【作者】 陈欣；

【导师】 张洪斌；

【作者基本信息】 电子科技大学 ， 电子与通信工程（专业学位）， 2015， 硕士

【摘要】 随着智能信息技术的快速发展,人工智能正在逐渐改变人们的生活方式。人脸识别作为人工智能领域的一个重要分支,近年来受到广泛的关注和研究。在一些实际应用场景中,每个人只能获得一张人脸图像作为训练样本,但大多数人脸识别方法在单样本条件下识别率低,甚至有些根本不再适用。本文研究单样本条件下的人脸识别,从Gabor小波变换入手,针对光照、表情和遮挡等可变因素寻求有效的解决方案。本文主要的工作如下:1.首先介绍了一维、二维Gabor小波变换,分析了二维Gabor小波函数中的参数对Gabor核函数的影响,并给出了提取人脸图像的Gabor小波特征的具体过程。2.针对人脸图像中的可变光照问题,利用无直流分量的DCT系数作为特征,以消除线性光照的影响,再将提取的DCT特征与最近邻分类器相结合进行分类。同时将Gabor小波特征与最近邻分类器相结合,利用Gabor小波的局部描述特性,增强识别方法对非线性光照变化的适应能力。实验证明,两种方法都对一定程度上的光照变化有很好的鲁棒性。3.根据通用学习方法的核心思想,将人脸图像的Gabor小波特征用于自适应回归分类(ALRC),利用主成分分析(PCA)对Gabor特征进行降维。并提出一种改进的增强类模型(ECM)方法,利用与测试人脸图像有最相似变化的辅助样本子集提取通用人脸变化特征,再与单幅训练人脸图像一起构成类模型,最后使用线性回归分类(LRC)方法进行分类识别。实验表明,所提方法对人脸图像中的光照、表情和遮蔽变化有很好的鲁棒性,并且构建增强类模型所需要的辅助样本图像数量比ALRC方法中的更少。4.根据核方法的基本思想,对非线性回归分类方法进行改进,提出一种基于Gabor小波特征的自适应非线性回归分类方法(GANRC),首先提取测试人脸图像的Gabor小波特征,接着构建自适应类模型,再利用核函数将测试人脸图像的Gabor特征和自适应类模型映射到高维空间,最后根据最小距离准则判断测试人脸图像所属的类。实验证明,该方法对人脸图像中的表情、光照和遮蔽变化有很好的鲁棒性,并且在特征维数较少的情况下就能达到较高的识别准确率。更多还原

【Abstract】 With the rapid development of information technology, artificial intelligence is gradually changing the way people live. As an important branch of artificial intelligence, face recognition has been extensively reviewed and researched in recent years. In some practical application scenes, each person can only gets one face image as the training sample,but most of face recognition methods will get low recognition rate under the single sample condition, some of them will even simply no longer apply. This paper focuses on research of single sample face recognition, which starts from the Gabor wavelet transform,and devises effective solutions in the light of the variable factors such as illumination,facial expression and occlusion. The main work of this paper are listed as follows:1. First of all, the one-dimensional and two-dimensional Gabor wavelet transform are introduced in this paper, and the influence of the parameters of two-dimensional Gabor wavelet function on the Gabor kernel function are also analyzed, then the specific process of Gabor wavelet feature extraction of face images are presented.2. Aiming at the variable illumination of face image, the DCT coefficients with no DC component as features is used to remove the influence of linear light, then the extractive DCT features are combined with nearest neighbor classifier for classification.Meanwhile, the Gabor wavelet features and the nearest neighbor classifier is combined,which utilize the local describe characteristic of Gabor wavelets to enhance the adaptability of recognition approach against the nonlinear illumination variation. As the experiment proved, all the two methods have good robustness to a certain degree of illumination variation.3. According to the core idea of general learning method, Gabor wavelet features is adopted into adaptive linear regression classification(ALRC) scheme, which use the principle component analysis(PCA) to reduce the dimension of Gabor feature. And an improved method based on enhanced class model(ECM) is put forword. The method uses the auxiliary sample subset which has the most similar variations with test face images to extract the generic face variation characteristics, then constructs the class model with both the single training sample and the generic face variation characteristics. Finally, the linear regression classification approach is used to classify and recognize. Experiments shows that the methods proposed have good robustness to the facial variations caused by illumination, expression, and occlusion. In addition, the auxiliary sample images needed to construct the enhanced class model are less then that needed in ALRC.4. According to the basic idea of kernel method, the nonlinear regression classification method is improved, and the Gabor wavelet feature based adaptive nonlinear regression classification(GANRC) approach is proposed. The approach firstly extracts the Gabor wavelet feature of the test face image, then constructs the adaptive class model,and then maps the Gabor wavelet feature of the test sample and adaptive class model to a high dimensional space, finally determines the class which the test face image belongs to according to the minimum distance criterion. As the experiment proved, the method proposed has good robustness to the facial variations caused by illumination, expression,and occlusion, which can achieve high recognition accuracy rate in the case of relatively less feature dimension.更多还原