基于改进的监督LLE人脸识别算法

摘要

LLE is an unsupervised nonlinear dimensionality reduction method, widely used in facial feature extraction, but the method is lack of class label information of sample points. A new method is proposed based on the LLE with introducing supervised learning mechanism and increasing class label information of the sample points, by shrinking the intraclass distance while expanding the interclass distance to get the enhanced supervised locally linear embedding and minimizing the global reconstruction error of local data, combined with kernel neighborhood preserving projection method ( KN PP), using the method to extract high-dimensional nonlinear features of face data. The method is conducive to classification, and has better noise ro bustness. The experiment shows that recognition performance is better than LLE, SLLE and KNPP.%LLE是一种无监督的非线性降维方法,广泛应用于人脸特征提取,但是该方法缺乏样本点的类别信息.提出了一种新方法,在LLE的基础上引入有监督的学习机制和增加样本点的类别信息,通过减少类内距离而增加类间距离和最小化局部数据的全局重构误差,同时结合核邻域保持投影方法(KNPP)来提取高维人脸数据的非线性特征.算法有利于分类识别,并对噪声有较好的稳健性.实验表明,该方法的识别性能优于LLE,SLLE和KNPP.