摘要
分析了训练样本对于识别系统性能的影响,将训练样本分为三种:好样本、差样本和边界样本,分析了它们在训练中所起的作用,并结合基于HMM的手写数字识别系统,给出了一种简单的边界样本定义和选择的方法;通过实验证明了采用边界样本训练可使系统误识率降低17.51%,并证明了边界样本的重要性,且指出非边界样本的存在会影响训练的效果.
In the paper, the influence of training sample to the performance of recognition system is analyzed. The training samples are classified into three classes: good sample, poor sample and boundary sample. Combined with the handwritten numeral recognition system based on HMM, a simple method of definition and selection of boundary sample is given. The experimental result shows the miss-recognition rate is reduced by 17.51% by introducing boundary sample training, which verify the importance of boundary sample, the existence of non-boundary sample will influence the training effects.
出处
《计算机学报》
EI
CSCD
北大核心
2005年第11期1923-1928,共6页
Chinese Journal of Computers
基金
国家"八六三"高技术研究发展计划项目基金(2001AA114080)
国家自然科学基金(60475007)资助
关键词
训练样本
样本分类
边界样本
HMM
training sample
sample classifying
boundary sample
HMM