Local Subspace Classifier with Prior Knowledge for Image Classification(Pattem Recognition)

概要

論文の詳細を見る
Linear subspace methods are widely used technique in pattern recognition systems because they can achieve dimension reduction and classification concurrently. However, the recognition rates of them decrease when data are not separable linearly. In order to overcome this difficulty, a local subspace classifier (LSC) and kernel nonlinear subspace methods have been proposed in the past. We can achieve high accuracy using these methods even if data are not separable linearly. In this paper, we combine linear LSC and prior knowledge about transformation-invariance for handwritten digit pattern classification. In addition, we derive kernel LSC for incorporating prior knowledge into LSC via kernel mappings. The performance of the proposed method is verified with the experiments on handwritten digit and color image classification.
社団法人電子情報通信学会の論文
2006-11-17