【作者】 高云飞；

【导师】 李红信；

【作者基本信息】 兰州大学 ， 电子与通信工程（专业学位）， 2014， 硕士

【摘要】 经典的自动人脸识别算法在理想情况下的人脸图像识别具有很好的适应性,其识别率也比较理想。可是在面对复杂非理想情况下人脸图像的识别问题时,其识别率下降明显,不能满足一定的要求。SIFT算法是基于物体的局部属性特征来描述物体的,对于物体的形状变化,平移变化,旋转变化具有很好的鲁棒性。应用在人脸识别邻域对非理想情况下的人脸识别具有很好地处理效果。在基于SIFT的经典人脸识别算法中,需要遍历数据库中的所有人脸图像寻找特征点匹配个数最多的人脸图像,完成人脸识别,这样构成了闭集的人脸识别系统。算法计算过程比较繁琐,并且算法中的一些阈值设定也比较具有先验性,能否满足一些复杂情况下的人脸识别有待验证。本文提出了通过对数据库中的人脸图像提取SIFT特征点,把属于同一类别的人脸图像上提取到的所有的SIFT特征点聚集在一个行向量中,此行向量作为这一类别的人脸对象的人脸模型特征集合。由于人脸的轮廓具有某些类似性,比如头发、额头与背景部分具有相似性,灰度级变化比较平滑,根据这些部分来判断人脸对象身份可信度会比较小,所以在这些部分上提取的特征点对人脸分类的贡献度比较小,需要把它们去除掉。每一类的人脸模型特征集合去除掉分类性比较小的特征点后,把数据库中的所有图像作为训练样本训练出个体分类器。由于数据库中的训练样本集有限,所以很难把各种非理想情况下的人脸图像都包括在训练样本中。设计个整体分类器,搜集大量非数据库中人脸对象的各种非理想情况下的图像,寻找各种人脸变化情况下的共性,训练出整体的分类器,是对个体分类器的一种补充同时构成判断输入的人脸图像是否属于数据库的开集人脸识别系统。输入的人脸图像投影到数据库每个人脸对象的人脸模型特征集合上生成有相关性的相似度向量,然后通过个体分类器与整体分类器加性融合后输出最终结果,完成人脸识别。最后通过实验对比,证明基于SIFT算法的人脸识别方法比经典的人脸识别方法在非理想情况下的人脸识别能力拥有更好的性能。更多还原

【Abstract】 The classical auto face recognition algorithms have good adaptability for human face image in the ideal condition, and its recognition rate is also good. But when dealing with the problem of human face recognition in the complex and non-ideal condition, the recognition rate is low and can not meet certain requirements. SIFT algorithm is based on the local characteristics of the object attributes to describe the object, and it has good robustness for the change of the object shape, translation and rotation. In the classical face recognition algorithm based on SIFT, we need to traverse the database to find all of the face image feature point matching the most number of face images to complete face recognition, which constitutes a closed set of face recognition systems. The computation process of the algorithm is complex, and some thresholds set in the algorithm has a prior, which can satisfy the face recognition in the complex condition is verified.This paper extracts SIFT feature points from the human face image through the database, gathering all SIFT feature points from the same class people in a row vector, which is the set of human face model feature from the same class people. Since the outline of the face having a certain similarity, such as the hair, the forehead has a similarity with the background portion, a relatively smooth gradation level changes, according to the face portion to determine the identity of the object will be relatively small reliability, so the contribution of the extracted feature points from those portions to face classification is relatively small, they need to be removed. After getting rid of the small classification feature points from the each class set of human face model, all the rest images will be trained samples to train the individual classification. Due to the limited sample set of the training database, it is difficult for all the face images are included in the training sample under non-ideal conditions. Designing an overall classification and gathering a large number of non-database object images of human faces under various non-ideal situation, finding the common features of face changing from all kinds people and then, train the overall classifier is an supplement for the individual classifier, meanwhile, constitute an open set face recognition system for determining whether the input human face image belonging to the database. The inputting human face image projected to the every human model feature set of the database to generate the relevant similarity vectors, and then after the adding of the individual classifier and overall classifier, outputs a final result to complete the face recognition. Finally, through the experiment comparison, the face recognition based on the SIFT is proved has better performance for the face recognition than the classical face recognition in the nod-ideal condition.更多还原