【作者】 刘晓山；

【导师】 金连文；

【作者基本信息】 华南理工大学 ， 通信与信息系统， 2011， 博士

【摘要】 自动人脸识别(AFR)是生物识别领域的研究热点,它涉及到图像处理、模式识别、机器学习和计算机视觉等多个学科的研究内容。人脸识别技术因其具有良好的非接触、非打扰的优势在身份鉴定、门禁系统、监控、法律实施和人机交互等方面有着广泛的应用前景。经过多年的发展,研究者已经提出了很多性能卓越的人脸识别算法,也出现了一些成功的商用系统。但是当人脸图像出现光照、姿态、表情、年龄的变化、遮挡问题或少样本问题的时候,识别率会急剧下降,特别是光照的变化对识别率的影响最大。设计出对光照变化具有高识别率和高鲁棒性的人脸识别算法是人脸识别领域中的一个难点和热点。本论文针对人脸识别技术中的光照变化问题进行了研究,重点研究在光照变化情况下的人脸特征提取方法。论文的主要研究工作包括以下几个方面:1.提出了基于Wavelet多尺度LBP/LTP人脸特征提取算法。局部二值模式(LBP)算子具有良好的灰度不变性、旋转不变性和尺度不变性,能够适应人脸图像光照的变化。局部三值模式(LTP)算子在LBP基础上发展而来,在保持了LBP优良的性质之外,还具有抗噪性能,能够更加精细的表达图像的纹理特征。本文利用Wavelet多分辨分析的思想构建了一个人脸图像的塔式结构,在不改变LBP/LTP算子的前提下可以使提取出的人脸特征不仅包含了局部信息,还包含了人脸图像的全局信息,提高了人脸特征的分辨能力。2.对双树复小波变换(DTCWT)的各子带系数在人脸识别中的重要性进行了研究,同时还研究了DTCWT高频系数的分布情况,并根据分布情况提出了基于系数分割方法的DTCWT边缘增强人脸识别算法。双树复小波变换(DTCWT)具有多分辨和多方向能力,通过提取6个方向的高频系数作为人脸的特征可以去除部分的光照影响。而且DTCWT可以采用传统的离散小波来实现,结构简单,运算速度快,同Gabor变换相比具有明显的优势。本文提出的DTCWT边缘增强算法能够很好地增强人脸图像的边缘信息,提高了人脸的分类能力。3.在辐照度模型的基础上提出了基于DTCWT的光照不变特征提取方法。辐照度模型是朗伯光照模型的泛化。它没有将光照的影响进行精细的划分,因而分析简单,常常被研究者用于二维人脸识别技术中来处理人脸图像的光照问题。辐照度模型存在一个通用的假设:人脸图像中受光照影响的部分属于图像的低频;而表现图像的纹理特征的部分属于图像的高频部分。本文就是基于这一假设,使用阈值收缩去噪的思想来获得光照不变的纹理特征。实验结果表明该方法提取出的光照不变特征能够滤除大部分的光照影响。4.提出了基于非下采样Contourlet变换(NSCT)的人脸光照不变特征提取方法。NSCT具有多分辨、多方向和各向异性的优点,具有更强的图像纹理表达能力,提取出的光照不变特征带有更多的信息量,图像的熵更高。本文还结合全变差(TV)模型提出了一种NSCT特征和TV特征融合的人脸识别算法,将算法的识别能力进一步提升。最后从朗伯光照模型出发,从理论上解释了基于DTCWT方法和NSCT方法只能处理部分光照的影响的原因。更多还原

【Abstract】 Automatic Face Recognition (AFR) is an active topic in Biometric Technology, which involves a surprising number of scientific disciplines, such as image processing, pattern recognition, machine learning and computer vision. Due to its advantage of being natural and non-intrusive, face recognition technology have a wide range of applications in identification, access control systems, surveillance, law enforcement and human-computer interaction. Lots of face recognition algorithms have being proposed. And some successful commercial systems also emerged. Current systems perform well under relatively controlled environments but tend to suffer when variations in different factors (such as pose, illumination, expression etc.) are present. Especially large illumination variations would seriously affect face recognition algorithms. It is still a great challenge to build high performance and robust algorithms for face recognition. At present, recognition under illumination variations is an active topic and difficult task in the field of face recognition technology.The face recognition algorithms under various lighting conditions are studied. And the focus of our work is extracting face features under illumination variations. The main work and innovation of this dissertation include the following aspects:1. An algorithm of facial features extraction based on multi-scale LBP/LTP is proposed. Local binary pattern (LBP) operator has good properties of gray-scale invariance, rotation invariance and scale invariance. And it is robust to illumination variations in face images. Local ternary patterns (LTP) which develops from the LBP operator maintains the good properties of LBP as well as noise immunity. A tower structure of face images is produced using the idea of Wavelet multiresolution analyse. The facial features extracted by this way contain the local and global information of face image, while with no need for changing the LBP/LTP operator’s parameters. This structure can enhance the identification ability of facial features.2. The contribution of each sub-band coefficients of dual-tree complex wavelet transform (DTCWT) to the face recognition performance is studied. And an edge enhancement face recognition algorithm based on DTCWT’s coefficient distribution is proposed. Due to DTCWT’s multi-resolution and multi-directional properties, the facial features extracted from the high-frequency coefficients in 6 directions can partly remove the effect of illumination. Furthermore DTCWT can be implemented by traditional discrete wavelet transformation, which has simple structure and fast speed. The proposed DTCWT edge enhancement algorithm have better recognition rate and higher computation speed than Gabor-based face recognition technology.3. A novel method of exacting illumination invariant features based on DTCWT is proposed using irradiance lighting model. The irradiance lighting model is the generalization of Lambertian illumination model, which is often used in 2D face recognition technology to deal with face illumination because of its simpleness. There is a common assumption in irradiance model: the illumination in face image is regarded as the low frequency of images, while the texture features of the image belong to the high frequency part. Based on this assumption, the illumination invariant texture feature can be extracted from the face image with the idea of threshold shrinkage denoising. Experiment results show that illumination invariant features extracted by this method can filter out most of the illumination effects.4. A method of facial feature exaction based on nonsubsampled Contourlet transform (NSCT) is proposed. NSCT has better performance to express the image texture than wavelet with superiority of multi-resolution, multi-direction and anisotropy. The extracted illumination invariant features have more information and higher entropy. A new recognition algorithm fusing the NSCT’s features with TV’s features is proposed, which has better performance. Finally, the reason why the illumination invarian features extracted by DTCWT and NSCT can only deal with part of the illumination is explained using Lambertian illumination model.更多还原