中文摘要：

局部保持投影(LPP)算法,作为Laplacian特征映射的线性扩展表示,解决了非线性算法难以获得新样本低维投影的缺点。但是,由于它是一种没有标签信息的无监督方法,不能利用已知的类别信息提高分类率。为此,引入类别信息重新定义权值矩阵,提出带有局部类别信息的局部鉴别投影,使相邻近的同类样本尽量紧凑,相邻近的异类样本尽量分开。同时,针对局部鉴别投影不能充分利用全局类别信息的缺陷,提出最大类间相斥的局部鉴别投影(MRLDP),通过最大程度地扩大异类间均值中心距离,带动异类间样本分散分布,最终得到既保持内在流形信息又使异类距离最大化的投影矩阵,并将此算法用于人脸识别。该方法在ORL人脸库和Yale人脸库上进行了验证,取得了较好的人脸识别效果。

英文摘要：

As the Iinear representation of LapIacian Eigenmap, the IocaI preserving projection (LPP) is used to achieve the Iow dimensionaI projection of the new sampIe. However, LPP is a kind of unsupervised method that ignored the IabeI information. To compact the sampIes of same cIass and separate the sampIes of dif erent cIass, the IocaI dif erentiaI projection is proposed to take the category information into consideration. At the same time, according to deficiency of gIobaI category information, LocaIity discriminate projection with maximum intercIass repuIsion (MRLDP) is put forward. Maximizing the center of the sampIes of dif erent cIass wiI cause the scat er of the whoIe sampIes. FinaI y, the projection matrix with inner manifoId structure and Maximum distance of dif erent cIass is obtained. This method achieved bet er ef ect on the ORL data set and YaIe face database.