摘要
核匹配追踪算法在生成函数字典的过程中常采用贪婪算法进行全局最优搜索,导致算法学习时间过长。该文针对这一缺陷,提出一种基于近邻传播(Affinity Propagation,AP)聚类与核匹配追踪相结合的分类方法(AP-Kernel Matching Pursuit,AP-KMP),该方法利用聚类算法来优化核匹配追踪算法中的字典划分过程,使用近邻传播聚类将目标数据集划分为若干小型字典空间,随后KMP算法在小型字典空间进行局部搜索,从而缩短学习时间。针对部分UCI数据集和遥感图像数据集,分别采用AP-KMP算法与另4种经典算法进行分类比较实验,结果表明该文算法在时间开销和分类性能上均有一定的优越性。
The processing of generating dictionary of function in Kernel Matching Pursuit (KMP) often uses greedy Mgorithm for global optimal searching, the dictionary learning time of KMP is too long. To overcome the above drawbacks, a novel classification algorithm (AP-KMP) based on Affinity Propagation (AP) and KMP is proposed. This method utilizes clustering algorithms to optimize dictionary division process in KMP algorithm, then the KMP algorithm is used to search in these local dictionary space, thus reducing the computation time. Finally, four algorithms and AP-KMP are carried out respectively for some UCI datasets and remote sensing image datasets, the conclusion of which fully demonstrates that the AP-KMP algorithm is superior over another four algorithms in computation time and classification performance.
出处
《电子与信息学报》
EI
CSCD
北大核心
2014年第12期2923-2928,共6页
Journal of Electronics & Information Technology
基金
国家自然科学基金(11078001)
国家863计划项目(2012AA121602)资助课题
关键词
目标识别:近邻传播
核匹配追踪
分类
Target recognition
Affinity Propagation (AP)
Kernel Matching Pursuit (KMP)
Classification