【作者】 张壮；

【导师】 尹宝才；

【作者基本信息】 北京工业大学 ， 计算机应用技术， 2006， 硕士

【摘要】 经过三十几年的发展,人脸识别技术取得了巨大的进步,在理想的成像条件下,已经取得了显著的成果。但非理想成像条件下(如光照条件不理想,人脸的姿态发生变化等)的人脸识别技术还远未成熟,有许多科学问题需要研究。本文基于三维形变模型,针对多姿态人脸识别问题,主要做了以下几方面的工作:提出了一种基于三维部件技术的多姿态人脸识别方法,该方法根据面部部件对姿态变化的不敏感性,利用部件的三维信息作为特征进行人脸识别。识别过程中结合人脸的局部特征和全局特征,统计各个部件的识别率,根据单个部件的识别率确定其在整体分类中的权值,基于整脸信息进行识别,与基于二维部件技术的识别方法相比,进一步改善了识别效果。提出了基于部件的三维形变模型。在模型匹配过程中,定义相同视点下获得的三维部件图像与给定部件图像的误差作为目标函数,由于每个部件维数比较低,与整个面部的三维形变模型相比,目标函数的规模相对较小,优化效果得到改善,提取的三维部件特征更准确。采用上面提出的基于三维部件技术的多姿态人脸识别方法,结合两层分类方法,将部件重建结果应用与人脸识别,取得了更好的识别效果。更多还原

【Abstract】 After thirty years of research, face recognition techniques have been made enormous progress. In well-controlled environments, the face recognition were successful in terms of their recognition performance. However, the pose and illumination variation is still very challenging for face recognition. Our work concentrates on the 3D Morphable Model and recognizing faces from different poses. The main study includes the following contents.In this thesis, we present a method to pose variant face recognition that combines two recent advances: Component-based face recognition and 3D Mophable Model. In this method, 3D components is robust to pose changes and are extracted as the input feature to face recognition. For classification, we combine the local feature and global feature of face and the whole face are used as input to the final classifier ,where each component is verified its weight based on its recognition rate. Comparing with Component-based face recognition in 2D, this method is robust to pose variant.A 3D Morphable Model based on components is present in the thesis. In the process of matching a Morphable Model to image, the Euclidean distance between the reconstructed 3D component image and the input component image is defined as the goal function. Comparing with the whole face 3D Morphable Model, the dimension of each component is low and the scale of the goal function of each component is relatively small, so the optimized result is improved and the 3D component is more precise. The reconstruction result, combined with Two-Layer classification, is better to face recognition.更多还原