基于K最近邻的L2,1范数稀疏回归分类器

在模式分类中．基于旋转不变范数的回归分类器（RRC）最近得到广泛的应用．然而RRC的稀疏重构是建立在全体训练样本之上．当训练样本的数量很大时，计算的时耗比较大．同时，对稀疏程度的过度追求也在一定程度上影响了分类性能．基于K最近邻分类器提出了一类局部的基于K最近邻的L2,1范数稀疏回归分类器（KNN—SRC），该分类器比全局的RRC计算速度快，同时。利用K最近邻点代替全体训练样本，在一定程上避免了非同类的相似样本对测试样本的过度稀疏表示，从而提高分类性能．KNN—SRC的分类性能在UCI的Wine数据集和Yale人脸数据库上作了检测．测试结果表明KNN—SRC分类性能优于RRC．

K-Nearest-Neighbor-based L2,1 norm sparse regression classifier

The Rotational-invariant-norm-based Regression for Classification （RRC） has been developed and shows great potential for pattern classification. RRC is a global representation based method in that a testing sample is represented by all training samples. Thus, on the one hand, it is time-consuming when the number of training samples is large; on the other hand, with the extremely sparse reconstructive coefficients, RRC sometimes will lead to misclassifications. This paper presents a local RRC method, called KNN-SRC, which chooses K nearest neighbors of a testing sample from all training sample to represent the testing sample. Since K is much smaller compared to the total number of training samples, KNN-SRC is much faster than the global RRC. More importantly, when there exists a class formed by parts of objects among many classes of objects, taking the K nearest neighbors as the training samples can avoid the misclassification. The proposed KNN-SRC is tested using the UCI Wine dataset and the Yale face database. The experimental results show KNN-SRC is more effective and efficient than RRC and other competitive methods.