改进的K均值聚类红外目标检测方法

利用图像方差能很好地反映目标边缘信息的特点,提出一种基于方差的K均值聚类红外目标检测算法。利用形态学方法对红外图像进行预处理,运用相应的模板计算得到红外图像的方差图像,利用K均值聚类算法对方差图像进行聚类,从而分离出目标类别和背景类别。实验表明,该算法提取的红外图像中目标信息的兰德指数最高,说明该算法能有效地提取红外图像中目标信息,从而达到目标检测的目的。     

一种基于非参数贝叶斯模型的聚类算法

鉴于聚类分析是机器学习和数据挖掘领域的一项重要技术, 并且与监督学习不同的是聚类分析中没有类别或标签的指导信息, 所以如何选择合适的聚类个数(即模型选择)一直是聚类分析中的难点. 由此提出了一种基于Dirichlet过程混合模型的聚类算法, 并用collapsed Gibbs采样算法对混合模型的参数进行估计. 新算法基于非参数贝叶斯模型的框架, 能够在不断的采样过程中优化模型参数并形成合适的聚类个数. 在人工合成数据集和真实数据集上的聚类实验结果表明: 基于Dirichlet过程混合模型的聚类算法不但能够自动确定聚类个数, 而且具有较强灵活性和鲁棒性.     

基于自适应径向基神经网络的类星体光谱自动识别方法

通过对光谱的研究来识别和认证类星体是天文学研究中的重要方法。文章提出了一种对类星体光谱进行自动识别的自适应径向基神经网络(RBFN)方法。该方法包括以下几个步骤: (1)先将训练样本归一化,再利用PCA变换进行降维,获得样本特征向量; (2)设计出K均值聚类算法与梯度下降法相结合的径向基神经网络结构的基本模型,再用SSE(sum of squares error)误差函数进行判断,对RBFN隐含层的神经元进行自动调节,直至满足给定误差阈值; (3)用训练得到的参数对用于测试的样本中的类星体光谱进行识别。该方法不但克服了经典RBFN算法选择隐层神经元数目的困难,而且还提高了对类星体识别的稳定性和正确率。研究结果对于大型光谱巡天所产生的海量数据的自动处理具有重要意义。     

Quasispecies distribution of Eigen model

We have studied sharp peak landscapes of the Eigen model from a new perspective about how the quasispecies are distributed in the sequence space. To analyse the distribution more carefully, we bring in two tools. One tool is the variance of Hamming distance of the sequences at a given generation. It not only offers us a different avenue for accurately locating the error threshold and illustrates how the configuration of the distribution varies with copying fidelity $q$ in the sequence space, but also divides the copying fidelity into three distinct regimes. The other tool is the similarity network of a certain Hamming distance $d_{0}$, by which we can gain a visual and in-depth result about how the sequences are distributed. We find that there are several local similarity optima around the centre (global similarity optimum) in the distribution of the sequences reproduced near the threshold. Furthermore, it is interesting that the distribution of clustering coefficient $C(k)$ follows lognormal distribution and the curve of clustering coefficient $C$ of the network versus $d_{0}$ appears to be linear near the threshold.     

Nonlinear Time Series Forecast Using Radial Basis Function Neural Networks

In the research of using Radial Basis Function Neural Network (RBF NN) forecasting nonlinear time series, we investigate how the different clusterings affect the process of learning and forecasting. We find that k-means clustering is very suitable. In order to increase the precision we introduce a nonlinear feedback term to escape from the local minima of energy, then we use the model to forecast the nonlinear time series which are produced by Mackey-Glass equation and stocks. By selecting the k-means clustering and the suitable feedback term, much better forecasting results are obtained.     

基于字符纹理的车牌定位算法

提出了一种基于小波变换和均值聚类的实时车牌定位算法,根据车牌字符的纹理特征,在小波变换域中结合均值聚类自适应动态阈值取得字符信息,然后通过连通区域分析,结合车牌的宏观特征对车牌进行定位,试验证明,该算法准确率高,抗干扰性强。     

Two Novel Methods to Enhance Network Synchronizability

In this paper, we propose two methods to enhance the synchronizability of a class of complex networks which do not hold the positive correlation between betweenness centrality （BC） and degree of a node, and observe other topology characteristics of the network affected by the methods. Numerical simulations show that both methods can effectively enhance the synchronizability of this kind of networks. Furthermore, we show that the maximal BC of all edges is an important factor to affect the network synchronizability, although it is not the unique factor.     

数学建模中的高维数据挖掘技术优化研究

高维数据挖掘由于特征空间占用开销较大,挖掘的复杂度较高,挖掘精度不高,为了提高对高维数据挖掘的准确性能,提出一种基于相空间重构和K-L变换特征压缩的高维数据挖掘数学建模方法。采用集成学习技术,对高维数据信息流进行相空间重构处理,考虑类间的数据不平衡性,求得高维数据的关联维特征参量,根据数据的链距离进行稀疏性融合,计算高维数据流模型的最大Lyapunove指数谱,根据谱分析方法实现数据聚类,对聚类后的数据采用K-L特征压缩方法进行降维处理,降低数据挖掘的内存及计算开销。仿真结果表明,采用该方法进行高维数据挖掘,数据挖掘的准确概率较高,占用内存消耗较少,计算开销较小。     

傅里叶变换近红外光谱结合RVM与新聚类算法鉴别灵芝孢子油多样掺假类别

食品掺假种类众多,手段隐蔽,成为食品安全检测一个重要难题。为摆脱传统模型识别食品中是否存在新掺假类别的局限性,实验以纯净的灵芝孢子油和掺杂不同比例花生油、玉米油、薏仁油、地沟油的五种类别为研究对象,采用傅里叶变换近红外光谱(Fourier transform near infrared spectroscopy, FT-NIR)收集12 400～4 000 cm-1范围内的近红外光谱。假设掺杂地沟油为新掺假类别,利用前四种类别的校正集样本构建相关向量机(RVM)多分类器,分别对建模的预测集样本和掺杂地沟油样本进行判别,并借助新聚类算法对判别为纯净的灵芝孢子油的样本做进一步分析验证。研究表明,RVM分类器对于建模的预测集样本判别准确率高达93.75%,说明模型有较强的判别能力,但由于模型局限性,掺杂地沟油样品被误判为纯净的灵芝孢子油;在新聚类算法的决策图上,纯净灵芝孢子油校正集和预测集混合样本的聚类中心数为1,而纯净灵芝孢子油校正集和掺杂了地沟油混合样本聚类中心数为2,直观验证判别结果的准确性。结果表明利用FT-NIR技术结合RVM分类器与新聚类算法对于灵芝孢子油掺假能够有效识别,并且能够定性识别新型掺假类型,为解决食品掺假多样化问题提供一种新思路。     

Stochastic techniques for global optimization: A survey of recent advances

In this paper stochastic algorithms for global optimization are reviewed. After a brief introduction on random-search techniques, a more detailed analysis is carried out on the application of simulated annealing to continuous global optimization. The aim of such an analysis is mainly that of presenting recent papers on the subject, which have received only scarce attention in the most recent published surveys. Finally a very brief presentation of clustering techniques is given.