Abstract: Nearest Neighbor Convex Hull classifier(NNCH) involves solving convex quadratic programming problems, which require large memory and enormous time for large-scale problem. Therefore, it is important for NNCH to reduce the computation complexity and memory requirement without degrading the prediction accuracy. In this paper, a sample selection method named Subspace Sample Selection(SSS) algorithm is used to select a subset of data for NNCH. The SSS algorithm is a one-class iterative algorithm, which selects the furthest sample to the subspace of the chosen set in one class at each step. The experiments on the training-synthetic subset of MIT-CBCL face recognition database show that a significant amount of training samples can be removed, and the computation time of NNCH can be significantly reduced without any loss in accuracy.

Key words: pattern recognition, face recognition, sample selection, nearest neighbor convex hull, subspace sample selection

摘要： 最近邻凸包分类器需要求解测试样本到训练集凸包距离的凸二次规划问题，对于训练集规模较大的情况，有必要在分类之前进行适当的样本选择。为此该文提出基于子空间样本选择的最近凸包分类方法。该方法首先采用子空间样本选择算法对训练集样本进行筛选，然后将各类选出的样本作为最近邻分类器的新的训练集。子空间样本选择方法的原理是在一类训练样本集内，迭代选择距离已选样本张成子空间最远的样本。在MIT-CBCL人脸识别数据库的training-synthetic子库的实验中，该方法只需5.6%的训练样本即可取得100%的识别率，并且执行时间较未经选样的最近邻凸包分类器也大为减少。

关键词: 模式识别, 人脸识别, 样本选择, 最近邻凸包, 子空间样本选择

CLC Number: 

• TP391