【作者】 范春年；

【导师】 张福炎；

【作者基本信息】 南京大学 ， 计算机应用技术， 2011， 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。近年来,人脸识别技术取得了很大进展,涌现出许多优秀的人脸识别方法,许多人脸识别系统表现优异。但是,人脸识别仍是一个没有彻底解决的难题,光照变化是其中关键问题之一。2006年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。针对光照变化问题,本文从光照归一化和提取光照不变量两个方面入手,将视觉模型、同态滤波、多尺度分析等理论引入光照变化问题的研究中,力图能够消除光照变化对人脸识别的影响。本文的贡献主要体现在以下五个方面：(1)提出了一种基于同态滤波的光照归一化算法依据朗伯光照模型,将同态滤波方法引入人脸识别光照处理中,同时将传统的同态滤波方法中使用的高通滤波器换成了具有带通特性的高斯差分滤波器,并在同态滤波后增加了一个直方图均衡化处理步骤。这样的光照归一化算法简单高效,能有效消除光照变化的不利影响。实验结果验证了算法的有效性。(2)提出了改进的基于小波的光照归一化算法小波变换是多尺度分析工具之一,能够从受光照变化严重影响的图像中提取关键的面部特征。现有的基于小波变换的光照处理算法处理高频分量时存在不足,针对这个问题,本文改进了该算法。首先在小波变换前增加了一个对数变换,此举将乘性朗伯光照模型转化为对数域加性模型,同时提高图像的对比度；接着在高频处理环节中增加了小波阈值去噪处理,以去除高频分量中混杂的噪声。实验结果表明改进后的算法识别率有了显著的提升。(3)提出了基于同态滤波和LoG算子的光照不变量提取算法研究表明图像的边缘对光照变化不敏感。现有基于图像边缘特征的算法均是直接对人脸图像进行边缘检测。当光照变化较大尤其是图像中含有阴影时,提取的边缘中会出现对识别不利的虚假轮廓。针对这个问题,本文将光照归一化预处理环节引入边缘图EM算法中,在归一化光照后再检测边缘,收到了较好的效果。(4)提出了基于Gabor相位特征的光照不变量提取算法Gabor小波变换对外界环境如光照、表情、姿态及遮挡等具有较强的不变性,能够提取鲁棒的人脸特征表示。现有的基于Gabor特征的方法大都使用Gabor幅值特征而丢弃了相位特征。但研究表明,相位信息包含了许多有效的图像特征,对光照变化不敏感。受此启发,本文提取Gabor小波相位特征作为光照不变量,同时,为了避免复杂的复数计算,且因为人脸图像本身具有对称性,选择了2-D对称实Gabor小波函数。算法取得了出众的光照处理效果,且计算复杂度也大为降低。(5)提出了改进的基于非下采样Contourlet变换的光照不变量提取算法Contourlet变换是一种新的多尺度几何分析方法,具有多方向选择性和各向异性,在表示图像的边缘和纹理等几何特征方面有着独到的优势。本文改进了NSCT和LNSCT算法中对低频分量的处理方法,将低频分量进行直方图均衡化处理,此举既保留了对识别有用的信息,又有效消除了光照变化带来的影响,取得了较好的去光照效果。实验结果验证了改进后算法的有效性。相对其他方法,本文提出的算法均具有以下优点：①不需要3D形状和光源等先验信息；②训练时不需要大量的训练样本,可直接应用于单训练样本的人脸识别场合；③算法的运行速度比较快,可应用于实际的人脸识别系统。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 received significant attention in pattern recognition, image processing, computer vision, cognitive science and neural networks. Related research in recent years has made great progress, a number of excellent face recognition algorithms have emerged, and a number of face recognition systems have achieved good performance. However, many issues still remain to be addressed and illumination changes remain one of the major challenges for current face recognition systems. The report of FRVT 2006 shows that varying illumination will seriously affect the performance of face recognition. To deal with the illumination variation problem, this dissertation focuses on illumination normalization and extracting illumination invariants, and introduces vision model, homomorphic filtering, and multiscale analysis theory into research on illumination problem, trying to eliminate the effects of illumination changes on face recognition. The main contributions are as follows.(1) A homomorphic filtering based illumination normalization method is presented.According to Lambert illumination model, the homomorphic filtering method is introduced to dealing with illumination variation problem for face recognition, in which the Difference of Gaussian filter with band-pass characteristics replaces the high-pass filter used commonly. Moreover, a histogram equalization process is added. This illumination normalization method is simple and efficient, and can effectively eliminate the adverse effects of illumination variation. Experimental results verify the effectiveness of the method.(2)An improved wavelet-based illumination normalization method is proposed.Wavelet transform is one of the multiscale analysis tools, and is able to extract key facial features from face images severely affected by illumination conditions. There are deficiencies in existing wavelet-based illumination processing algorithms when dealing with high frequency component. To address this issue, an improved algorithm is presented. First, a logarithmic transformation is added before the wavelet transform, and which will not only take the multiplicative Lambert illumination model into additive model in log-domain, but also improve image contrast. Second, a wavelet threshold de-noising is used in high frequency processing part in order to remove mixed noise. Experimental results demonstrate the significant performance improvement of the proposed method.(3) A homomorphic filtering and LoG operator based illumination invariants extracting method is presented.Studies have shown that the image edge is not sensitive to illumination variation. In addition, existing illumination processing algorithms based on edge features are always applied edge detection on face images directly. The extracted edge using these methods will have aliasing artifacts disadvantageous to face recognition when the light is changed greatly, especially when images contains shadows. To deal with this issue, the illumination normalization preprocessing is introduced to the edge map algorithm. Therefore, the edge detection is used after illumination normalization. This obtains good results.(4) An illumination invariant method based on Gabor phase feature is proposed.Gabor wavelet transform is insensitive to external environment such as illumination, facial expressions, gestures, and occlusion, etc., and can extract robust facial feature representation. Most existing methods based on Gabor features always use the Gabor magnitude features and discard the phase features. However, studies have shown that the phase information contains a number of effective image features, and is insensitive to illumination variation. Inspired by this, the Gabor phase features are extracted as illumination invariant in this dissertation. The 2-D symmetric real Gabor wavelet is chosen in our method in order to not only avoiding the complexity of complex calculations, but also fitting the symmetry of the face image itself. A superior light treatment effect is achieved, and the computational complexity is greatly reduced.(5)An improved illumination invariant extracting algorithm based on nonsubsampled Contourlet transform is presented.Contourlet transform is a new multiscale geometric analysis method with multi-directional selectivity and anisotropy, and has a unique advantage in representing the geometrical features such as edges and texture of images. This dissertation improves the NSCT and LNSCT algorithm in the way of processing the low frequency component. The low frequency component is applied histogram equalization, which not only retains the useful information for identification, but also effectively eliminates the impact of illumination variation. This results in good illumination eliminating effects. Experimental results verify the effectiveness of improved method.Compared with other methods, the proposed methods have the following advantages.①There is no need to any prior information on 3D face shape and light sources assumption;②there is no need to many training samples, thus our methods can be directly applied to single training image per person condition; and③they are simple and computationally fast, and can be applied to the real face recognition system.更多还原