【作者】 李欣；

【导师】 王科俊；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2011， 博士

【摘要】 单训练样本人脸识别,是指每人仅存储一幅人脸图像作为训练集去识别姿态、光照等可能存在变化的人脸图像的身份。由于单训练样本问题给人脸识别系统带来的巨大挑战及本身具有的重要意义,它已成为人脸识别研究中的一个重要的研究方向。本文分析了单训练样本人脸识别的研究现状,指出样本扩张和多特征融合是解决单训练样本问题的有效途径；本文主要在理论和算法上对二维以及双模态生物特征融合的单训练样本人脸识别进行深入的研究,重点研究二维人脸图像特征提取以及2D人脸与3D鼻型双模态信息融合。论文的主要工作有：1.本文对单训练样本人脸识别的样本扩张技术进行研究,提出了应用于的单训练样本人脸识别的基于单张正面照片的三维人脸模型合成方法,利用合成的三维人脸模型生成姿态、光照、表情变化的虚拟人脸图像,以达到扩充训练样本库的目的,从而将单训练样本问题转化为多训练样本人脸识别。由二维图像恢复对象的三维模型是计算机视觉领域的一个基本问题,传统的方法需要多幅人脸图像、图像序列,或者需要限定条件下的立体图像对、正面和侧面图像对等,不利于实际应用。本文提出的算法只需要一幅正面人脸图像,降低了对使用条件的要求,便于实际应用,具有广阔的应用前景,该方法合成的三维人脸模型可以满足人脸识别、表情动画、人机交互的需要。基于虚拟图像的人脸识别实验结果表明,样本扩张法可以有效地解决单训练样本人脸识别中光照,姿态,表情变化的问题。同时,从生成的三维人脸模型上提取三维鼻型并利用三维鼻型进行身份识别,实验论证了3D鼻型作为一种新兴的生物特征识别模式的可行性。2.从提高单一模态——二维人脸识别的识别性能的角度出发,对传统的二维人脸特征提取算法进行改进,将核技巧和流形学习算法融合,分别针对人脸图像向量和人脸图像矩阵,提出了非局保投影算法和融合核技巧的二维局保投影算法,从识别率和训练以及识别时间上对基于人脸图像向量和人脸图像矩阵的两类特征提取算法进行了比较。实验结果表明提出的方法具备了提取人脸图像的非线性特征和邻域保持特性的两大优.势,因而取得更好的识别效果。3.对基于图像矩阵的子空间方法进行了较为深入的研究,提出了分块双向加权的二维主成分分析,分块多投影双向二维线性判别分析、分块二维判别监督局部保留主成分分析三种特征提取方案。实验结果表明提出的算法都具有比改进前更好的特征提取能力。说明对算法的改进是有效的,有利于模式识别。4.从信息融合的角度出发,提出将二维人脸识别和三维鼻型识别在决策层融合以解决单训练样本人脸识别识别率低,可信度低的新思路,分别针对1：1验证和1：N辨识两种工作模式提出了基于改进的加权投票法的“并行结构”和基于双层筛选模型的“串行结构”两种融合策略。实验证明,基于2D人脸与3D鼻型信息融合的人脸识别框架是解决单训练样本问题的有效途径之一。较基于单一模态的身份识别方法有很多优势。首先,2D信息在获取上非常便捷,可以利用现有很多相关的成熟技术进行处理；同时,3D信息包含了物体的三维形状属性,与2D信息相互补充,可以较好地解决人脸识别中饰物和表情变化的问题。更多还原

【Abstract】 The face recognition with a single training sample means that each person only save one face image as training set to recognize the identity of face image whose attitude and light can be change. The problem of single training sample has important significance and brings large challenge, so it has become an important research direction of face recognition. This paper analysed the research status of face recognition with a single training sample, and indicated that sample augment and multi-feature hybrid are efficient way to solve the problem of single training sample. This paper mainly focuses on theory and algorithm to research the face recognition with a single training sample of 2D and bimodal hybrid, especially research the 2D face image feature extraction and 2D face and 3D nose shape bimodal information hybrid.The main research work is as follows:1. This paper proposed a synthesis method of 3D face model based on single frontal face image through studying the sample augment of face recognition with a single training sample. This method use synthetical 3D face model to creat virtual face images whose attitude, light and expression can change to expand training sample set, so the problem of single training sample is translated into face recognition with multi-training sample. The 3D model recovering by 2D image is a basic problem in computer vision filed. Conventional method is impractical because it needs many face images and image sequences, or 3D image pair, frontal and side image pair in limited condition. The method proposed in this paper is suitable for practical application and has a broad application prospects because it only needs one frontal face image which reduces the demand for use condition. The synthetical 3D face model method can satisfy the demand for face recognition, expression animation and human-computer exchange. The experiment based on virtual image proves that the sample expanding method can solve the problem of illumination, gestureand and expression in face recognition.with a single training sample efficiently. Meanwhile, it extracts 3D nose shape from the generated 3D face model and using it to recognize people, the experiment proves the feasibility of 3D nose shape as a novel biometric identification mode.2. From the view of increasing single mode, which is called 2D face recognition performance, this paper improves the conventional 2D face feature extraction method, combines kernel trick and manifold learning algorithm, proposes Non-locality Preserving P(?) ction(NLPP) algorithm and 2DLPP algorithm aiming at face image vector and matrix and (?) (?)pares two feature extraction algorithms based on face image vector and matrix respective (?) from the recognition rate, training and recognition time. The experiment proves that the methods proposed can not only extract the non-linear information from the image but also hold the neighbour-keeping characteristic. Thus, better recognition results can be gotten.3. This paper has deeply researched the subspace method based on image matrix, proposed three feature extraction methods which are modular bi-direction weighting 2D major component analysis, modular multi-projection bi-direction 2D linearity discriminant analysis and modular 2D discriminant locality preservation major component analysis. The experimental results indicated that the proposed algorithms have better feature extraction performance than previous algorithms. It means that it is benefit to pattern recognition by changing the conventional algorithm.4. From the view of information hybrid, this paper proposed a new way that 2D face recognition and 3D nose shape recognition should be fused in decision level to solve the problem that the face recognition with a single training sample has low recognition rate and credibility. Two hybrid methods namely the parallel structure based on improved weighting voting method and serial structure based on double-layer screening model have proposed focusing on 1:1 Authentication and 1:N identification model. The experiment shows that face recognition framework based on 2D face and 3D nose shape information hybrid is a suitable method to solve single training sample problem and it has many advantages comparing with single identity recognition method. Firstly,2D information can be easy to catch, and can be treated by many related mature techniques; simultaneity,3D information including 3D shape properties of the object which can be supplement with 2D information and preferably solve the problem of accouterment and expression in face recognition.更多还原