硕士研究生课程论文（或读书报告）课程名称： 模式识别 题 目： 人脸识别技术研究 摘要人脸识别是计算机视觉和模式识别的一个研究热点。但是在复杂光照条件下，如何快速自动识别人脸，仍然是一个富有挑战性的问题。基于图像处理的知识，研究在复杂光照下利用计算机自动识别人脸的技术。在系统设计中详细阐述整个人脸识别系统的处理流程，比较系统地介绍了该系统的图像预处理、人脸检测、人脸特征定位、人脸特征提取、人脸识别等组成部分。通过对彩色图像的偏色进行分析，提出一种结合偏色纠正和改进Retinex的彩色图像增强算法。利用灰度世界和完美反射理论建立偏色纠正的数学模型，通过线性拟合对偏色图像进行偏色纠正。对彩色图像进行亮度和色度分离，多尺度Retinex算法对亮度分量增强并进行自适应调整，通过获取的亮度增益矩阵对彩色图像的ROB三分量进行逐点增强。该算法解决了彩色图像增强后色彩变化的问题，对于存在偏色、低亮度等复杂光照下的彩色图像均有较好的增强效果。人脸图像具有稳定的肤色特征和灰度分布，运用结合肤色检测的AdaBoost算法检测人脸。利用肤色检测算法获得肤色区域信息，去除大量非人脸的背景部分，通过对肤色块的统计分析，得到可能人脸的尺寸范围。将人脸尺寸范围及肤色区域二值图像提供给AdaBoost人脸检测算法，从而减少搜索区域及搜索尺度范围。该人脸检测方法克服了人脸类肤色和检测速度慢的问题，能够快速有效地检测人脸。人脸特征定位容易受到光照的影响。针对灰度图像，提出一种新的基于各向异性滤波的人眼定位方法。构造各向异性滤波器对图像进行滤波，消除光照影响；运用形态学操作突出眼睛的特征区域，并采用相关系数法对特征区域块进行匹配，获得眼睛粗定位；对粗定位区域进行重定位校正获得精确的眼睛中心点。对于彩色图像，提取三分量差分特征，二值化并滤波后，通过特征区域的相关匹配定位眼睛中心。根据彩色图像的眼睛中心点位置初步确定嘴的区域，提取红色度信息和RGB差分特征信息，定位嘴巴。基于各向异性滤波的人眼定位方法解决了复杂光照环境下的人脸定位问题。有效地提取人脸特征是人脸识别成功的关键。运用多尺度局部二进制模式提取人脸纹理特征。对图像进行小波分析，并运用局部二进制模式方法在不同尺度的分块图像上提取人脸特征。多尺度局部二进制模式能够全面、准确地表达人脸图像的纹理特征，解决特征描述的准确性问题。人脸识别的核心在于寻找最优的分类特征。提出一种改进的正交拉普拉斯特征脸识别算法。该算法在保局投影的目标函数中融入类间离散度，运用Schur分解实现基向量的正交化。这种改进的算法利用类别信息提高分类性能。对人脸识别系统进行了测试，在光照变化的Yale B人脸库上，人脸的识别率达到9475，实验证明，人脸识别系统能够达到复杂光照下人脸识别的要求。关键词：图像增强；人脸检测；人脸特征定位；人脸特征提取；人脸识别ABSTRACTFace recognition technology is a hot topic for researchers on computer vision and pattem recognitionHowever,how to quickly and automaticly recognize face remains a challenging problem in complex illuminationAutomatic face recognition by computer technology is researched based on image processingIt elaborated the processing flow of the face recognition system in system designComponents of the system，such as image preprocessing，face detection,facialfeature location,facial feature extraction，face recognition，were systematically introduced An color image enhancement algorithm combined color offset corrected and the improved Retinex was proposed on the analysis of color offset for color image.The mathematical model for t11e color offset corrected was estabished by gray world and the perfect reflection theory，and the color offset of original color image was corrected by linear fittingThe brightness was separated from color imageBrightness component Wasenhanced by multi-scale Retinex enhancement algorithm and adapted itself。and the RGB three components of color image were enhanced point by point by brightness gain in matrix.The algorithm proposed solved the problem that the color Was changed after the color image enhancedThe algorithm proposed had a better performance for color image enhancement with the images under variable illumination such aS color offset，low-intensitylight and SO onThere are stable skin color feature and gray distribution for face imagesAdaBoost algorithm combined with skin color detection was used to detect faceSkin color detection algorithm was used to obtain skin color information of skin regions and to remove a large number of non-face background，and the size range of faces Was got by analysis of skinblocksSize range of the human face and binary image of skin color region were provided to AdaBoost algorithm for face detection，thereby the search area and search for scales reducedThe method for face detection solved the problem that non-face taken as face and low testingIt Can rapidly and effectively detect human facesFacial feature location is effected easily by illuminationAn new method of eye location based on anisotropic filtering Was proposed for gray imageAnisotropic filters were constructed,and the image input was filtered to eliminate the influence of variant illumination；the features of eye areas were highlighted by morphological operation；the method of correlation coefficient was used to match the feature area blocks to obtain eye rough location；to obtain the accurate center of the eyeme coarse positioning region was corrected and relocatedFor color images，differential features for RGB three component were extracted，and then eye centre points were located by the related matching of feature area after binarization and filterAccording to the center of eyes in color image，the initial location of the mouth was determinedAnd the mouth was located by combined with the information of red degree and RGB differential feature Extracting facial features effectively is the key to recognize face successfullyThe method of extracting face texture feature by multi-scale local binary pattern was usedIt used wavelet to transform image，and facial features were extracted by local binary pattern on image blocks of different scalesMulti-Scale local binary pattern can ben fully and accurately expresses the texture features of face images，solving the problem that features are accurately describedThe core technology of face recognition is to find the optimal classification featuresAn improved algorithm of orthogonal Laplace Eigenface was proposedThe algorithm added class scatter fused into objectiv