【作者】 安高云；

【导师】 阮秋琦；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2009， 博士

【摘要】 进入二十一世纪,人脸识别迎来了一个至关重要的攻坚阶段。经过二十世纪近40年的发展,人脸识别领域已经积累了丰富的理论和大量成功算法,也已经基本解决了可控条件下的人脸识别难题,但非理想条件下、用户不配合、大规模人脸数据库上的人脸识别则是当前人脸识别攻坚阶段所要完成的任务。到目前为止虽然相继有一砦成果被公布,但应当说人脸识别所面临的难题均未被解决,我们目前仍处在此次攻坚阶段的起步时期,仍处在对各难题的探索时期。本文对复杂条件下的人脸识别技术进行了研究,重点对其中如何解决光照影响、如何消除人脸图像中的遮挡以及如何提取强普适性的人脸特征三个关键问题进行了全面的综述和系统的研究,并提出了若干新模型和算法。主要贡献如下:1.提出光照不变面像合成模型。所提模型以偏微分方程中的全变分模型为基础,通过兼顾大小尺度特征可以实现在对原始人脸图像进行光照预处理的同时尽量保证预处理后的图像信息不受损失。实验部分对所提模型的预处理结果进行了定性分析,同时采用图像熵作为度量标准对其进行了定量分析。与常用的直方图均衡化算法以及新近提出的商图像算法相比,所提模型无论从视觉感官方面还是从图像熵的保持方面均优于现有算法。为进一步验证所提模型的有效性,本文对其预处理后的结果采用目前已有的典型子空间分析算法进行特征提取,实验结果表明:经所提模型预处理后的人脸图像可以提升整个人脸识别算法的识别率。2.提出抗遮挡干扰的有监督最佳表示而像合成模型。所提模型包含两个重要模块:最佳面像合成系数的获取模块和自学习有监督遮挡掩模的生成模块。从处理效果看,所提面像合成模型可以有效去除样本中所存在的遮挡,从而生成一个无遮挡人脸图像。不过合成后的人脸图像中存在较大的引入噪声,为进一步消除引入噪声,本文进而提出一种自适应特征保持面像复原算法。从定性分析的结果看,最终复原出的人脸图像在感官上非常自然;从定量分析的结果看,最终复原出的同一人的样本间以欧氏距离为标准的相似性度量明显缩小,即达到归一化的目的,可为后续算法提供无遮挡干扰的样本输入。3.对典型子空间分析系列算法进行系统的普适性评测。在这一部分,本文首次提出普适子空间分析算法这一概念,并对目前的子空间分析系列算法的普适性(相关概念参看本文第四章第3节)进行了系统研究。实验部分设计了狭义开集测试和狭义闭集测试(相关概念参看第四章第3节)两种测试方法,选用目前常用的四个评测人脸库对代表性算法进行综合评测。经过大量的综述、分析及实验验证,得出如下几点对今后子空间分析算法的研究具有参考意义的结论:1).有监督算法的泛化能力弱于无监督算法的,主要表现在只有当不出现训练集之外个体需要识别且训练样本具有代表性时,有监督算法才能取得较好的识别性能。2).在所有无监督算法中,独立分量分析相关算法的普适性最优,表现在其整体识别性能无论在狭义闭集测试还是狭义开集测试中均最优。3).独立分量分析相关算法为典型的普适子空间分析算法,其中的ICA算法可作为普适子空间分析算法的基准算法。4.提出非线性固有码提取模型与普适人脸识别算法。为解决现有算法未能同时克服光照、表情以及遮挡等外部干扰对算法性能的影响以及识别率偏低的问题,这一部分的主要工作将在前三章工作的基础上,提出一种系统的普适人脸识别算法。在新提出的普适人脸识别算法中,除了借鉴前三章的工作之外,还提出了一种最佳表示特征分析(Expressive Feature Analysis,EFA)模型和非线性固有码提取(Non-linear Intrinsic Codes Extraction,NICE)模型。EFA模型主要是用来替代传统的PCA模型并实现对高维原始人脸图像进行降维,EFA模型可以保证整个算法在对高维面像样本进行降维的同时突出由变量的幅度值所体现出来的有效信息,保证后续操作能够提取到可分性特征;而NICE模型则用来对输入样本进行编码,NICE模型最终生成的人脸图像的固有码既有较好的普适性又有较好的可分性,可以作为人脸图像的固有码而应用于识别任务。在目前常用的CAS-PEAL、FERET、Yale B以及CMU PIE人脸库上的狭义开集测试和狭义闭集测试的实验结果进一步证实所提算法具有较好的普适性,直接表现为所提算法在各测试中的识别率均高于采用Gabor特征的普适子空间分析算法(ICA)的识别率。更多还原

【Abstract】 As the coming of the 21st century,face recognition is facing an important and difficult period.During the passed 40 years,a lot of theories and successful algorithms have been studied and proposed,and most key issues of face recognition under controlled conditions have been solved.While face recognition under nonideal condition, uncooperative condition,or large scale face databases,is still an unsolved problem.Till now,many algorithms were proposed for face recognition under complex conditions, but the research on this topic is just the beginning and all the new algorithms are at the exploratory stage.In this paper,we give deep analysis on face recognition under complex conditions. We specially focus on solving the effect from lighting and occlusion,and extracting the robust & adaptive face features.We proposed several new models and algorithms.The novel models and algorithms proposed in this paper are illustrated as follows:1.An illumination invariant face reconstructing model is proposed.Based on the Total Variation theory,this model extracts illumination invariant face features from both large and small scale information;also,it preserves all the key features.In the experimental parts,we give qualitative analysis about this reconstructing model and using image entropy as the quantitative analysis.Compared with conventional histogram equalization algorithm and the newly proposed quotient image algorithm,our model outperforms them based on both the visual sense and the image entropy.To further prove the validity of our model,we use some famous subspace analysis algorithms to extract the features of face images reconstructed by the proposed model.The experimental results prove the reconstructing model could largely improve the face recognition rates.2.A supervised occlusion invariant face reconstructing model is proposed.The proposed model is composed by two important parts:occlusion invariant reconstructing coefficients extracting and adaptive supervised mask of occlusion generating. From the visual evaluation point,the proposed model could effectively remove the occlusion in the original face samples,and then generate an ideal occlusion invariant face image.Since the model would induce noise,we further propose an adaptive face reconstructing model which could preserve the facial features while removing the noisy information.Based on the qualitative analysis,the recon- structed face images have good visual perceptions and they are close to the true face samples collected under normal conditions.Based on the quantitative analysis, the Euclidean distance between the reconstructed face samples of the same person is obviously reduced.Therefore,the lace samples are the normalized result and they could be used as the occlusion invariant samples for the subsequent recognition algorithms.3.Testing and Evaluation of the robust & adaptive property of current standard subspace analysis algorithms.The conception of robust & adaptive subspace analvsis algorithm is firstly proposed.We give deep analysis about the robust & adaptive property(definition referred to section 3 in chapter 4) of current standard subspace analysis algorithms.In the experiments,we propose two new testing methods for face recognition:the narrow-sense open set testing and the narrow-sense close set testing(definition referred to section 3 in chapter 4).The most popular four face databases are adopted to evaluate the recognition performance.Based on the summary,analysis and experimental results in this paper,we conclude the following conceptions which would provide reference significance for further research on subspace analysis algorithms.1).The unsupervised algorithms have better robust & adaptive property compared with the supervised algorithms.Only when no individual not belong to the training set needs to be recognized or the training set are representative,the supervised algorithms outperform the unsupervised ones.2).Among all the unsupervised algorithms,the independent component analysis algorithm has the best robust & adaptive property.It achieves the best performance in both the narrow-sense close set testing and the narrow-sense open set testing.3).The face recognition algorithms based on the independent component analysis are the typical robust & adaptive subspace analysis algorithms,and the ICA algorithm could be used as the reference algorithm.4.Non-linear intrinsic codes extraction model and robust & adaptive face recognition algorithm.Up to now,no algorithm could resolve all the problems of illumination effect,expression,occlusion,and low recognition rate etc.at the same time. Based on the research work in the first three sections,we propose a novel systematic robust & adaptive face recognition algorithm which could solve most existing problems.In the algorithm,we propose an Expressive Feature Analysis(EFA) model and a Non-linear Intrinsic Codes Extraction(NICE) model.The EFA model is mainly used to replace the conventional PCA model to reduce the dimensions of the original face samples.It could enhance the effective information represented by the scale of variables in statistics while reducing the dimensions.Therefore,it guarantees that the most discriminate features would be extracted in the following steps.The NICE model is used as the coding part of the input face samples.The intrinsic codes extracted by it have good robust & adaptive property;also,they have good discriminant property.These intrinsic codes used for face recognition could achieve good performance.We execute both narrow-sense open set testing and close set testing experiments on the famous CAS-PEAL,FERET,Yale-B and CMU PIE face databases.The experimental results further prove the robust & adaptive property of our algorithm.The recognition rates of our algorithm are higher than that of the ICA algorithm employing Gabor analysis.更多还原