【作者】 马媛媛；

【导师】 马小虎；

【作者基本信息】 苏州大学 ， 计算机应用技术， 2013， 硕士

【摘要】 人脸识别是图像处理和模式识别领域的一个重要研究课题，人脸识别和认证技术在公共安全、智能监控、多媒体等领域有着广阔的应用前景。经过数十年的研究，在理想情况下人脸识别技术已能取得较好的识别性能。但在不可控环境中易受到光照、姿态、表情、遮挡等因素的影响，使识别性能急剧下降。让人脸识别系统走向实用仍然是一个极具挑战性的课题。本文针对光照问题进行研究，以提高人脸识别系统对光照的鲁棒性和识别率为主要目标，主要对人脸图像预处理、提取光照不变特征、分类识别等关键阶段展开研究，探讨人脸识别问题的研究方法。本文的主要工作和研究成果概括如下：1.首先介绍了人脸识别课题的研究背景和意义及国内外发展情况，针对光照问题的三种研究方法：光照建模、光照补偿、光照不变特征提取法进行分类总结，分析其优缺点。2.提出一种基于非下采样Contourlet变换（Nonsubsampled Contourlet Transform，NSCT）和邻域去噪的光照不变人脸识别算法。光照问题是影响识别效率的重要因素之一。本文在分析了小波变换提取人脸特征的基础上，采用多尺度的NSCT，它不仅具有小波变换的多分辨率和时频局域特性，还具有很强的方向性和冗余性，可以更完备的提取光照不变特征，同时在图像表示上能更好的描述人脸细节信息。采用邻域去噪方法去除光照不变特征中的投射阴影等噪声，因为投射阴影出现在局部的可能性最大，邻域去噪符合这点，其只在小范围内进行去噪，相比于一般的去噪方法，能保留更多的边界信息。经实验证明，该方法有效的提取光照不变特征，显著提高了人脸识别率。3.人脸识别系统分为预处理、特征提取、分类识别三个关键环节，每一阶段对人脸识别系统性能都有所提升。本文提出的基于局部二值模式（Local Binary Pattern，LBP）和线性回归的可变光照人脸识别算法同时涉及这三个阶段，在每一阶段中对光照进行处理，得到更完备的可变光照算法。在预处理阶段采用Gamma校正、DoG滤波、对比度均衡化方法，降低光照敏感度，采用具有光照鲁棒性的分块LBP提取光照不变特征，最后使用改进的线性回归模型进行分类识别，即消除受光照影响最大的主成分系数。所提方法能有效的消除光照对人脸识别的影响，提高人脸识别系统的鲁棒性和识别率。更多还原

【Abstract】 Face recognition is an important research topic in image processing and computervision, which has promising applications in public security, smart surveillance, multimediaand so on. Through face recognition has achieved great progress in the decades, when itcomes to uncontrolled conditions, such as different illumination conditions, pose variation,mixture of emotions and object shelter, and the accuracy of face recognition willdramatically decline. Therefore, how to build a real-life face recognition system is a sharpchallenging topic.This paper focuses on the problem of face recognition in variant illuminationenvironment. The main purpose is to improve the recognition accuracy and robustness offace recognition system under various lighting condition. To achieve this purpose, thefocus of our work is on the image preprocessing, illumination invariant feature extraction,classification and recognition. The major work and contributions of this paper are follows:1. The research background and significance of face recognition both at home andabroad are introduced firstly in this paper. Then the methods dealing with lighting problemare summarized, that is illumination model, preprocessing and normalization, invariantfeature extraction, and the advantages and disadvantages of these methods are analyzed.2. An illumination invariant algorithm based on Nonsubsampled contourlet transform(NSCT) and NeighShrink denoise is proposed. Illumination is one of the factors that affectthe recognition efficiency. On the analysis of wavelet transform, we extract illuminationvariant feature through NSCT, which is a fully shift-invariant, multi-scale, multi-directiontransform and not only can extract more effective illumination invariant facial features butalso can get a clearer positive light image of face. A NeighShrink-based denoising model isapplied, which considers the correlations of sub-band coefficients. Thus, more usefulinformation can be restored in the high frequency sub-band coefficients, unlike some of the other approached in which too many coefficients that might contain useful informationtend to be killed. Experimental results showed that our method could extract invariantfeature more effective and obviously improve the recognition accuracy.3. Face recognition system is divided into three key points, image preprocessing,feature extraction, classification and identification. Each part helps to improve the systemperformance. A variable light face recognition algorithm involved three points based onlocal binary pattern (LBP) and linear regression model is proposed. It can obtain betterperformance, through processing the illumination at each stage. In the image preprocessingphase, an efficient preprocessing chain is adopted which contains Gamma correction,difference of Gaussian(DoG) filtering, contrast equalization and can eliminates most of theeffects of changing illumination. Then block LBP is applied to extract invariant featurewhich is robust to variant illumination. Finally, improved linear regression model is used toclassification which drops the first n principal components. The proposed approach canreduce the effect of illumination, and improve robustness and the recognition accuracy offace recognition system.更多还原