基于流形学习与SVM的手写字符识别方法

Manifold Learning and SVM Based Handwritten Numeral Recognition

本文结合核方法、局部线性嵌入(LLE)和支持向量机等机器学习方法,提出了一种集成手写字符维数约简、特征提取及识别方法。鉴于LLE方法对其近邻个数太过敏感,以及要求流形上的数据分布比较均匀,难以实现手写字符维数约简。本文提出的基于核局部线性嵌入方法(KLLE),能够选择最优的近邻个数、构造分布均匀流形,并克服了手写字符识别中由于书写习惯和风格不同造成字符模式不稳定的问题。使用MINST数据库中的手写数字进行仿真实验并利用PCA、LLE进行维数约简比较,验证了KLLE算法的有效性及优势。

This paper investigates kernel method, locally linear embedding （LLE） and support vector machine algorithms for dimensionality reduction, features extraction and recognition for handwritten numeral. Because of the LLE impressionable the number of nearest neighbors and uniform distribution manifold, this paper proposed the kernel method based locally linear embedding to select- ing the optimal number of nearest neighbors, constructing uniform distribution manifold, and overcome the instability of handwritten character model that is caused by different writing style. Furthermore, we have realized the emulation experiments of MINST data base. A full comparison of dimensionality reduction by PCA, LLE, and KLLE demonstrates that KLLE is effective approach.