摘要
针对单核学习方法不能充分获取对象非线性特征的问题,提出一种核采样空间中的多核融合模型。与工作于隐式核空间的常见多核融合模型不同,该融合模型本质上是一种矩阵融合模型,其融合参数不受融合核矩阵半正定性要求的约束。在该模型基础上,进一步提出一种多核正则化Ho-Kashyap分类器,并设计了相应的迭代优化算法。最后,将该多核融合算法应用到水下钴结壳超声识别领域。实验结果表明,与单核学习方法相比,采用核采样空间多核信息融合模型的钴结壳超声识别分类正确率提高了7%,说明了该融合模型的有效性。
Single kernel learning method cannot obtain full nonlinear features of the objects to be recognized.A kind of multiple kernel fusion model in kernel sampling space is proposed.Unlike common multiple kernel fusion model in implicitly kernel space,the new fusion model is a kind of matrix fusion method in essence and the fusion parameters are not restricted by semi-definiteness of fusion kernel matrix.Based on the new fusion model,a kind of multiple kernel regularized Ho-Kashyap classifier is proposed and the related optimization algorithm is designed.At last,the fusion model is used in underwater cobalt crust echo recognition and experiment results show that the proposed multiple kernel fusion learning method is effective,which improves the performance of cobalt crust echo recognition by 7% in contrast with single kernel learning method.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2010年第2期248-252,共5页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(50474052
50875265)资助项目
