摘要
针对基于内容的信息检索中负样本抽样效率低的问题,设计了1 5类支持向量分类器.在训练过程中利用正样本对分类线建立初始模型,在保证总体泛化能力的基础上,用所能获得的负样本修正分类线,以提高其检测精度;通过对比标准序列最小优化方法,得到快速训练算法.在美国邮政数据库(USPS数据库)与麻省理工大学人脸数据库(CBCL数据库)上的实验结果表明,与传统的支持向量分类器相比,这种方法能取得更高的检测精度.
To solve the problem of the low sampling efficiency in the content-based information retrieval, a new classifier named 1.5-class support vector classifier (1.5 SVC) is proposed. To improve the detection rate, the positive samples are adopted to model the initial boundaries, and the available negative samples are added to refine the boundaries on the basis of keeping a good global generalization performance. The fast training algorithm of the method is also given by contrast with standard sequential minimal optimization. Compared with the traditional support vector classifiers, the experimental results on USPS and CBCL database demonstrate that the new classifier yields a higher detection rate.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2003年第6期581-585,共5页
Journal of Xi'an Jiaotong University
基金
国家创新研究群体科学基金资助项目 (6 0 0 2 43 0 1 )
国家自然科学基金资助项目 (6 0 1 750 0 6 ).
关键词
信息检索
数据库
支持向量机
支持向量分类器
Classification (of information)
Database systems
Optimization
Sampling
