一种用于人脸识别的非线性鉴别特征融合方法

最近，在人脸等图像识别领域，用于抽取非线性特征的核方法如核Fisher鉴别分析(KFDA)已经取得成功并得到了广泛应用，但现有的核方法都存在这样的问题，即构造特征空间中的核矩阵所耗费的计算量非常大．而且，抽取得到的单类特征往往不能获得到令人满意的识别结果．提出了一种用于人脸识别的非线性鉴别特征融合方法，即首先利用小波变换和奇异值分解对原始输入样本进行降雏变换，抽取同一样本空间的两类特征，然后利用复向量将这两类特征组合在一起，构成一复特征向量空间，最后在该空间中进行最优鉴别特征抽取．在ORL标准人脸库上的试验结果表明所提方法不仅在识别性能上优于现有的核Fisher鉴别分析方法，而且，在ORL人脸库上的特征抽取速度提高了近8倍．

Nonlinear Discriminant Feature Fusion Method for Face Recognition

Recently, kernel methods for nonlinear feature extraction have achieved success and been widely applied in image recognition area such as face recognition. However, for current kernel methods, construction of kernel matrices in feature space has to consume enormous time. Moreover, only using one class of features for recognition can't usually achieve satisfied results. Considering both weakness mentioned above, a novel nonlinear discriminant feature fusion method for face recognition is developed in this paper. The main idea of this method is that wavelet transform and singular value decomposition are first employed to preprocess the original training images for reducing its dimensionality and two groups of reduced dimensional feature vectors from a same sample space are obtained. Then, the two sets of feature vectors are combined together through complex vector, which is used to establish a complex feature vector space. Finally, a modified kernel Fisher discriminant analysis(KFDA) is adopted to extract optimal nonlinear discriminant features in the complex vector space.. The experimental results on ORL face database indicate that the proposed method is more effective than the current KFDA. And, more importantly, its consumed time for feature extraction is only one eighth of that of KFDA.