投影残差分类器

针对声音、图像等高维数据的分类问题,提出了一种快速算法。首先通过非线性特征映射,将各个类别的训练样本集转换到特征空间中,构造相应的特征子空间,然后提取它们的主要特征。特征映射能够降低特征子空间的维数,并增强它们之间的两两正交性,提高了分类的准确性。在进行分类时,该方法将测试样本向各个特征子空间投影,并计算投影残差,测试样本即为投影残差最小的特征子空间的样本。与传统的分类方法不同,快速算法能一次区分多个类别,并具有与支持向量机相同的准确率。又使用了流形学习理论对快速算法进行改进,在保持准确率的前提下,极大地降低了特征子空间的维数,验证了流形学习理论的应用价值。

Projection Residual Classifier

A fast algorithm for the classification of high dimensional vectors is proposed in this paper. Given training sample sets of multi classes, special nonlinear feature function is used to reshape these sets to low dimensional and orthogonal feature sub-spaces. By projecting a new coming vector on each feature sub-space we could compute the projection residuals. The new vector will be regarded as a sample of the feature sub-space with the smallest residual. This algorithm can distinguish high dimensional vectors of multi classes by one comparison and has good accuracy. Furthermore, manifold learning theory is added to the feature function. This keeps the accuracy but greatly reduces the dimensionality of feature sub-spaces. The validity of manifold learning theory is verified by doing that.