一种硬件友好型自适应K–均值学习算法  

An Hardware-friendly Adaptive K–means Learning Algorithm

作　　者：侯作勋[1] 韩培[2] 张宏伟[1] 安然[1] HOU Zuoxun HAN Pei ZHANG Hongwei AN Ran(Beijing Institute of Space Mechanics ＆ Electricity, Beijing 100190, China Technology and Engineering Center for Space Utilization, Chinese Academy of Sciences, Beijing 100190, China)

机构地区：[1]北京空间机电研究所,北京100190 [2]中国科学院空间应用工程与技术中心,北京100190

出　　处：《航天返回与遥感》2017年第3期68-77,共10页Spacecraft Recovery & Remote Sensing

基　　金：国家重点研发计划(2016YFB0501300;2016YFB0501302)

摘　　要：文章提出了一种适合于嵌入式平台实现的自适应K–均值学习算法,用于解决标准K–均值算法中存在的无法自主确定类属数量、难以确定合理的初始化种子集和运算时间过长的问题。算法通过引入变异比准则(VRC)对聚类结果进行定量评估,并通过迭代运算寻找VRC最大值的方法有效解决了类属数量的自主确定问题;提出了一种分布式最大–最小初始化种子选择方法,利用渐进寻找类内距离最大样本的方法解决了K值递增时初始化种子集的确定问题;并给出了利用FPGA实现该算法的有效途径。仿真实验结果表明,该算法针对各种类型的样本向量均能够准确高效的完成聚类处理任务,VRC评估结果与理论预期一致,初始化种子集选择正确。为进一步实现目标分类、图像分割等智能图像处理任务奠定了基础。This paper proposes a hardware-friendly adaptive K–means learning algorithm to solve the basic problems of the standard K–means algorithm which include determining the cluster number and the reasonable initial seeds automatically, and improving the computing speed effectively. The proposed algorithm uses the variance ratio criterion （VRC） to quantatively evaluate the clustering result, and finds the optimized cluster number by seeking the maximal value of the VRC. The distributed max–min initial seeds selection method is proposed to find the optimized initial seeds for differentKby searching for the sample with maximal inner-cluster distance gradually. Also, this paper introduces the possible scheme of implementing the algorithm by FPGA. The simulations show that the proposed algorithm can finish the clustering tasks for different kinds of samples accurately and efficiently. The VRC evaluating results completely satisfy the theoretical prospective, and the found initial seeds are reasonable. It can be used in some intelligent image processing tasks, such as the object classification and image segmentation.

关 键 词：自适应 K–均值 硬件友好 图像处理 航天遥感 

分 类 号：TP72[自动化与计算机技术—检测技术与自动化装置]