【作者】 钟清流；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 待识对象的特征检测是各种智能系统(如机器人,医疗诊断仪器等)实现智能信息处理的基础。其应用的日益广泛性、任务的复杂性、工作环境的不确定性和特殊性、其自身资源的有限性及特征检测的时效性、决定了其研究任务的挑战性和艰巨性。然而,对特征检测可靠性和安全性日益增加的需求,又不断加大了研究解决这一问题的紧迫性。因此,探索准确高效的特征检测理论、方法和技术已经成为研究的重要内容之一。本文主要探讨如何以基于机器学习及相关途径相结合的方式,利用有限系统资源高效实现实时特征检测的相关理论与技术问题。本文着重针对时序和图像特征检测方面的问题,提出了一些相应的算法。并验证了其有效性。本文的主要创新点包括(1).在研究分析OCSVM及PSO模型特点和综合前人研究成果基础上,提出OCSVM_CPSO组合式异常检测模型,实现了系统的自适应调节,解决了检测系统的在线运行问题,从而为其现实应用扫除了障碍。并将其应用于解决机器人传感器故障检测的实际问题。取得较好效果。(2).针对SAX模型容易丢失边界区信息问题,提出时序数据的DLS模型。它根据时序极值确定划分的上下边界,并根据最大熵确定最佳描述字符集,进而确定最佳划分间隔。从而能有效减少边界区的信息丢失;针对EXT_SAX模型缺陷,提出VSB模型,采用增加分量而非增加符号的途径来降低计算代价。且用实验证实它的有效性。(3).提出时序矢量符号的SFVS模型和相应的确定时序数据最大压缩比的方法,此模型能够比SAX提供更全面的描述信息,这有利于在时序特征检测的应用中实现更精确的分析。通过理论分析和实验证实了它的有效性。(4).针对自调节谱聚类缺陷,提出一种新的ASC(自适应谱聚类)算法。它用全局平均N近邻距离作为比例参数σ,利用本征矢差异来估计最佳聚类分组数k,这可在构造亲和力矩阵时减少计算代价,提高效率,并且更容易实现。在彩色图像检测与分割实际应用中的实验结果证实了其有效性。在此基础上提出改进聚类性能的相关半监督学习算法,且用实验证实了它的有效性。更多还原

【Abstract】 Character_detection for identifying object is foundation on which intellective Systems such as robot and medical diagnose actualize intellective information disposal. It’s increasingly universality in application,complexity in role, uncertainty and particularity in environmental, restricted self resource and real time demand of character detection , which have determined its research tasks are challenge and difficult . However , the demand ,for reliability and security of Character detection to be increasingly enhanced ,continuing enhance the pressure to resolve it.Thus,exploring true and resultful theory,method and technique about Character_detection has become one of important research contents.In this paper, the relative theory and technique how to effectively implement real time Character_detection are discussed by restricted system resources and based on conjunction machine learning with others approach.For issue of Character detection in time series and image , some relevant algorithm have been put forward and its validity have been validated. Main contributions of the dissertation are shown as following,By means of research and analysis the characteristic about the model of OCSVM and PSO and based on integration of father fruitm,an model of anomaly detection based on OCSVM_CPSO is put forward,which actualize that system adaptive adjustment and solve the problem about detecting system online run, and clean off obstacle for its real application. It is used as solving the real problem of fault detection of robot sensor,and a good result is abtainedThe DLS model is put forword in order to overcome the bug that SAX easily lose information on boundary .DLS partition fluctuate boundary according to extremum of the time series ,and select optimization character set based on the most entropy ,more, neatly set optimization partition interval,thus can effectively reduce losing information on boundary .Another Model, for solving the bug of the EXT_SAX,the VSB model is put forword, which increase the component to reduce calculating cost. Its availability is proved by the experiments.A vector symbolic model for Time Series Data Based on Statistic Feature,SFVS,and relavant mothod for estimating the most compress ratio about time series data, are put forward in order to surmount the bugs with which SAX Algorithm can not describe time series information fully,which is helpful to implement more accurate analysis in application of feature detection of time series.Its validity have been proved by experimentsFor the bug of Self-tuning spectral clustering, A new algorithm ASC(called as the adaptive spectral clustering), is put foward,which takes average distance of N-near-neighbour as scaling parameterσ, automatically estimates optimal clustering grouping k by means of information about enginvector difference .It can reduce calculating cost for constructing appetency matrix as well as get highter efficiency and implement easyer .Farther, the correlative semi supervised algorithm for improving its’performance are put forward on the base.The results of experiments in application about color image detection and segmentation shown that the algorithm is valid .更多还原