【作者】 赵磊；

【导师】 董吉文；

【作者基本信息】 济南大学 ， 计算机应用技术， 2013， 硕士

【摘要】 伴随尖端科技日新月异，人脸检测和识别已然成为科研领域的热门研究对象，人脸检测作为人脸识别的第一步，在人机交互、视频监控、视频处理、安全安防等领域中具有广阔的应用前景。在实际应用中问题较多，人脸检测需要考虑光照、遮挡、倾斜、背景[4]等的影响，从复杂变化的环境中检测出的人脸框并不能很好的代表真实人脸位置。由于该领域的特殊性，通常需要对图像序列进行分析，对算法的实时性、准确性要求较高。在人脸识别环节中，又涉及到神经网络和模式识别，语音识别、虚拟现实等领域，目前的生物识别技术包罗万象，从开始的指纹识别，虹膜识别，到现在的人脸识别甚至是DNA生物信息对比识别等，可以说人脸检测跟踪及人脸识别在当今科研的众多领域都有涉及并且占据相当重要的位置。本文首先对目前的人脸检测算法进行对比和探讨后，对Paul Viola等人提出的AdaBoost算法的基础上进行研究，考虑到AdaBoost算法对于后续人眼检测工作实时性的影响，本文引入了CamShift算法。基于颜色模型的跟踪算法CamShift算法具有很好的实时性、较小的计算量等优点，可以较好的跟踪前后两帧人脸框旋转，倾斜角度等信息的跟踪，为后续人眼检测提供了先验知识。在人脸检测的后续阶段，又引入了归一化的NMI特征方法，在人脸框中进行人眼定位，准确的人眼定位增加人脸检测的精度。并对人眼定位进行改进，在NMI空间中提出上下限筛选人眼特征等方法，提高了人眼定位的识别率，进而提高人脸检测的精度。本文在人脸识别方面从基于kernel方法识别算法入手，并对比了核主成分分析和传统线性判别方法效果，此外，在之前的特征提取空间内利用零空间的有效信息提取有用的鉴别信息。在人脸特征分类识别阶段，实验对比了欧氏距离和余弦角距离对人脸分类识别的效果，最后对预处理后的人脸采用基于零空间的核主成分分析，线性鉴别分析以及进行欧氏距离和余弦角距离的对比。大量实验表明，基于零空间的人脸识别能把以往方法被排除的有效信息重新利用，增强了人脸特征的有效性，并且作为最近邻分类器的余弦角方法的识别率要好于欧氏距离方法，充分发挥了余弦角距离的优势，得到了较满意的识别率。更多还原

【Abstract】 With the most advanced technology ever-changing,face detection and recognition hasbecome a popular research objects in the field of scientific research, face detection as the firststep in face recognition of human-computer interaction, video surveillance, video processing,safety and security and other fields, has broad application prospects. There are many problemsin practical applications, face detection needs to consider lighting, occlusion, tilt, thebackground, face frame detected from the complex changes in the environment and can’t be agood representative of the true position of the face. Due to the special nature of the field, itneeds generally image sequence analysis, real-time algorithm, higher accuracy requirements.Face Recognition related to neural networks and pattern recognition, speech recognition,virtual reality and other fields, biometric technology all-inclusive, even from the beginning ofthe fingerprint recognition, iris recognition, to the face recognition DNA bio informationcontrast identification, face detection and tracking and face recognition are involved in manyareas of research today and occupy a very important position.Firstly, we discuss the face detection algorithm to compare and explore research, takinginto account the AdaBoost algorithm proposed by Paul Viola et al AdaBoost algorithm basedon real-time detection work for the follow-up of the human eye, this paper introduces theCamShift algorithm. CamShift tracking algorithm based on the color model has theadvantages of good real-time, the smaller amount of computation can better track around thetwo face frame rotation, tilt angle information tracking for follow-up eye detection whichprovides priori knowledge. Face detection in the follow-up phase, we introduce the NMIFeature normalization method used for human eye positioning in the face frame, accuratehuman eye positioning lead to increase the accuracy of the face detection. After that, weimprove the positioning of the human eye by introducing upper and lower filter characteristicsof the human eye in the NMI space to raise the recognition rate of the human eye positioning,thereby increasing the accuracy of the face detection.Start with recognition algorithm based on kernel methods in face recognition andcompare the effects of the kernel principal component analysis and traditional lineardiscriminant method, effective information extraction using zero space to identify useful information in the space of the feature extraction before. In facial feature recognition andclassification stage, experience compares the effect of the classification and identification ofEuclidean distance and the cosine angular distance, at last,apply zero space-based kernelprincipal component analysis, linear discriminant analysis to the image after face pretreatmentas well as Euclidean distance and cosine angle distance comparison. A large number ofexperiments show that zero space-based face recognition re-use the useful information theconventional method excluded where it can enhance the effectiveness of the facial features,and the recognition rate of cosine angle distance methods is better than Euclidean distance,and it gives full play to the advantages of the cosine angle distance, obtain a satisfactoryrecognition rate.更多还原