基于多尺度稀疏近邻图的近邻保持嵌入算法

针对近邻保持嵌入算法NPE中构造近邻图所存在的缺陷,提出了基于多尺度稀疏近邻图的近邻保持嵌入算法.对于每个待识别的人脸图片,该方法都建立一个具有九个尺度的图像金字塔,并且计算金字塔中每个尺度的图片与其他图片金字塔对应尺度的稀疏近邻.利用稀疏表示算法抗遮挡的特性,通过计算样本多尺度近邻的方法克服了传统方法丢失人脸图片二维结构的缺点.结果表明,该算法具有较强的鲁棒性,比传统的NPE算法具有更好的识别效果.

Neighborhood preserving embedding algorithm based on multi-scale sparse neighbor graphs

Aiming at the deficiencies for establishing neighbor graphs in the neighborhood preserving embedding(NPE)algorithm, a NPE algorithm based on multi-scale sparse neighbor graphs was proposed. For each face image to be recognized, an image pyramid with nine scales for the image was established in the proposed method. The sparse neighbor between the image of each scale in the pyramid and the corresponding scale of other image pyramid was calculated. With the anti-overlap characteristics of sparse representation algorithm, the defects of losing the two-dimensional structure of face images with the method through calculating the sample multi-scale neighbor, were overcome. The results show that the proposed algorithm has stronger robustness, and possesses better recognition effect than the traditional NPE algorithm.