Abstract: This paper proposes simple SVM based on virtual vector transformation and local relationship feature according to example invariance features, and compares with tangent distance kernel, virtual SVM, human recognition etc by recognizing handwritten numbers in USPS. The method can acquire good raw error rate and its computation time is least. The results indicate that the method could sufficiently abstract invariance features in example data, and could be an effective measurement for pattern classification.

Key words: Invariance feature, Simple support vector machine, Virtual vector, Local relationship

摘要： 应用样本数据具有不变性特征的特点，提出基于虚拟向量变换和局部相关性特征的简易SVM(支持向量机)方法，并以USPS手写体数字识别为例子，与切距核、虚拟SVM、人工识别等方法进行对比，发现该方法能获得较好的粗识别率且计算时间最少。结果表明，该方法能充分提取样本数据中的不变性特征，是研究模式分类问题的有效方法。

关键词: 不变性特征, 简易支持向量机, 虚拟向量, 局部相关性