【作者】 王智飞；

【导师】 苗振江；

【作者基本信息】 北京交通大学 ， 人机交互工程， 2013， 博士

【摘要】 经过四十多年的发展,人脸识别的各种算法层出不穷,从研究初期只针对单一简单背景发展到目前应对各种复杂条件,如姿态、光照、表情、噪声、遮挡、化妆、年龄、种族、性别等。尽管已有的人脸识别系统在特定约束环境下的正确识别率令人满意,但在实际环境尤其在视频监控应用中,由于监控对象的不配合及距离监控摄像头较远等问题引起图像质量较低,导致识别性能很不理想,我们把这种情形下的人脸识别统称为低分辨率人脸识别。本文针对远距离监控带来的人脸小尺寸和低质量问题,对低分辨率人脸识别算法进行系统研究,全面综述低分辨率人脸识别算法的研究现状与发展趋势,重点对分辨率稳健特征表达能力的局限性、高低分辨率统一特征空间表达能力的不足、人脸超分辨率增强与识别目标的不一致等三个关键问题进行深入研究,并提出若干新的模型和算法,为自动人脸识别系统走向实际应用提供理论依据和技术方法。本文的主要贡献如下：(1)首次对低分辨率人脸识别算法进行全面的综述。给出低分辨率人脸识别的相关概念,并从整体上归纳解决问题的系统框架和基本策略。提出超分辨率增强和分辨率稳健特征表达等两大类、面向视觉&面向识别的超分辨率增强、基于特征&基于结构的分辨率稳健特征表达等四小类的分类体系,并对代表性算法对比分析和综合评测。总结已有算法存在的问题并给出将来的研究方向。(2)提出一种基于特征的分辨率稳健特征表达算法：基于分辨率级差概率准则的图嵌入算法(FisherNPE),致力于解决分辨率稳健特征表达能力的局限性问题。所提算法包括两个重要模块：基于FisherNPE算法的特征提取模块和基于分辨率级差概率准则的特征分类模块。所提特征提取模块将线性鉴别分析(LDA)的全局描述能力和邻域保持嵌入(NPE)的局部描述能力结合,按照一定的权重因子加权形成一个新的关系权重矩阵,构建FisherNPE算法的图嵌入模型,实现分辨率稳健特征提取。所提特征分类模块创造性地引入分辨率级差的概念,构建高低分辨率样本差值图像的特征空间,并结合贝叶斯概率准则,将传统“一对一”的特征分类模式拓展到“多对多”模式,极大地提高了可用于分类的有效特征数量。实验结果表明：与LDA、NPE、BayesianFace等已有算法相比,所提算法在ORL、YALE、 CMU PIE三个数据库单一分辨率和多重分辨率上的识别性能均有较大的提升,尤其在超低分辨率上如7×6、5×5、8×8等,识别率平均提高10%。(3)提出一种基于结构的分辨率稳健特征表达算法：基于核耦合交叉回归的分辨率空间映射算法(KCCR),致力于解决高低分辨率统一特征空间表达能力的不足问题。结合核方法在描述非线性空间的强大能力和谱回归理论带来的低计算复杂度的优势,创造性地提出耦合交叉回归的思想,构建高低分辨率样本的低维嵌入与原始数据样本之间的交叉映射关系。所提算法不仅充分利用高分辨率样本之间的邻近关系,也利用了低分辨率样本之间的邻近关系,有效地提升统一特征空间的表达能力。实验结果表明：所提的CCR/KCCR及其改进算法ICCR/KICCR基于单一核和组合核在FERET和CMU PIE两个数据库上进行测试,与CLPMs、 CLDMs、KECLPMs等已有耦合映射相关算法相比,其不仅在超低分辨率如8×8的识别性能上有了较大的提升,识别率平均提高6%,更是将计算复杂度由立方级下降到平方级甚至线性级,提升了系统运行速度。(4)提出一种面向识别的人脸超分辨率增强算法：基于张量特征转换的分辨率增强算法(TET),致力于解决人脸超分辨率增强与识别目标的不一致问题,并从系统构建上考虑高分辨率人脸的获取问题。所提算法包括两个重要模块：面向分辨率增强的远距离人脸检测模块和基于张量特征转换的分辨率增强模块。所提远距离人脸检测模块结合改进肤色模型如H-SV和C’bC’r在较远距离和改进AdaBoost算法在较近距离检测上的互补优势,实现在室内环境中不同距离条件下姿态和光照稳健的人脸检测,为分辨率增强模块提供输入样本。所提分辨率增强模块利用张量分析进行特征转换,并引入奇异值分解特征,构建输入样本在高低分辨率训练样本中的组合权重分布,得到全局增强人脸,并引入残余补偿技术,构建高分辨率残余信息,实现正面视角和良好光照的人脸分辨率增强,并改善了增强图像的识别性能。此外,在室内环境中对本文所提算法(FisherNPE、KCCR、 TET)和两种代表性的算法(CLPMs和S2R2)进行综合评测,实验结果表明：本文所提算法在距离(分辨率)、光照、配准等条件变化下取得较好的性能。更多还原

【Abstract】 Face recognition (FR) has been widely studied for over40years due to its great potential applications. Currently, the advance of technology is able to perform the recognition of complex conditions such as pose, illumination, expression, noise, disguise, aging, race, gender and so on, which is a significant improvement than decades ago. Although the accuracy of face recognition for subjects under controlled conditions can be reached in a satisfactory manner, the performance in real applications such as surveillance remains a challenge. The low-resolution (LR) of the images caused by unperceived surveillance where subjects are far away from the cameras and face regions tend to be small is the major issue, which is classified as low-resolution face recognition (LR FR) problems in the area of study.In this dissertation, LR FR algorithms were systematically investigated for the purpose of solving small size and low quality problems brought up by surveillance at-a-distance. A comprehensive literature review on LR FR algorithms was conducted. Three key issues about LR FR system, namely the limitation of resolution-robust feature representation, the deficiency of unified feature space between LR and HR, and the inconsistency between enhancement and recognition based on super-resolution, were explored in depth. More importantly, a set of new models and algorithms were developed and proposed in the chapters. The major contributions of this dissertation are summarized below:(1) A comprehensive literature review was conducted for LR FR systems, which is the first attempt in this area. First, it gave an overview on LR FR, including concept description, system architecture and algorithm categorization. Second, many representative algorithms were broadly reviewed and discussed. The algorithms were classified into two different categories, super-resolution for LR FR and resolution-robust feature representation for LR FR, which are further classified into four small groups, namely vision-oriented&recognition-oriented super-resolution and feature-based&structure-based resolution-robust feature representation. Their strategies, advantages and disadvantages were discussed in detail. Some relevant issues such as databases and evaluations for LR FR were presented as well. By generalizing their performances and limitations, promising trends and crucial issues for future research were summarized. (2) A new feature-based resolution-robust feature representation algorithm was proposed, namely a graph embedding algorithm (FisherNPE) based on resolution level difference probabilistic similarity measure, which aims to improve the limitation of resolution-robust feature representation. The proposed algorithm includes two modules: feature extraction based on FisherNPE algorithm and feature classification based on resolution level difference probabilistic similarity measure. The feature extraction module buids a new relationship weight matrix by using a weight factor to combine the global and local relationship representation ability of LDA and NPE respectively, which is embeeded into a new graph embedding model FisherNPE, thus realize resolution-robust feature extraction. The feature classification module creatively introduces the concept of resolution level difference combining with the probabilistic similarity measure to build the feature spaces of the differences between HR and LR face images. It would expand the traditional way of "one-to-one" into "many-to-many" in feature classification and greatly incease the amouts of the effective features. Finally, the proposed algorithm was comprehensively tested on three public databases namely ORL, YALE, CMU PIE with the cases of single resolution and multiple resolutions. Experimental results showed that the proposed algorithm had improved recognition performance by10%, especially on the very low resolutions such as7×6.5×5.8×8. in comparison with the traditional algorithms such as LDA, NPE, BayesianFace. And the system maintains a relatively stable performance.(3) A new structure-based resolution-robust feature representation algorithm was proposed, namely a LR and HR feature space match algorithm based on kernel coupled cross-regression (KCCR), which aims to improve the deficiency of the unified feature space. The idea of kernel coupled cross-regression was proposed for building the coupled mappings between LR and HR feature space based on the advantages in the power ability of kernel technique to describe nonlinear space and the low computational complexity produced by spectral regression theory, which is a creative point in this study. Meanwhile, cross-regression was developed to build the coupled mappings between the low-dimensional embeddings of HR/LR samples and themselves. The proposed algorithm not noly utilize the relationships of HR samples but also ones of LR samples, thus effectively improve the representation ability of the unified feature space. Finally, the proposed CCR/KCCR and its improved algorithm ICCR/KICCR were comprehensively tested on two public databases named FERET and CMU PIE with the cases of single kernel and compound kernel. Experimental results showed that the proposed algorithm had a great improvement on the very low resolution such as8×8in recognition performance with the average improvement of6%, in comparison with the existing related algorithms based on coupled mappings model such as CLPMs, CLDMs, KECLPMs. Moreover, the computational complexity of the system was declined from cubic level to square level and even linear level. Thus, the speed of the system was greatly enhanced.(4) A new recognition-oriented face super-resolution algorithm was proposed, namely a resolution enhancement algorithm based on tensor eigen-transformation (TET) with the consideration of the acquisition of HR face image at-a-distance, which aims to improve the inconsistency between enhancement and recognition. The proposed algorithm includes two modules:resolution enhancement-oriented face detection at-a-distance and resolution enhancement based on tensor eigen-transformation. The proposed detection module combines the improved skin models such as H-SV and C’bC’r color spaces and the improved AdaBoost algorithm to realize face detection at-a-distance with pose and illumination variation cases, thus provide the input face samples for the next module. The proposed resolution enhancement module uses tensor analysis to perform eigen-transformation based on the feature vectors of singular value decomposition. It would build the combining weight distribution of the input LR face samples in the HR/LR training samples, thus achieve the global enhanced face images. Face residue compensation was then used for further processing the enhanced face images. Resolution enhancement with frontal view and good illumination cases was finally achieved, thus being beneficial to improve the recognition performance. In addition, the proposed three algorithms in this dissertation (FisherNPE, KCCR, TET) and two representative algorithms (CLPMs and S2R2) were comprehensively tested on indoor environment condition. Experimental results demonstrated that the proposed three algorithms achieved relatively better performance under the various combinations of distance (resolution), illumination and alignment.更多还原