基于类矩阵和特征融合的加权自适应人脸识别

为了准确快速地进行人脸识别,提出了一种基于类矩阵和特征融合的加权自适应人脸识别算法,该算法首先,提取人脸的全局特征和6个关键部分的局部特征,同时给出了局部特征权值的动态选择方法,由于该法可以根据不同的训练集得出不同的权值,因而增强了算法的自适应能力;然后通过将全局和局部特征加权融合来得出样本的特征矩阵;接着设计出了一种加权PCA方法用于对样本矩阵进行降维;再进一步提出类矩阵的概念,同时给出并证明了类矩阵的推导公式,并据此得出一种新的投影准则;最后,将类矩阵和试验样本分别进行投影,并根据其欧氏距离的大小得出试验人脸的最终类别。试验表明,该算法不仅计算速度快、识别率高,而且能有效解决LDA小样本空间问题,应用前景良好。

Weighted Adaptive Face Recognition Based on Class Matrix and Feature Fusion

A new weighted adaptive algorithm of face recognition based on class matrix and feature fusion was proposed. Firstly, global features and local features of six key parts of faces were extracted respectively. Dynamic method of how to choose the weights of local features was given. Different weights could be gained for different training sets according to this method. So, the adaptive ability of algorithm was enhanced. Then, global and local features were fused with weights to get the eigen-matrix of samples. Secondly, a new weighted principal component analysis (PCA) method was designed to lower dimension for sample matrixes. Thirdly, the concept of class matrix was proposed, and formula of how to obtain the class matrix was given and proved. According to class matrix, a new projected rule was given. Finally, class matrix and tested samples were projected respectively through the proposed rules. Then, the final class that tested faces belonged to was declared according to the Euclidean distance. Experiments show that the proposed algorithm can deal with small sample problems in LDA effectively, and the results also indicate that it has good performance on speed and recognition rate.