【作者】 蒲晓蓉；

【导师】 章毅；

【作者基本信息】 电子科技大学 ， 计算机应用技术， 2007， 博士

【摘要】 多模态生物特征识别,作为一种新型的、凭借用户自身固有的属性特征进行身份识别,近年来日趋成为国际上的研究热点。其应用涉及商业、安全和司法等领域,可用于自动视频监控、访问控制、身份鉴定、计算机人/机界面设计、银行ATM机等。作为生物特征识别的主要方法之一,人工神经网络自80年代初复苏以来一直是科学与工程上的一个热点研究学科,吸引了包括MIT、Harvard大学等国际一流学术机构的科学家从事此领域研究,并取得了大量的研究成果,《Science》、《Nature》等许多国际一流学术刊物常有相关研究成果发表。本论文重点研究多模态生物特征融合的神经网络方法,主要研究了如何利用神经网络方法提取人脸的局部特征和全局特征,并进行身份识别。其中主要研究了PCA主元分析、带稀疏度约束的非负矩阵分解NMFs、径向基神经网络RBF、Fisher线性判别法等方法,并用于含噪声的人脸图像识别(本文不涉及人脸检测)。研究了基于固定吸引点的联想神经网络,构建了能学习连续吸引子的子空间联想存储器,用于局部残缺、局部遮挡的人脸图像识别。研究了连续粒子群优化算法PSO模型及二进制离散PSO模型,并利用人工免疫阴性选择机制控制粒子的最低及最高飞行速度。研究了自适应模糊神经推理、自适应Bayes决策融合策略、以及自适应Bayes优化的人工免疫二进制PSO模型(用于寻找最优融合规则),并利用自适应模糊神经推理和自适应Bayes优化的人工免疫二进制PSO算法解决人脸全局特征与局部特征融合、人脸与指纹融合等多模态生物特征融合的身份识别问题。本论文的主要研究成果和创新如下:1.提出并改进了基于神经网络的人脸图像的局部特征提取与识别方法,将带稀疏度约束的非负矩阵分解NMFs与Fisher线性判别法结合起来,利用径向基神经网络RBF作为分类器,对识别不同光照、采自不同头部姿势、不同表情的、含局部遮挡噪声的人脸图像,具有显著的优势,系统性能明显提高;2.建立了具有连续吸引子的子空间联想神经网络模型,与具有固定吸引点的联想神经网络相比,子空间联想神经网络能恢复填充输入人脸图像的残缺或遮挡部分,并能显著提高部分残缺/遮挡的人脸图像的识别性能;3.提出了最低/最高速度约束的连续粒子群优化算法,并用于训练多层感知机,能有效保证在搜索函数的最优解时,算法的收敛速度更快,收敛能力更强,并且能避开局部最优,收敛到全局最优;4.提出了人脸全局特征与局部特征融合的自适应模糊神经推理,以及自适应Bayes优化的人工免疫二进制PSO算法,融合人脸的局部和全局特征进行身份识别,显著提高了人脸识别性能。同时,融合人脸与指纹特征进行身份识别,也获得了较好的性能。更多还原

【Abstract】 As a novel personal identification technology using certain physiological or behavioral traits associated with the person, Biometrics has always been an attractive topic of research. Biometric systems make use of fingerprints, hand geometry, iris, retina, face, hand vein, facial thermograms, signature or voiceprint to verify a person’s identity. They have an edge over traditional security methods in that they cannot be easily stolen or shared. Examples of such applications include secure access to buildings, computer systems, laptops, cellular phones and ATMs.Neural networks approach has been regarded as an interesting and powerful tool for biometrics identification. Since 1980s, more and more experts from some famous universities such as MIT and Harvard universities have been doing research on the theories and the applications of neural networks. Many important achievements have been reported on some first class international journals such as Science and Nature.This thesis studies the problem of multi-modal biometric fusion by using neural networks. The main contributions of this thesis are as follows:1. Using neural networks to study holistic and partial based face recognition. It combines Neural Networks, Principal Component Analysis (PCA), Non-negative Matrix Factorization with Sparseness constraints (NMFs), Radial Basis Function (RBF), Fisher’s Linear Discriminant (LDA or FLD), to investigate face recognition for face images with large variations in lighting, pose, facial expression and partial occlusion noise or partially damaged facial images. A novel partial facial features extracting method by combining NMFs with FLD (FNMFs) is proposed, and the RBF classifier is then applied to classify the facial images with large variations. A comparative analysis engages PCA-FLD (FPCA) method and FNMFs method for both parts-based and holistic-based face recognition. The comparative experiments show that FNMFs has better performance than FPCA-based method for face recognition.2. Using subspace associative memory with continuous attractors to study face recognition. The traditional associative memories with fixed-point attractors and associative memories with continuous attractors are studied. A subspace associative memory with continuous attractors is proposed. It is applied to recognize partially damaged or occluded facial images. The theoretical expressions are plotted, and the comparative experiments are carried out. It shows that partial-feature-based subspace associative memory outperforms holistic-feature-based subspace method significantly in recognizing partially damaged faces, and the subspace associative memory can learn and store some continuous attractors for completion partially damaged face images.3. Some algorithms are proposed on continuous and discrete binary particle swarm optimization (PSO). In these algorithms, both maximum and minimum velocities are controlled to improve the abilities of the convergence by applying the theory of negative selection in Artificial Immune System (AIS). Two multilayer perceptron networks are successfully trained by the PSO with minimum and maximum velocity constraints in order to overcome premature convergence and alleviate the influence of dimensionality increasing.4. A novel binary PSO algorithm based on Adaptive Neuro-Fuzzy Inference System and Artificial Immune System for face recognition is proposed to select the fusion rules by minimizing the Bayesian error cost. Such fusion rules are applied to face recognition as well as fusion face and fingerprint. Experimental results show that the proposed fusion algorithm outperforms individual algorithms that based on PCA or NMFs.更多还原