摘要
局部保持投影算法仅能保持近邻样本的局部结构,无法保证提取的特征有利于后续分类识别。为此,提出一种半监督保持投影特征提取算法。SPP算法能够利用标记样本所携带的类别信息来约束未标记样本,从而提高样本的可分性;同时,还在目标函数中加入一正则项,避免了因矩阵奇异导致算法无法求解的问题。利用实际高光谱数据进行对比实验,结果表明,用SPP算法进行特征提取后的分类精度较LPP算法有显著提升,验证了它的有效性。
Since locality preserving projections (LPP) only preserves the local structure and cannot guarantee the extracted features helpful for classification, a feature extraction algorithm of semi-supervised preserving projections (SPP) is proposed. The proposed method can use the classification information carried by the labeled samples to restrain the unlabeled samples, so as to improve the divisibility of samples. Moreover, the problem of singular matrix is avoided by adding a regularization term to its objective function. Experiments on hyperspectral data demonstrate that the classification accuracy of SPP is significantly higher than that of LPP.
出处
《现代电子技术》
2011年第13期74-77,80,共5页
Modern Electronics Technique
关键词
局部保持投影
特征提取
半监督
高光谱
locality preserving projection
feature extraction
semi-supervised learning
hyperspectral data