人脸识别中基于图像局部结构的特征提取与分类研究

发布时间：2018-01-15 20:23

本文关键词：人脸识别中基于图像局部结构的特征提取与分类研究　出处：《南京理工大学》2015年博士论文　论文类型：学位论文

【摘要】：过去几十年中,人脸识别一直是计算机视觉和模式识别领域一个研究热点,因而吸引了大量关注并取得了巨大进展。但大多数人脸识别方法都严重依赖于训练样本的个数,在面临单样本人脸识别问题时性能严重下降甚至无法工作。为了解决单样本人脸识别问题,本文从人脸图像的局部结构出发提出了新的特征提取和分类算法,所提出的算法不仅获得了很好的识别效果而且对表情、光照、遮挡、时间变化也具有鲁棒性。主要研究工作和创新性成果如下：(1)提出了基于心理学定律的局部特征提取算法—韦伯局部二元模式(WLBP)。 WLBP包含两个部分：差励(Differential Excitation,DE)算子和LBP算子,差励算子根据韦伯定律模拟人类对图像的感知提取感知特征,而LBP算子可以有效地提取图像局部特征,通过二维直方图将两种特征结合从而形成更具鉴别性的统计直方图特征。差励算子提取的感知特征通过使用LoG(Laplacian of Gaussian)算子有效地提升了WLBP对噪声、光照等变化的鲁棒性。此外,WLBP还成功应用于了纹理分类、人脸识别、人眼状态检测等应用中。(2)提出了基于局部特征的单样本人脸识别。为解决单样本人脸识别问题的特例—草图人脸识别问题,我们提出了局部几何(LG)特征提取算法,假设人脸草图与人脸照片中的局部块在不同的图像空间形成了具有相似几何结构的流形,利用图像局部块之间的几何结构作为特征可以将不同的模态的人脸草图和人脸照片转换到了同一特征空间中,从而降低草图与照片之间的模态差异。此外,本文还将局部几何特征、WLBP、LBP、WLD等局部特征分别应用于单样本人脸识别中,与全局特征相比局部特征对单样本人脸识别问题更加鲁棒。(3)提出了基于局部相似性的子空间学习算法。PCA、LDA、LPP等传统的子空间学习算法在单样本情况下性能下降严重甚至无法工作问题,为解决此问题本文将人脸图像分块并对每个局部块单独分类然后合并所有块的分类结果。为了分类每个局部块,我们将每个局部块进一步划分为重叠的子块并根据子块间的局部相似性提出了一系列基于局部相似性的子空间学习算法：基于局部相似性的线性鉴别分析(LS_LDA)、基于局部相似性的中值费舍尔鉴别器(LS_MFD)、基于局部相似性的主成分分析(LS_PCA)、基于局部相似性的局部保持投影(LS_LPP)、基于局部相似性的边界费舍尔分析(LS_MFA)。相比于PCA、LDA、LPP、MFA等传统的子空间学习算法,LS_PCA、LS_LDA、LS_MFD、LS_LPP和LS_MFA不仅解决了PCA、LDA, LPP和MTA在单样本情况下性能下降或无法工作的问题而且对光照、表情、遮挡、时间变化具有很好的鲁棒性。此外,本文还将上述基于局部相似性的子空间学习方法都统一到了基于局部相似性的线性图嵌入(LS_LGE)算法框架中。LS_LGE具有很好的泛化能力可以根据需要泛化出更多的基于局部相似性的子空间学习算法。(4)提出了基于局部结构的稀疏表示分类算法和基于局部结构的协同表示分类算法。针对稀疏表示分类(SRC)在单样本情况下性能下降问题,本文提出了基于局部结构的稀疏表示分类(LS_SRC),将人脸图像划分为包含重叠子块的局部块并假设重叠的子块位于同一线性子空间内。子空间假设不仅反映了局部结构还保证了SRC适用于单样本人脸识别问题。为了提升LS_SRC的计算效率,本文进一步提出了基于局部结构的协同表示分类(LS_CRC)算法。考虑到不相关的训练样本对协同表示的干扰而影响分类性能,本文接着提出了基于局部结构的两阶段协同表示分类(LS_TPCRC)和基于局部结构的多阶段协同表示分类(LS_MPCRC),通过样本选择机制去除无用的训练样本,可以在保持协同表示性能的同时引入了有监督的稀疏性。此外,为了进一步提升LS_SRC和LS CRC性能,本文还提出了基于混淆矩阵的贝叶斯推理,利用混淆矩阵描述分类器误差并根据贝叶斯推理推导出测试样本属于各类的更加准确的概率。
[Abstract]:Over the past few decades, face recognition has been a hot research field of computer vision and pattern recognition, which has attracted a lot of attention and made great progress. But the number of most face recognition methods rely heavily on training samples, in the face of single sample face recognition problem when the performance decline even not work. In order to solve the single the sample problem of face recognition from the local structure of face images is proposed based feature extraction and classification algorithm, the proposed algorithm not only achieves good recognition effect and the expression, illumination, occlusion, time change is also robust. The main research work and innovative achievements are as follows: (1) proposed the extraction algorithm of Webb local two yuan local features based on the law of Psychology (WLBP). The WLBP consists of two parts: differential excitation (Differential Excitation, DE and LBP) operator The difference, according to Webb's law to simulate the human operator excitation perception on the extraction of image perception features, while the LBP operator can effectively extract the local features of images, the two-dimensional histogram will thus form histogram feature is more discriminative features. Combining the two difference perception features extracted by using LoG operator excitation (Laplacian of Gaussian) operator to enhance the robustness of WLBP noise and illumination change. In addition, WLBP also successfully applied to texture classification, face recognition, eye state detection application. (2) proposed a single sample face recognition based on local features. As a special case of sketch face recognition problem solving single sample face recognition problem and we propose a local geometry (LG) feature extraction algorithm, local block hypothesis face sketch and photographs of human faces in the image space of different form with similar geometric structure of the manifold, The local geometric structure of image blocks between different modes can be characterized as the face and face sketch photo conversion into the same feature space, thereby reducing the differences between the sketch and photo mode. In addition, this paper will also local geometric features, WLBP, LBP, WLD and other local features are respectively applied to the single sample face recognition compared with the global features, local features of the single sample face recognition problem is more robust. (3) is proposed based on local similarity subspace learning algorithm of.PCA, LDA, LPP and other traditional subspace learning algorithm performance in single sample case serious decline and even unable to work, in order to solve this problem this paper will face the image blocks and each block separately and then merge the local classification classification results of all blocks. In order to classify each local block, we will each local block is further divided into overlapping blocks according to the The local similarity between blocks and put forward a series of local similarity based on subspace learning algorithm based on local similarity of linear discriminant analysis (LS_LDA), based on local similarity value (LS_MFD), Fisher discriminator based on local similarity of principal component analysis (LS_PCA), based on local similarity preserving local the projection (LS_LPP), based on local similarity analysis of the boundary of Fisher (LS_MFA). Compared with PCA, LDA, LPP, MFA and other traditional subspace learning algorithm, LS_PCA, LS_LDA, LS_MFD, LS_LPP and LS_MFA not only solve the PCA, LDA, LPP and MTA in a single sample worse performance or unable to work and the problem of light, expression, occlusion, robust time changes. In addition, this paper also the similarity of the subspace based on local learning methods have the same based on local similarity of linear graph embedding (LS_LGE) algorithm framework .LS_LGE has good generalization ability can be generalization more local similarity based on subspace learning algorithm. (4) propose a classification algorithm based on local structure and collaborative representation classification algorithm based on sparse representation of local structure. According to the sparse representation classification (SRC) in a single sample case performance decline. This paper proposed a classification based on sparse local structure (LS_SRC), the face image is divided into local block overlapping sub block and assuming the overlapping sub block is located in the same line of subspace. The subspace assumption not only reflects the local structure but also ensure that the SRC is suitable for single sample face recognition problem. In order to enhance computing efficiency LS_SRC, this paper proposes a local structure based on Collaborative representation classification (LS_CRC) algorithm. Considering the interference is not related to training samples and the influence of collaborative representation classification Can, then the paper presents the collaborative representation classification based on two stage local structure (LS_TPCRC) and multi stage local structure based on Collaborative representation classification (LS_MPCRC), through the sample selection mechanism of removing useless training samples, can keep said performance at the same time we introduce some sparse supervision. In addition, in order to further improve the LS_SRC the performance of CRC and LS, this paper also put forward the Bias reasoning based on confusion matrix, using confusion matrix description and classification error are derived according to the Bias reasoning test sample belongs to all kinds of more accurate probability.

【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】