·300 智能系统学报 第7卷 的综述.通过研究相关向量机的模型、回归算法以及 sing RVM with pre-segmentation and RANSAC[C]//2007 分类算法，有效地将相关向量机应用到多个领域当 IEEE Interational Geoscience and Remote Sensing Sympo 中.尤其着重介绍了相关向量机分类方法在高光谱 sium (IGARSS).Barcelona,Spain,2007:1763-1766 图像分类中的应用. [6]BISHOP C M,TIPPING M E.Variational relevance vector 高光谱数据的特点决定了相关向量机非常适合 machines[C]//Proceedings of the 16th Conference on Un- certainty in Artificial Intelligence.Catalina Island,USA, 其应用，无论是检测还是分类问题，相关向量机都有 2010:4653 着较好的表现.高光谱图像分类的评价标准有许多 [7]LUKIC A S,WERNICK M N,TZIKAS D G,et al.Bayes- 种，但是最主要的还是分类精度和分类效率.分类精 ian kernel methods for analysis of functional neuroimages 度主要是体现了该方法对高光谱数据的分类准确 [J].Medical Imaging,2007,26(12):1613-1624. 度，在这方面，相关向量机可以达到较高的分类精 [8]GHOLAMI B,HADDAD W M,TANNENBAUM A R.Rel- 度，满足分类的要求；分类效率主要是体现该方法在 evance vector machine learning for neonate pain intensity as- 分类过程中的空间复杂度以及时间复杂度，相关向 sessment using digital imaging[J].Biomedical Engineer- 量机的空间复杂度非常低，是支持向量机的15%左 ing,2010,57(6):1457-1466. 右，但其时间复杂度却比较高，尤其体现在大训练样 [9]AGARWAL A,TRIGGS B.Recovering 3D human pose 本的训练时间上.因此对于如何减少训练时间，提高 from monocular images[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2006,28(1):44-58. 训练效率也是今后进一步研究相关向量机的重点方 [10]GUO Feng,QIAN Gang.3D human motion tracking using 向之一 manifold leamning[C]//2007 IEEE Interational Confer 在现有的改进方法中，主要分为两大类，一是对 ence on Image Processing.San Antonio,USA,2007:357 算法本身的改进，也叫改进型相关向量机；另外一类 360. 是对算法中应用到的核函数进行改进.2种改进方 [11]SEDAI S,BENNAMOUN M,HUYNH D.Context-based 法都可以提高相关向量机分类的效率，并且保证了 appearance descriptor for 3D human pose estimation from 较高的分类精度.对于整个高光谱图像分类研究领 monocular images[C]//Proceedings of the 2009 Digital 域，如何进一步挖掘各种方法间的优点，拓展其应用 Image Computing:Techniques and Applications.Mel- 范围是该领域未来发展的重点 boure,Australia,2009:484-491. [12]GUO Weiwei,PATRAS I.Discriminative 3D human pose 目前相关向量机的研究已经取得了一定进展， estimation from monocular images via topological preserving 但其方法还是存在着一些不足，如何改进这些算法， hierarchical affinity clustering[C]//2009 IEEE 12th Inter- 突破制约技术发展瓶颈，是未来其理论研究的目标. national Conference on Computer Vision Workshops (ICCV 本文仅对其进行综合分析与论述，以达到抛砖引玉 Workshops).Kyoto,Japan,2009:9-15. 之功效 [13]OUYANG Yi,ZHANG Sanyuan.RVM-based classification 参考文献： of the network video surveillance system[C]//WASE In- terational Conference on Information Engineering. [1]TIPPING M.Sparse Bayesian learing and the relevance Taiyuan,China,2009:144-148. vector machine[J].Machine Learning Research,2001,6 [14]OLIVERI G,ROCCA P,MASSA A.A Bayesian compres- (1):211-244. sive sampling based inversion for imaging sparse scatters [2]DUAN Qing,ZHAO Jianguo,MA Yan.RVM and SVM for [J].IEEE Transactions on Geoscience and Remote Sens- classification in transient stability assessment[C]//2010 A- ing,2011,49(10):39934006. sia-Pacific Power and Energy Engineering Conference (AP- [15]LI Xiangru,LI Xiaoming,CHENG Xuezhen.Rejecting PEEC).Chengdu,China,2010:1-4. mismatches between fish-eye camera images by RVM [3]BILGIN G,ERTURK S,YILDIRIM T.Segmentation of hy- [C]//5th International Conference on Natural Computa- perspectral images via subtractive clustering and cluster vali- tion.Tianjin,China,2009:293-295. dation using one-class support vector machines[J].IEEE [16]WANG Xiaofei,ZHANG Junping,ZHANG Ye.Registra- Transactions on Geoscience and Remote Sensing,2011,49 tion of remote sensing images based on the relevance vector (8):2936-2944. machine [C]//5th International Conference on Natural [4]CAMPS V G,MARTINEZ R M.Nonlinear system identifica- Computation.Tianjin,China,2009:547-551. tion with composite relevance vector machines[].IEEE [17]DEMIR B,ERTURK S.Hyperspectral image classification Signal Processing Letters,2007,14(4):279-282. using relevance vector machines[J].IEEE Geoscience and [5]DEMIR B,ERTURK S.Hyperspectral data classification u- Remote Sensing Letters,2007,4(4):586-590