摘要
在图像信息处理中视觉词典生成过程需要对高维数据进行聚类操作.但这些高维数据不可避免会对计算机内存和计算能力提出更高要求.本文针对聚类过程中可能产生的内存耗尽以及初始聚类质心设置问题,对现有K-means算法加以改进.通过建立初始聚类质心与各类场景中的特定语义的关联,使之体现图像各类场景的类别特征集合,进而用于指导K-means过程中的初始质心设置.此外,在迭代过程中通过批次读入特征描述子,采用K近邻进行簇分配,从而避免了一次性读入全部特征描述子而造成的内存耗尽问题.同时,对于新的簇质心生成采用综合判别均值与中位值的办法来提高各族的聚合度.本文方法与Oxford University提出的K-means进行了对比,实验结果表明本文算法在性能与收敛上更具优势.
In order to generate a codebook for Image Information Index task,cluster has been done to deal with high dimension data. But it may give rise to problems,because it demands high memory capacity and fast operation of computers,in which computers will risk running out of memory. Since proper initial centers are very important to K-means which is used to do cluster. The paper proposes a novel way to adapt classic K-means to solve those problems by building correlation between the initial centers and special sematic contained in each image category scene. We read all image feature descriptors in groups, and distribute them into its individual group by KNN in the each iteration. After comparing mean vector and median vector, new cluster centers have been chosen in the iterations, with which can converge fast. We compare with the K-means proposed by Oxford University, our algorithm is jump on the rival to converge.
出处
《小型微型计算机系统》
CSCD
北大核心
2016年第8期1854-1856,共3页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61362024
61562035)资助
关键词
K均值聚类
视觉词典
图像高维特征描述
K近邻
K-means
visual dictionary
image high dimension feature descriptor
K-nearest neighbor
