核映射和Rank-Order距离的局部保持投影相似性度量方法

针对近红外光谱高维、高冗余、非线性和小样本等特点导致光谱相似性度量时出现的“维度灾难”,提出一种基于核映射和rank-order距离的局部保持投影(KRLPP)算法。首先将光谱数据经过核变换映射到更高维空间,有效保证了流形结构的非线性特征。然后改进局部保持投影(LPP)算法对数据进行降维操作,将rank-order距离替代传统的欧氏距离或测地线距离,通过共享邻近点的信息,得到更加准确的局部邻域关系。最后在低维空间通过距离的计算实现光谱的度量。该方法不仅有效解决了高维空间存在的“距离失效”问题,同时还提高了相似性度量结果的精度。为了验证KRLPP算法的有效性,首先根据降维前后数据集信息残差的变化确定了最佳参数近邻点的个数k和降维后的维数d。其次,从光谱降维投影效果和模型分类效果两个角度与PCA,LPP和INLPP算法进行了对比,结果表明KRLPP算法对于烟叶的部位有较好的区分能力,降维效果以及对于不同部位的正确识别率明显优于PCA,LPP和INLPP。最后,从某品牌卷烟叶组配方中选取了5个代表性烟叶作为目标烟叶,分别采用PCA,LPP和KRLPP方法从300个用于配方维护的烟叶样品中为每个目标烟叶寻找相似烟叶,并从化学成分和感官评价两方面对替换前后的烟叶及叶组配方进行了评价分析。其中LPP和KRLPP用于降维的参数选择保持一致,PCA选择前6个主成分。结果表明,由KRLPP选出的替换烟叶与替换配方在总糖、还原糖、总烟碱、总氮等化学成分以及香气、烟气、口感等感官指标上较PCA、LPP方法差异最小,相似性度量准确度最高。该方法可应用于配方产品替换原料的查找,辅助企业实现产品质量的维护。

基于通勤时间距离与Rank-Order距离的LLE算法改进

局部线性嵌入(Local linear embedding,LLE)算法作为一种经典的非线性降维算法,在图像识别等领域取得了很好的应用效果,但仍存在一些缺陷,如在构造邻域图时使用欧氏距离,可能会出现“短路边”的情况,同时,会受到离群点的影响,导致鲁棒性较差。为解决以上问题,论文基于通勤时间距离(commute time distance,CTD)和Rank-Order距离提出了CRLLE(LLE based on CTD and Rank-Order distance)算法,并在ORL人脸数据集和IMM人脸数据集上进行实验。实验设置CRLLE算法与LLE算法、等距特征映射(Isomap)算法和主成分分析降维(PCA)算法三种维数简约方法进行比较,得出改进后的CRLLE算法的降维效果优于其他三种算法的结论。

基于改进局部保持投影算法的轴承故障特征提取

在局部保持投影算法中常采用欧氏距离来度量线性数据间的相似性,而在实际工程中欧氏距离并不能准确度量非线性数据间的相似性,从而使算法对高维数据的特征提取性能不佳。针对以上问题提出了基于相关性Rank-order度量的局部保持投影算法,首先应用相关系数计算样本间相关距离实现数据相似性的初步度量,在此基础上再计算Rank-order距离用于评估具有共享邻居的样本间相似性从而得到更有效的相似度矩阵并构建降维模型,最后通过在公开轴承故障数据集和实际采集的轴承数据集上的实验表明,与传统方法相比改进算法对非线性数据有较好的特征提取效果,证明了改进算法的有效性。

A detail-preserving random-valued impulse noise removal algorithm based on S-ROAD

In a random-valued impulse noise corrupted image, in order to remove impulse noise and, meanwhile, efficiently preserve image edges and details, a novel two-phase detail- preserving random-valued impulse noise removal algorithm is proposed. At the noise detecting phase, an image statistic called S-estimate based rank-ordered absolute difference （S- ROAD） is presented to distinguish image edge and detail pixels from impulse noise pixels in a noise corrupted image. By introducing S-estimate into ROAD statistic, the interference caused by the image edges and details in the ROAD statistic is eliminated. With the S-ROAD statistic, most of the noise pixels, including the noise at edges and details, can be distinguished. At the noise pixels filtering phase, a two-threshold iterative method is used to restore the identified noise pixels and the estimate precision is improved; thus, the image details can be efficiently preserved. Experimental results show that the proposed method provides a significant improvement over many existing filters in terms of both subjective and objective evaluations.

Power Law Distribution of Broadband Radiated Energy for Global Seismicity

The claim that there exists a transition of earthquake energy distribution between small and large earthquakes is checked using broadband radiated energy of earthquakes for global seismicity. Scattering of the relation between the magnitude and the broadband radiated energy makes it necessary to use the energy data directly. Rank-ordering statistics is applied to enhance the resolution in retrieving the power law distribution with undersampled data, namely, a few tens of events. Seen in the perspective of broadband radiated energy with higher resolution, there is no evidence for the kink in the frequency-energy distribution for large and small earthquakes. Instead, a single power law can well explain the data. For earthquakes with energy larger than 1014 J, we find that the number N of events with energy E depends on E via N ∝ E-B, with the scaling constant B = 0.64 ± 0.04, corresponding to b = 0.95± 0.06.