【作者】 李全彬；

【导师】 刘锦高；

【作者基本信息】 华东师范大学 ， 通信与信息系统， 2011， 博士

【摘要】 自动人脸识别的研究具有重要的理论价值和广阔的应用前景。自本世纪以来,相关技术已经取得了长足的进步,在约束环境下已经取得了满意的识别效果,一些商用系统也已经开始在某些领域得到一定的应用。但实践表明,非约束环境下自动人脸识别系统的广泛应用,还面临很多需要解决的技术难题,本文对其中涉及到的部分关键问题进行了相关研究。论文的主要研究成果总结如下：一、全面综述了人脸检测和人脸识别技术的研究历史和现状本文将现有的人脸检测方法分为基于知识、基于统计和基于肽色模型三类进行了综述,全面介绍了该方向的最新研究成果,并将人脸识别分成三个阶段进行了综述,对各个阶段代表性的算法进行了分析,对人脸识别国内外的研究现状和研究机构(团队)进行了全面的介绍。同时,对人脸识别相关的重要资源进行归纳整理,对重要的公共人脸库和人脸识别领域重要的国际会议和期刊进行了分类,并对推动人脸识别技术发展产生重要影响的相关测试进行了概括性介绍,最后结合测试的结论分析了当前自动人脸识别技术在应用上所面临的关键技术难题。二、研究了非约束环境下的人脸检测问题1.提出了一种光照鲁棒的肤色模型构建方法该方法提取复杂光照下肤色样本的YCbCr特征值进行训练,得到光照鲁棒的肤色模型。实验结果表明,该模型在检测各种复杂光照的彩色人脸肤色区域时均表现出良好的性能,配合4-连通区域筛选和肤色区域还原技术,能够实现准确的肤色区域检测和定位。2.提出了基于SMQT+SNoW+SVM的复杂光照人脸检测方法为了解决SMQT+SNoW人脸检测方法在检测复杂背景和复杂光照的人脸时存在的误检率高、检测时间过长的问题,本文引入了肤色预检和支持向量机分类策略,提出了基于SMQT+SNoW+SVM的人脸检测方法,该方法首先利用肤色模型对彩色图像进行人脸候选区域的分割,然后利用SMQT方法计算相应区块的特征,最后利用SNoW+SVM的方法实现了快速准确的人脸检测。针对1000张复杂光照图像人脸检测的实验结果表明,该方法在速度和准确率上都取得了良好的表现,误检率也下降到了可以接受的水平,满足了系统的实时运行需求。3.提出了基于FloatBoost的复杂光照多姿态人脸检测方法该方法首先利用光照鲁棒的肤色模型进行肤色分割,进而搜索可能的人脸区域,然后在人脸特征定位的基础上,确定候选人脸的特征块,并将这些候选区域利用FloatBoost进行分类,最终实现了快速准确的多姿态人脸检测。与其他已有算法的对比实验表明,所提方法不仅明显提高了复杂光照下多姿态人脸的正确检测率,缩短了检测时间,而且将可检测人脸姿态的范围扩大到[-90,+90]。同时,提出的特征搜索策略明显改善了最终检测出的人脸区域的分割效果,为后期人脸识别提供了更准确的人脸特征信息。三、研究了非约束环境下的人脸识别问题1.提出了基于统一模式LBP(ULBP)和SVM的复杂光照人脸识别方法在对人脸特征提取方法进行综述分析的基础上,将最近提出的LBP特征算法应用到人脸的纹理特征提取中,采用两级LBP算子级联的方法扫描经3*5非规则分块的人脸图像,并将扫描结果的直方图按顺序组合起来作为最终的鉴别特征,然后利用训练的SVM分类器实现了复杂光照的人脸识别。在YaleB库的实验结果表明,这种分块、分级的人脸特征提取方法,兼顾了人脸图像的细节和整体特征,可有效消除光照的影响,增强所提取特征的可鉴别性,能够有效提高复杂光照下的人脸识别率。2.提出了一种准确的人脸特征定位和姿态估计方法该方法首先利用色度信息产生人脸的映射图,然后提取二值化后人脸图像的4-连通区域信息,并利用设定的规则消除误检的特征块,最终实现了绕Y轴旋转角度在[-90,+90]之间的多姿态人脸关键特征点的准确定位,并结合定位结果给出了人脸各种姿态旋转角度的计算方法。3.提出了一种基于水平镜像和决策融合的多姿态人脸识别方法该方法利用水平镜像技术产生更多的训练样本,并将人脸姿态从[-90,+90]划分为7个姿态空间,然后利用Gabor小波提取各个姿态空问下样本的Gabor特征,并采用2DPCA降维,形成对应的7个特征子空间。识别时,抽取输入图像及其水平镜像图像的特征向已训练的7个特征空间投影,然后根据投影的欧式距离,采用决策融合的方法得到最终的识别结果。在ORL、ColorFeret和Cas-Peal人脸数据库上的实验结果表明,该方法在少量训练样本的情况下,即可对姿态跨度[-90,+90]的多姿态人脸取得满意的识别结果。四、提出了一个网络人脸识别系统的实现方案从数据库设计、服务程序开发、网页设计等方面详细阐述了网络人脸识别系统的实现过程,给出了具体的设计方案,并在实际应用中取得了很好的效果。本文的上述研究内容为非约束环境下的人脸检测和人脸识别提供了相关的解决方案,并已在网络人脸识别系统中得到应用。更多还原

【Abstract】 The study on Automatic Face Recognition(AFR) has both significant theoretic values and bright future of applications. Since this century’s beginning, AFR technology has made great progress and obtains satisfactory results under limited conditions, some AFR commercial systems have successfully applied in some fields. However, practice has proved that AFR still has many existing technical problems to be solved under unlimited conditions. This paper does some research on several key technologies of AFR about the above problems.The main contributions of this thesis are as follows:1. Provided a thorough survey of face detection and recognition on history and research situationThis paper provides a detailed survey of research in the area of face detection on three aspects:knowledge-based approach、statistical-learning-based approach and skin-color-model-based approach, and reviews the latest findings of face detection generally. Then, the research of face recognition is reviewed from three historical stages, classical algorithms in every stage are analysised. Moreover, generally introduce the famous research institutes(groups) of AFR both abroad and in China and summarizes the important resources related to AFR, such as face databases, the top international conferences, authoritative journals and famous tests on AFR. Finally, introduce the key technical difficulties in the applications of AFR depending on those tests results.2. Studied face detection under unlimited conditions(1) Proposed a new method for skin color model under illumination variationsThe proposed method trains a robust skin color model using the YCbCr values of selected samples. The experimental results show that this model achieves a good performance on face skin region detection under illumination variations, with the support of 4-connected regions detection and the face skin region recovery technology, proposed skin color model can detect and locate the face regions accurately.(2) Proposed a face detection method based on SMQT+SNoW+SVMThe SMQT+SNoW method has some disadvantages such as low speed and high false positive rate when applied to face detection. In order to solve those problems, this paper presents a modified method to detect faces using the strategy of pre-detection of skin region and SVM classification. First, search the potential face regions using the proposed skin color model, then, calculate the regions’ feature values by SMQT, finally, detect the real faces accurately using the classification of SNow+SVM. The experiment on 1000 face images under illumination variations shows that proposed method performances very well on speed and correct rate, at the same time, the false positive rate also has a marked reduction and can meet demand of practical applications.(3) Proposed a multi-view face detection method under illumination variations based on FloatBoostThis method firstly search the potential face regions using the proposed skin color model, then, chroma map is adopted to obtain the four-connected components from the skin color segmentation blocks, label them, and identify the center of each block, finally, the faces verification is performed through the classifier based on FloatBoost. Comparing with some other previous algorithms for multi-view face detection, our method not only effectively improves the right detection rate of multi-view faces under illumination variations, but also obviously decreases the time consumption and operation complexity, and at the same time, the located faces position are more accurate and be good for improving the accuracy of feature extrication in next face recognition step.3. Studied face recognition under unlimited conditions(1) Proposed a new method for face recognition under illumination variations based on ULBP and SVMFirst, the proposed method applies two-degree ULBP to extract Illumination invariant feature on multi-block face, second, combines those histogram features into the final identification characteristics in the right order, then, SVM is applying to feature classification to realize the face recognition under illumination variations. The experiment on YaleB database shows that, the combination of multi-degree and multi-block feature extraction method based on ULBP performances well both details and universe on extracting the illumination invariant face feature and achieves a high face recognition rate.(2) Proposed an efficient method to locate key face feature points and estimate the head pose accuratelyThis method first calculates a face map using chroma information, binary it and search the four-connected regions on binary result. Second, eliminates the regions which do not follow proposed rules, then the key face feature points are located accurately. Finally, the head pose is estimated by using the above feature points. (3) Proposed a Multi-view face recognition method based on horizontal mirror technology and decision-level image fusionThe proposed method first generates more face training samples by horizontal mirror technology, then, estimates the head pose of each of them and classifies it to one of seven corresponding pose spaces decomposing from the pose range of-90 to +90. Second, extracts the face feature by Gabor wavelet and reduces the feature dimension using 2DPCA to make seven feature sub-spaces. when recognition begin, firstly, makes input face’s horizontal mirror image, secondly, extracts their face feature and estimate their head pose using the same method, and projects them to corresponding feature sub-space, finally, calculates the projection Euclidean distance and achieves the recognition result by decision-level image fusion. The experiments on ORL、ColorFeret and Cas-Peal database show that it can obtain satisfactory result on pose range of-90 to+90 with few training samples.4. Design and realize an online face recognition systemThis paper describes in detail how to design an online face recognition system from the construction of information database、the development of client/server background server programs and the design of website, gives the process flow diagram and experiment result. The proposed system has been applied to smart access and performance well.This thesis provides some new methods to solve the problems of face detection and recognition under unlimited conditions, the proposed methods have been applied to an online face recognition system.更多还原