱,鲁棒性有所增强。
[Abstract]:Face recognition is a hot subject in the field of computer technology. It belongs to biometric recognition technology. It uses biological individuals' own characteristics to distinguish and recognize organisms.Due to its superiority and its wide application in multimedia, pattern recognition, image processing and computer vision, face recognition technology has made great progress in recent years.A large number of face recognition systems are used in identification.The theory of compressed perception is a new theoretical framework proposed by Donoho and Candes et al in 2006. It was quickly introduced into the field of face recognition and triggered a wave of research.In this theory, the most successful face recognition algorithm based on sparse representation is Sparse Representation-based Classification (SRCCs), which makes use of the sparsity of high-dimensional data distribution in comparison with most existing algorithms.It can deal with the problem of high dimension of face image effectively.In addition, the SRC method uses the original pixels of the image, which greatly reduces the loss of information caused by some preprocessing operations.However, the change of face pose and expression is easy to cause alignment error, and the existing SRC methods require strict alignment of training and test images, which also affects the performance of SRC method and becomes the main factor that hinders the application of SRC method.In this paper, the algorithm of face recognition based on sparse representation is studied, and the PCA-LDA method, which is commonly used in traditional face recognition, is introduced, and some simulation experiments are carried out.Then a face recognition algorithm based on sparse representation is implemented by introducing the theory of compressed perception and good recognition results are obtained.In this paper, the basic principle and realization method of the above sparse representation methods are given, and the ORL face database is simulated by this method and the traditional method. The recognition rate is calculated, and the differences and relations between these methods are compared and revealed.On the basis of the above research, face recognition based on two-stage sparse representation is realized in view of the inherent shortcomings of SRC method. It is proved by experiments that two-stage sparse representation method is superior to traditional SRC method.Then, a two-stage sparse representation method based on improved face database is implemented to solve the problem of insufficient training samples in the application of two-stage sparse representation.In this paper, the effectiveness of the proposed method is proved by simulation experiments on the Yaalehnet and AR libraries, and the dependence of the method on the nearest sample number is weakened, and the robustness is enhanced.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41

【参考文献】