基于综合集成的人脸识别

【作者】 赵桂敏；

【导师】 夏利民；

【作者基本信息】 中南大学 ， 交通信息工程及控制， 2004， 硕士

【摘要】 人脸识别在模式识别领域的发展和应用方面都有着重要意义，目前是一个非常活跃的研究方向。人脸识别问题一般可以描述为：从背景图像中检测人脸；抽取人脸识别特征；最后进行匹配和识别。本文在人脸识别的三个主要环节上均进行了研究工作，主要工作体现在以下几个方面： (1)提出了基于子空间特征的人脸检测方法。论文中利用K-L变换得到特征脸子空间，进而得到样本在该空间上的投影系数。在此基础上，采用模糊聚类和加权K近邻分类器相结合的方法进行人脸检测，使得人脸检测速度明显提高。 (2)在基于子空间方法的人脸识别中，为了克服NMF方法不能提取图像局部特征信息的缺点，提出用局部NMF方法提取人脸子空间特征，该方法能够提取图像的局部特征信息，有利于提高人脸识别率。将Bagging思想用于神经网络，进一步提高神经网络的分类准确率和泛化能力。 (3)在基于几何特征的人脸识别中，为了克服眼睛定位不准确的缺点，提出一种根据几何特征定位人眼，然后采用特征眼模板进行二级验证的定位方法。 (4)考虑到子空间特征和几何特征反映的是不同模式的特征信息，提出综合集成的人脸识别方法。在综合集成过程中，为克服分类器权系数确定的不灵活性，提出基于神经网络和遗传算法动态确定权系数估计器的方法。 实验证明，基于以上方法的人脸识别，比单独采用子空间特征或几何特征进行识别，识别率明显提高。更多还原

【Abstract】 Face recognition plays an important role in the development and application of pattern recognition. At present, it is an active research topic. A general statement of the problem can be formulated as three steps: face detection, facial feature extraction, match and recognition. This paper deals with all above three steps, most work is described as follows:(1) We present a method of face detection based on subspace feature. In this paper, K-L transformation is used to get subspace of eigenfaces. Then projection coefficients on the subspace can be obtained. Based on this, blur cluster and weighted k-nearest neighbor algorithm are adopted to detect faces, which greatly enhances the speed of face detection.(2) In face recognition based on subspace feature, we propose local NMF for extraction of subspace of eigenfaces in order to solve the problem that NMF can’t extract local feature. Our method can fully extract local feature of image, which is helpful in enhancing recognition rate. Meanwhile, Bagging algorithm is used, which improves classification accuracy and generalization of neural network.(3) In face recognition based on geometrical feature, we present a method of localization in order to overcome the problem that eyes can’t be located accurately. Firstly we localize eyes according to geometrical feature, then we use eigeneyes template to validate the result.(4) We present a metasynthesis method of face recognition, considering that subspace and geometrical feature reflect different feature information. In metasynthesis, we present a method of deciding dynamically estimator of weights based on neural network and genetic algorithm, in order to overcome the problem that weights are decided unskillfully.The experimental results prove that recognition rate is higher basedon our methods than that based on individual geometrical feature or individual subspace feature.更多还原