摘要
传统的多流形判别分析(MMDA)方法要求每类样本数目必须相同,这在实际中往往很难满足,因此限制了它的应用。针对此问题,提出一种改进的多流形判别分析(IMMDA)方法。该方法去除了MMDA中的限制条件,用类内图和类间图来描述类内紧凑度和类间离散度,类内图可以代表子流形信息,类间图可以代表多流形信息,从而更好地实现分类。在FERET、ORL人脸库及UCI数据集上的实验证明了该方法的有效性。相比其他几种子空间学习方法,该方法取得了更好的识别效果。
Traditional multi-manifold discriminant analysis （MMDA） usually requires the number of samples in each class must be the identical, which is hard to be satisfied in real-world applications, so its application is restrained. To tackle this problem, we propose an improved multi-manifold discriminant analysis （IMMDA） method. IMMDA removes the limitation in MMDA. It uses within-class graph and between-class graph to describe the within-class compactness and the between-class separability. In addition, the within-class graph can represent the sub-manifold information, while the between-class graph can represent the multi-manifold information, so that better classification is achieved. The experiments carried out on FERET, ORL and the UCI datasets proves the effectiveness of IMMDA. It is superior to other learning methods in terms of the recognition performance.
出处
《计算机应用与软件》
CSCD
2015年第9期175-180,共6页
Computer Applications and Software
关键词
多流形学习
线性判别分析
局部保持投影
特征提取
Multi-manifold learning Linear discriminant analysis Locality preserving projection Feature extraction
