Abstract: For the robust face recognition problem with high-dimensional small sample, the algorithm of spectral regression classification optimized by local linear embedding is proposed. Firstly, feature vectors of training samples are calculated. Then, local linear embedding is used to construct embedding needed by classification and embeddings needed by sub-manifold of each classification is learned. Finally, spectral regression classification algorithm is used to compute project metrics, and nearest neighbor classifier is used to recognize face. Experimental results on the common face datasets FERET, AR and Extended YaleB show that proposed algorithm has better recognition efficiency than several other spectral regression algorithms.

Key words: robust face recognition, illumination variation, local linear embedding, spectral regression, manifold learning

摘要： 针对高维小样本鲁棒人脸识别问题，提出了一种局部线性嵌入优化光谱回归算法。计算出训练样本的特征向量，然后用局部线性嵌入算法构建分类问题所需的嵌入，并学习每种分类的子流形所需的嵌入；利用光谱回归计算投影矩阵，最近邻分类器完成人脸的识别。在人脸数据库FERET、AR及扩展YaleB上的实验结果表明，相比其他几种光谱回归算法，该算法取得了更好的识别效果。

关键词: 鲁棒人脸识别, 光照变化, 局部线性嵌入, 光谱回归, 流形学习