判别稀疏保持嵌入及其在人脸识别中的应用

Discriminant Sparsity Preserving Embedding with Application to Face Recognition

稀疏保持投影(SPP)是最近提出的稀疏子空间学习方法,该方法的目标是保持数据的稀疏结构关系。然而,SPP是无监督并且不适合分类任务。为了提取判别人脸特征,提出一种新的判别稀疏保持嵌入(DSPE)算法。DSPE在SPP的目标函数中引入了Fisher准则,强调了判别信息;另一方面,用施密特正交化获得人脸子空间的正交基向量,进一步增强了算法的识别性能。在ORL和FERET人脸库上的实验结果表明,本文提出的DSPE在特征提取和分类识别方面有更好的效果。

Sparsity Preserving Projections （SPP） is a recently proposed sparse subspace learning method which aims to preserve the sparse reconstructive relationship of the data. However, SPP is unsupervised and unsuitable for classification tasks. To extract the discriminant feature, Discriminant Sparsity Preserving Embedding （DSPE） is proposed. DSPE introduces Fisher criterion into the objective of SPP and emphasizes the discriminant information. On the other hand, Schmidt orthogonalizaiton is used to obtain the orthogonal basis vectors of the face subspace, which further enhances recognition performance. Experiments results on ORL and FERET face database indicate that the proposed DSPE has better effect on feature extraction and classification recognition.