【作者】 程勇；

【导师】 赵春霞；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2010， 博士

【摘要】 人脸识别作为一种非接触式、友好的生物特征识别技术,在军事、公安、经济等安全领域具有广阔的应用前景。目前人脸识别已成为模式识别、图像处理、计算机视觉、认知科学和神经网络等领域的一个研究热点。近几十年来,国内外研究者提出了各种人脸识别方法,并已经生产出许多自动人脸识别系统,但是,FERET和FRVT测试结果表明光照变化会严重影响自动人脸识别系统的识别性能。针对光照问题,本文以减弱或消除光照变化严重影响人脸识别的问题为目标,以提取光照不变量(光照不敏感特征)为研究主线,从视觉模型、多尺度几何分析、朗伯光照模型和光照不敏感特征的角度研究复杂光照情况下人脸识别问题,提出了相应的光照不变量提取算法。下面概述本文的主要研究内容和进展。最近,人类视觉系统处理图像信息的模式受到图像处理、图像理解和模式识别等研究领域的广泛关注。2007年Meylan等通过Naka-Rushton方程对人类视网膜的非线性信息处理进行建模,提出了一种基于人类视网膜模型的图像局部对比度增强方法。2009年Vu等将该对比度增强方法应用在可变光照的人脸识别中,提出了一种基于视网膜模型的光照人脸识别方法。本文对人类视网膜模型和Meylan等的图像对比度增强方法进行研究,发现Meylan等的图像对比度增强方法在估计局部光照时,仅考虑图像像素点的位置相似度未考虑像素点的强度相似度,在图像边缘或纹理区域不能准确估计局部光照,因此,在此基础上增强的图像会在图像边缘或纹理区域产生扭曲现象。针对该问题,本文将双边滤波(Bilateral Filtering)引入到人类视网膜模型中,提出了一种基于双边滤波和人类视网膜模型的光照不变量提取算法,取得了较好的效果。光照人脸识别中基于光照模型提取光照不变量的方法主要采用朗伯光照模型。朗伯光照模型是一种经典的经验模型。陔模型假设目标物体表面具有朗伯反射特性,即当光线照射到物体时,物体表面向各个方向都有相同的散射,漫散射分量仅与物体表面和光源的入射角有关,与观察者(成像设备)的位置无关。一幅灰度图像F如果符合朗伯反射特性则可以简单的描述为：F=R×Ⅰ,其中,只是图像的内在特征,它取决于物体的反射率和表面法向量；Ⅰ是成像过程中的光照情况。经典的光照不变量提取方法需要假设光照分量Ⅰ变化缓慢,而物体的反射率和表面法向量尺变化较大。因此,Ⅰ对应图像的低频成份；R对应图像的高频成份,提取图像光照不变量归结为如何从图像中将R分离出来。根据朗伯光照模型研究者提出了MSR (Multi-scale retinex, MSR)、SQI (Self Quotient Image. SQI)和MFSR (Multiscale facial structure representation, MFSR)等光照不变量提取算法。MSR和SQI通过权重的高斯滤波器获取平滑图像,无法保持良好的图像边缘信息,无法准确估计光照分量,因此,这类方法不能准确获取光照不变量。MFSR首先对对数域图像通过小波变换消噪模型获耳义图像的平滑成份,然后用对数域图像减去平滑成份获取图像的光照不变量,取得了不错的效果。但是,小波变换是一种各向同性的多尺度分析技术,只能描述点状奇异性,对轮廓与纹理这样的线状奇异性的表征则无能为力,因此,MFSR方法会产生较强的伪Gibbs现象,无法获取图像准确的光照不变量。针对上述方法的不足,本文对多尺度几何分析技术进行研究,将非下采样轮廓波变换和自适应NormalShrink降噪技术应用于图像光照不变量的提取过程中,提出了一种基于NSCT和自适应NormalShrink滤波的光照不变量提取算法及一种改进的SQI算法。实验结果表明本文提出的算法不仅改善了人脸图像光照不变量的视觉效果,而且提高了复杂光照情况下人脸识别的识别精度。图像轮廓作为一种主要的高频信息受光照变化影响较小,包含着图像的大部分信息,是图像的一种重要的内在特征。轮廓特征在图像处理及模式识别领域受到广泛关注,已经被应用于立体匹配、图像拼接、图像检索和图像识别等方面。另外,研究者发现图像的方向信息比图像的强度信息包括了更多的识别信息；人类的视觉皮层的感受野具有方向性,有效的图像表征方法应该是多方向和多尺度的。因此,本文对图像的轮廓信息进行研究,首先,从多尺度、多方向、轮廓特征的角度入手,提出了一种图像光照不变量——多尺度主轮廓方向(Multiscale Principal Contour Direction, MPCD)。然后,以非下采样轮廓波变换为基础,对图像进行轮廓分析,构造多尺度多方向轮廓信息(复信息)。最后,根据多尺度主轮廓方向的定义提取出图像的多尺度主轮廓方向。实验结果表明本文提出的多尺度主轮廓方向是图像的一种光照不敏感特征。近年来,基于图像梯度分析的特征被应用于图像分割、图像识别及动念目标跟踪等方面。研究者们指出梯度方向是图像的一种重要的光照不敏感梯度特征,并被应用于复杂光照情况下人脸识别中。最近,梯度脸(Gradientfaces)通过高斯函数一阶导数与图像作卷积求取图像的梯度域,在图像的梯度域求取图像的梯度方向,在复杂光照人脸识别中取得了较好的结果。本文在梯度方向和梯度脸的启发下,提出了一种图像的光照不敏感梯度特征——梯度最大分量方向(Gradient Maximum Component Direction, GMCD),实验结果表明GMCD算法优于Gradientfaces算法。更多还原

【Abstract】 Face recognition, a non-contact and friendly biometric identification technology, has broad application prospects in the military, public security and economic security. Face recognition has become one of research focuses in the fields of pattern recognition, image processing, computer vision, cognitive science and neural networks. In recent decades, a variety of face recognition methods have been proposed, and many automatic face recognition systems have been produced by domestic and foreign research institutions. However, test results of FERET and FRVT show that varying illumination would seriously affect performance of face recognition. In order to weaken or eliminate the problem, this paper focuses on how to extract illumination invariants (i.e., illumination insensitive features) from vision model, multiscale geometric analysis, the Lambert illumination model and illumination insensitive features, and proposes some illumination invariant methods. Its main works and research results are as follows:Recently, the model processing image information via human vision system is concerned by image processing, image understanding and pattern recognition. In 2007, Meylan et al. modeled human retina model by Naka-Rushton function, and propose a method for image local contrast enhancement. In 2009. Vu et al. applied image local contrast enhancement into face recognition with varying illumination, and developed an algorithm based on retina modeling. In this paper, we find that Meylan et al. only consider neighborhood pixels’ geometric closeness without their photometric similarity when estimating the local illumination. This produces inaccurate local illumination estimation in some edges and texture regions of an image, which would cause distorted result in the subsequent process of local illumination compression. To solve the problem, this paper introduces bilateral filtering into human retina model, and proposes an algorithm for extracting illumination invariant based on bilateral filtering and human retina model. The experimental results are satisfying.Methods for extracting illumination invariant based on illumination model in illumination face recognition mainly use the simple Lambert illumination model. Lambert illumination model is a kind of classic empirical model. The model assumes that object has a Lambertian surface reflection characteristics, that is. when the light shines to the object, its surface in all directions has the same scattering, and diffuse scattering component only concerned with the surface and the incident angle of light source regardless of the observer position. A gray image F meeting with Lambertian surface reflection characteristics can be described as:F= R×I. where R is inherent characteristics of the image, depending on the object’ reflectivity and surface normal, and I is related to the lighting source. Generally, to extract R by the simple Lambert illumination model, there have hypothesis that I changes slowly and R varies abruptly. On the basis of the illumination model, researchers propose many methods, for example. MSR (Multi-scale retinex), SQI (Self Quotient Image) and MFSR (Multiscale facial structure representation), etc. MSR and SQI getting the smoothed image by using weighted Gaussian filter make it difficult to keep sharp edges in low frequency illumination fields and can not estimate accurately illumination component. Hence, those methods can not extract illumination invariant accurately. MFSR first performs logarithm transform on original face images, then obtains the smoothed image from the Log-domain image by wavelet denoising model, and finally extracts illumination invariant by the difference between the Log-domain image and the smoothed one. It achieves good experimental results. However, the wavelet transform as an isotropic multiscale analysis can only describe point-like singularity, and is powerless to express linear singularity liking contour and texture. Therefore, MFSR will face strong pseudo-Gibbs phenomenon, and can not obtain accurate illumination invariant. To address the deficiencies, this paper studies multiscale geometric analysis, and proposes an illumination invariant algorithm based on nonsubsampled contourlet transform and adaptive noise reduction technology of Normal Shrink and an improved SQI algorithm. Experimental results show that the proposed algorithms not only improve visual effects of illumination invariant features, but also improve accuracy of face recognition.Image profile, major high-frequency information, is little affected by light changes. It includes most of the information in an image, and is an important intrinsic characteristic of the image. Contour feature of an image has received extensive attention in image processing and patter recognition, and has been used for stereo matching, image stitching, image retrieval and image recognition, etc. Human visual cortex receptive fields are characterized as being localized and directional, and an effective image representation method should be based on multi-direction and multiscale. Therefore, this article first studies contour information of an image and proposes an illumination invariant feature (Multiscale Principal Contour Direction, MPCD) of an image, then performs contour analysis based on nonsubsampled contourlet transform and constructs multiscale multi-orientation contour information (complex information), and finally gets MPCD of an image by its definition. Experimental results show that MPCD is an illumination insensitive feature.In recent years, characteristics based on image gradient analysis are applied to image segmentation, image recognition and dynamic target tracking. Researchers point out that image gradient direction is an important gradient feature being insensitive to varying lighting, and has been used in face recognition under complex illumination. Recently, Gradientfaces first gets image gradient field via performing convolution between an image and the first derivative of Gaussian function, then obtains image gradient direction in the above field. Gradientfaces achieves better results in face recognition under complex illumination conditions. Inspired by image gradient direction and Gradientfaces, this paper proposes an illumination insensitive feature (Gradient Maximum Component Direction, GMCD). Experimental results show that GMCD is superior to Gradientfaces.更多还原