基于包围盒约束光谱聚类的红外目标识别算法

在红外成像过程中,目标边缘模糊化是影响红外目标识别效果的关键因素,也是红外目标识别算法的研究重点,故在光谱图像中合理补偿目标几何特征信息成为研究热点之一。结合包含目标几何特征信息的包围盒作为约束条件,对红外光谱图像进行分层限定滤波,降低原有图像数据中目标几何外形数据的丢失,提高目标可识别性。设计了在包围盒约束条件下的光谱聚类算法,设置参数η表征待测军用车辆目标的几何信息,设置参数m表征待测军用车辆目标的光谱特征信息。实验采用TEL-1000-MW型红外成像光谱仪获取多光谱图像,通过改变m和η值调整光谱特征值个数与包围盒范围,从而获得不同的目标识别图像。并与传统方法对同一幅红外目标图像的识别效果相比较,结果发现采用包围盒约束的待测目标图像几何边界信息保留效果明显优于传统方法,当m=10、η=0.7时,红外图像的目标识别效果最好,同时算法收敛速度也最优。由此可见,该算法在提高红外目标识别能力、避免误判伪目标和漏检目标方面具有很高的实用价值。

高光谱遥感数据波谱相似性模糊聚类方法研究

高光谱遥感数据的聚类是定量遥感研究的热点。为提高其聚类精度,将光谱相似性测度与模糊动态聚类法有机结合,构建四种模糊聚类新方法:光谱角度-模糊聚类法、光谱信息熵-模糊聚类法、光谱角度与光谱信息熵混合-模糊聚类法与海明距离-模糊聚类法。以实测植被高光谱信息为数据,利用该四种方法进行比较分析其聚类精度。实验结果表明光谱角度与光谱信息熵混合模糊聚类法以及光谱角度模糊聚类法的聚类精度均相对其他几种方法较好,是最有效高光谱遥感聚类方法。为今后该领域的方法的进一步研究提供一种借鉴与参考。

一种基于深度谱空网络的高光谱图像聚类方法

通常无监督算法在对高光谱数据进行聚类时仅使用光谱信息,忽略了空间信息,使得聚类准确率较低。针对上述问题提出一种基于深度谱空网络和无监督判别极限学习的高光谱图像聚类算法。利用深度谱空网络对高光谱数据进行光谱特征和空间特征的分层交叉学习,通过反复学习获得深度空谱特征,为后续无监督聚类提供方便。在三种高光谱图像上进行实验,结果表明,该算法获得的聚类效果优于其他基于极限学习机的方法和其他无监督方法。

基于运动轨迹聚类的运动分割

文中提出了一种基于子空间聚类的运动分割算法。针对基于因式分解的运动分割方法对噪声敏感的问题,引入一个基于形状相关矩阵的相似矩阵,然后将特征点映射到一个由相似矩阵决定的低维子空间中。在这个子空间中,可以通过光谱图聚类方法对特征点进行聚类。为提高光谱图聚类方法的分段常数特征向量条件,文中计算了一个相关的概率矩阵的较大特征值相对于边权重变化的敏感度,来切断类之间的连接。通过实验,可以看到文中的方法有较好的鲁棒性。

基于野外实测光谱统计分析的蚀变矿物填图

提出了一种新的基于野外实测光谱统计分析的蚀变矿物填图方法。该方法首先对野外测量得到的大量样本光谱数据进行聚类处理,从光谱的整体特征上将不同类型的样本区分开。第二步对各聚类结果中不同蚀变矿物分别建立其各自的判别函数。第三步在遥感影像上按照聚类得到的参考光谱分别进行大蚀变类的划分,最后在此基础上采用第二步得到的各蚀变类型的判别函数进行蚀变矿物的细化填图,得到最终的填图结果。该方法充分考虑了不同蚀变类型及蚀变组合的地区差异性,建立的蚀变矿物的判别函数更具有科学性,并且填图结果可在一定程度上进行可靠度评价。将该方法应用到新疆包古图地区的某一研究子区中,获得了较好的成果。

Study on FTIR Spectra of Corn Germs and Endosperms of Three Different Colors Combining with Cluster Analysis

[Objective] This research aimed to study the FTIR spectra of corn germs and endosperms so as to provide a scientific way for identifying corn of different types. [Method] The corn germs and endosperms of three types were studied by using Fourier transform infrared spectroscopy（FTIR） technology, combined with cluster analysis. [Result] The overall characteristics of original FTIR spectra were basically similar within the range of 700-1 800 cm^-1. The FTIR spectra were mainly composed by the absorption peaks of polysaccharides, proteins and lipids. Within the wavelength range of 700-1 800 cm^-1, there were only tiny differences in original FTIR spectra among the corn germs and endosperms of three different types. The spectra were then processed by using first derivative and second derivative. The second derivative spectra were used for hierarchical cluster analysis（HCA）. The results showed that with the wavelength range of 700-1 800 cm^-1, the second derivative spectra of the 52 samples could be better clustered according to the tree types and corn germ and corn endosperm. The clustering correct rate reached 96.1%.[Conclusion] FTIR technology, combined with cluster analysis, can be used to identify different types of corn germs and endosperms, and it is characterized by convenience and rapidness.

Terahertz spectrum clustering of traditional Chinese medicine based on first derivative characteristics

In order to deal with the unclear absorption peak caused by the absorption peak overlap of traditional Chinese medicine(TCM)and other mixtures,a method of three unsupervised clustering algorithms as K-means,K-medoids and Fuzzy C-means(FCM)combined with the first derivative characteristics of terahertz absorption spectrum,is proposed to perform the terahertz spectra clustering of Sanchi and other three kinds of TCM compared with their easily-confused products(ECPs).These three unsupervised clustering methods complement the scope of the supervised learning classification method.The first derivative of the spectrum could amplify the difference in the absorption coefficient with different substances,so that the obvious absorption peak can be revealed.Experiments shows that these three clustering algorithms can achieve good results by combining the origin absorption coefficient with its first-order derivative as the characteristic data,and among which K-means does the best with the accuracy of95.32%.Compared with pure absorption coefficient data clustering,the accuracy in this study has been significantly improved,especially for the non-absorption-peak TCM classification.And the accuracy of K-means algorithm is improved by5.38%.Besides,clustering algorithms in this study have strong anti-interference ability to the error data.

Response of fuzzy clustering on different threshold determination algorithms in spectral change vector analysis over Western Himalaya, India

Abstract： Change detection is a standard tool to extract and analyze the earth＇s surface features from remotely sensed data. Among the different change detection techniques, change vector analysis （CVA） have an exceptional advantage of discriminating change in terms of change magnitude and vector direction from multispectral bands. The estimation of precise threshold is one of the most crucial task in CVA to separate the change pixels from unchanged pixels because overall assessment of change detection method is highly dependent on selected threshold value. In recent years, integration of fuzzy clustering and remotely sensed data have become appropriate and realistic choice for change detection applications. The novelty of the proposed model lies within use of fuzzy maximum likelihood classification （FMLC） as fuzzy clustering in CVA. The FMLC based CVA is implemented using diverse threshold determination algorithms such as double-window flexible pace search （DFPS）, interactive trial and error （T＆E）, and 3x3-pixel kernel window （PKW）. Unlike existing CVA techniques, addition of fuzzy clustering in CVA permits each pixel to have multiple class categories and offers ease in threshold determination process. In present work, the comparative analysis has highlighted the performance of FMLC based CVA overimproved SCVA both in terms of accuracy assessment and operational complexity. Among all the examined threshold searching algorithms, FMLC based CVA using DFPS algorithm is found to be the most efficient method.

基于用户画像特征的优质电力园区多维度能耗预测模型

为了精准预测多维度能耗,文章研究基于用户画像特征的优质电力园区多维度能耗预测模型.利用光谱双聚类算法,按照用户用电类别维度,聚类用户用电数据,得到用户用电类别画像特征;按照用户需求响应维度,以负荷模式、分时用电量变异系数、光伏出力交易量为指标,建立用户需求响应评价模型,得到用户需求响应画像特征:利用模糊聚类算法聚类处理两个维度的画像特征,建立优质电力园区多维度用户画像,确定径向基函数神经网络结构:利用微粒群算法优化径向基函数神经网络的参数,确定径向基函数神经网络的多维度能耗预测模型:在多维度能耗预测模型内输入多维度用户画像,输出多维度能耗预测结果.实验证明:该模型可有效获取用户用电类别与需求响应两个维度的用户画像特征,建立多维度用户画像:该方法可精准预测优质电力园区多维度能耗.