基于虚拟样本图像集的多流形鉴别学习算法

Virtual sample image set based multimanifold discriminant learning algorithm

为了丰富训练样本的类内变化信息,提出了基于通用训练样本集的虚拟样本生成方法。为了利用生成的虚拟样本中的类内变化信息有效地完成单样本人脸识别任务,提出了基于虚拟样本图像集的多流形鉴别学习算法。该算法将每类仅有的单个训练样本图像和该类的虚拟样本图像划分为互不重叠的局部块并构建流形,然后为每个流形学习一个投影矩阵,使得相同流形内的局部块在投影后的低维特征空间间隔最小化,不同流形中的局部块在投影后的低维特征空间中间隔最大化。实验结果表明,所提算法能够准确地预测测试样本中的类内变化,是一种有效的单样本人脸识别算法。

This paper proposed a virtual sample generating method to enrich intra-class variation information of training samples based on generic training set. And then this paper proposed a virtual sample image set based multimanifold discriminant learning algorithm to utilize the intra-class variation information of generated virtual samples for efficiently performing single sample face recognition tasks. The algorithm divided the only single sample image of each class and the virtual sample images of the class into non-overlapping local patches firstly,and then modeled these local patches of each class as a manifold. Next,the algorithm learnt different projection matrices,which could make the distances between the local patches in same manifold minimized and the distances between the local patches in different manifolds maximized simultaneously in the low dimensional projection spaces,for different manifolds. Experiment results demonstrate that the algorithm can predict the intra-class variations of testing samples accurately and is an efficient single sample face recognition algorithm.