【作者】 冯悦；

【导师】 鲁宏伟；

【作者基本信息】 华中科技大学 ， 计算机系统结构， 2007， 硕士

【摘要】 光照、姿态和表情是人脸识别中亟待解决的核心问题。二维人脸识别算法在特征提取方面已达到相当成熟的程度,在图片中的人脸为正常光照和中性表情的正面人脸时能够取得较好的识别效果。但由于二维人脸识别算法没有考虑到人脸的三维形状信息,使得多数算法对于光照、姿态和表情具有很差的鲁棒性。同时由于三维模型给出了人脸真实的形状信息,不同于二维图像是人脸在某个方向上向平面的投影,信息更充分详细,从而有可能产生对于光照、姿态和表情具有更强鲁棒性的识别算法。在三维人脸模型重建的基础上,设计了一个基于二维Gabor特征的三维人脸识别算法。该算法基于能够精确并且稳定描述人脸特征的二维Gabor特征,依次采取三维人脸模型重建,对重建后的三维人脸模型进行模板匹配以及对匹配后的三维人脸模型进行线性判别分析的步骤完成三维人脸识别。基于ORL(Olivetti Research Laboratory)和UMIST(University of Manchester Institute of Science and Technology)两个人脸数据库的实验证明该算法具有良好的性能。同时定义了一个光照空间,通过设计人脸表面模型以及采用光照补偿和姿态识别最优化步骤有效地解决了光照和姿态对人脸识别造成的影响。通过引入一个三维网格模型建立整个面部的几何信息以及依次采取模型匹配、生成模板变形图、小波分析和识别认证阶段的数据处理完成人脸表情识别,有效地解决了人脸表情对人脸识别造成的影响。更多还原

【Abstract】 Illumination,gesture and expression are three core problems of face recognition that wait to be solved urgently these days.2-D face recognition algorithms can be quite accurate in extracting face features and obtaining preferable recognition results when the faces are those frontal faces under normal illuminations and with neutral expressions.However,in respect that 2-D face recognition algorithms do not take 3-D shape information into account,many algorithms have too bad robustness towards illumination,gesture and expression.In contrast,as 3-D model gives more detailed shape information of face and is quite different from the 2-D image which could only give the projection of the face into a surface in one direction,3-D model could engender those algorithms that are quite robust towards illumination,gesture and expression.Followed by 3-D model reconstruction,a 3-D face recognition algorithm based on 2-D Gabor feature extraction is brought forward.This algorithm is based on accurate 2-D Gabor features which could stably describe features on human face.The 3-D face recognition is performed by three steps:3-D face model reconstruction,pattern matching of reconstructed 3-D face model and Fisher’s Linear Discriminant Analysis of matched 3-D face model.The experiments developed on ORL(Olivetti Research Laboratory) and UMIST(University of Manchester Institute of Science and Technology) databases show that the proposed method has excellent performances.An illumination space is defined and applied to effectively reduce influences of illumination and gesture by designing a face appearance model and applying illumination compensation as well as optimization of gesture recognition.A 3-D grid model is introduced to construct the geometric information of the whole face.The whole process of expression recognition is put forward by model matching, creating the deformable model graph,wavelet analysis and data processing in the phase of recognition as well as verification.更多还原