针对非平行语料非联合训练条件下的语音转换,提出一种基于倒谱本征空间结构化高斯混合模型的方法。提取说话人语音倒谱特征参数之后,根据其散布矩阵计算本征向量构造倒谱本征空间并训练结构化高斯混合模型SGMM-ES(Structured Gaussian M...针对非平行语料非联合训练条件下的语音转换,提出一种基于倒谱本征空间结构化高斯混合模型的方法。提取说话人语音倒谱特征参数之后,根据其散布矩阵计算本征向量构造倒谱本征空间并训练结构化高斯混合模型SGMM-ES(Structured Gaussian Mixture Model in Eigen Space)。源和目标说话人各自独立训练的SGMM-ES根据全局声学结构AUS(Acoustical Universal Structure)原理进行匹配对准,最终得到基于倒谱本征空间的短时谱转换函数。实验结果表明,转换语音的目标说话人平均识别率达到95.25%,平均谱失真度为1.25,相对基于原始倒谱特征空间的SGMM方法分别提高了0.8%和7.3%,而ABX和MOS测评表明转换性能非常接近于传统平行语料方法。这一结果说明采用倒谱本征空间结构化高斯混合模型进行非平行语料条件下的语音转换是有效的。展开更多
标记分布学习是近年来提出的一种新的机器学习范式,它能很好地解决某些标记多义性的问题。现有的标记分布学习算法均利用条件概率建立参数模型,但未能充分利用特征和标记间的联系。本文考虑到特征相似的样本所对应的标记分布也应当相似...标记分布学习是近年来提出的一种新的机器学习范式,它能很好地解决某些标记多义性的问题。现有的标记分布学习算法均利用条件概率建立参数模型,但未能充分利用特征和标记间的联系。本文考虑到特征相似的样本所对应的标记分布也应当相似,利用原型聚类的k均值算法(k-means),将训练集的样本进行聚类,提出基于kmeans算法的标记分布学习(label distribution learning based on k-means algorithm,LDLKM)。首先通过聚类算法kmeans求得每一个簇的均值向量,然后分别求得对应标记分布的均值向量。最后将测试集和训练集的均值向量间的距离作为权重,应用到对测试集标记分布的预测上。在6个公开的数据集上进行实验,并与3种已有的标记分布学习算法在5种评价指标上进行比较,实验结果表明提出的LDLKM算法是有效的。展开更多
The texture vectors, as the fundatnental elements of the vector method (VM), are usually determined by the Durand's iterative method. In present paper, the texture vector is derived by two kinds of the maximum ent...The texture vectors, as the fundatnental elements of the vector method (VM), are usually determined by the Durand's iterative method. In present paper, the texture vector is derived by two kinds of the maximum entropy method (MEM), which choose pole figure data (MEM(I)) and the series coefficients of pole figures (MEM(II)), respectively, as a constrained condition. The detailed comparisons, including the texture vector and residual vector in the pole figure and ODF, among the results obtained by different methods are given through the ideal fiber texture simulation with Gaussian distribution. It is demonstrated that, although both methods the good results in the ideal texture simulation, the solution on assumption of maximum entropy displays more attractive results. In order to compare the sensitivity of the different methods to the experimental errors, the stochastical errors in pole figures are introduced by the computer random processes (Monte-Carlo simulation). The Monte-Carlo simulation shows that the MEM with the series coefficients as a constrained condition is rather sensitive to the 'experimental' errors, however, inversely the conventional VM and MEM with pole figure data as a constrained condition.展开更多
After the existence of phlebotomine sand flies was first reported in China in 1910,the distribution of different species and their role in the transmission of visceral leishmaniasis(VL)have been extensively studied.Up...After the existence of phlebotomine sand flies was first reported in China in 1910,the distribution of different species and their role in the transmission of visceral leishmaniasis(VL)have been extensively studied.Up until 2008,four species have been verified as vectors of VL,namely,Phlebotomus chinensis(Ph.sichuanensis),Ph.longiductus(Ph.chinensis longiductus),Ph.wui(Ph.major wui),and Ph.alexandri.The sand fly species vary greatly depending on the natural environments in the different geographic areas where they are endemic.Ph.chinensis is euryecious and adaptable to different ecologies,and is thus distributed widely in the plain,mountainous,and Loess Plateau regions north of the Yangtze River.Ph.longiductus is mainly distributed in ancient oasis areas south of Mt.Tianshan in the Xinjiang Uygur autonomous region.Ph.wui is the predominant species in deserts with Populus diversifolia and Tamarix vegetation in Xinjiang and the western part of the Inner Mongolia autonomous region.Finally,Ph.alexandri is steroecious and found only in stony desert areas,such as at the foot of the mountains in Xinjiang and the western Hexi Corridor,in Gansu province.This review summarized the relationship between the geographic distribution pattern of the four sand fly species and their geographical landscape in order to foster research on disease distribution and sand fly control planning.Furthermore,some problems that remained to be solved about vectors of VL in China were discussed.展开更多
Soil bulk density(BD) is an important physical property and an essential factor for weight-to-volume conversion. However, BD is often missing from soil databases because its direct measurement is labor-intensive, time...Soil bulk density(BD) is an important physical property and an essential factor for weight-to-volume conversion. However, BD is often missing from soil databases because its direct measurement is labor-intensive, time-consuming, and sometimes impractical, particularly on a large scale. Therefore, pedotransfer functions(PTFs) have been developed over several decades to predict BD. Here, six previously revised PTFs(including five basic functions and stepwise multiple linear regression(SMLR)) and two new PTFs, partial least squares regression(PLSR) and support vector machine regression(SVMR), were used to develop BD-predicting PTFs for coastal soils in East China. Predictor variables included soil organic carbon(SOC) and particle size distribution(PSD). To compare the robustness and reliability of the PTFs used, the calibration and prediction processes were performed 1 000 times using the calibration and validation sets divided by a random sampling algorithm. The results showed that SOC was the most important predictor, and the revised PTFs performed reasonably although only SOC was included. The PSD data were useful for a better prediction of BD, and sand and clay fractions were the second and third most important properties for predicting BD. Compared to the other PTFs, the PLSR was shown to be slightly better for the study area(the average adjusted coefficient of determination for prediction was 0.581). These results suggest that PLSR with SOC and PSD data can be used to fill in the missing BD data in coastal soil databases and provide important information to estimate coastal carbon storage, which will further improve our understanding of sea-land interactions under the conditions of ongoing global warming.展开更多
提出了一种分层的和多分辨的镜头边界检测方法。该方法对各种不同的镜头间过渡类型给出了用不同方法进行联合检测的方案,该方案主要分为突变镜头检测(即视频切分)、淡化过渡检测、溶解过渡检测及划变过渡检测4个部分。检测方案并不是简...提出了一种分层的和多分辨的镜头边界检测方法。该方法对各种不同的镜头间过渡类型给出了用不同方法进行联合检测的方案,该方案主要分为突变镜头检测(即视频切分)、淡化过渡检测、溶解过渡检测及划变过渡检测4个部分。检测方案并不是简单地将各种方法拼接在一起,而是通过小波变换的多分辨分析将它们有机地结合起来,相互关联,达到有效检测结果。首先用FCM聚类算法进行视频切分,然后根据聚类结果分别在整数小波分解后的高频部分用Gaussian加权Hausdorff距离结合边界改变率算法检测淡化过渡;对分解后的低频部分用所提出的SCD算法(Similarity of color distribution based method)检测溶解过渡,并通过自适应调节权系数(系数盲调节)使检测相异度函数更能适用于多种视频片段。最后根据切分以及前面两种过渡检测的结果,利用三维小波分解后高频成分中的运动部分所定义的运动矢量来检测划变过渡。用实际视频数据所做的仿真实验结果表明,该方法不但能同时检测突变过渡和渐变过渡,而且能准确地判断渐变过渡的类型及其位置。此外,还能有效地抑制闪光、运动等的影响,从而提高了检测精度。展开更多
文摘针对非平行语料非联合训练条件下的语音转换,提出一种基于倒谱本征空间结构化高斯混合模型的方法。提取说话人语音倒谱特征参数之后,根据其散布矩阵计算本征向量构造倒谱本征空间并训练结构化高斯混合模型SGMM-ES(Structured Gaussian Mixture Model in Eigen Space)。源和目标说话人各自独立训练的SGMM-ES根据全局声学结构AUS(Acoustical Universal Structure)原理进行匹配对准,最终得到基于倒谱本征空间的短时谱转换函数。实验结果表明,转换语音的目标说话人平均识别率达到95.25%,平均谱失真度为1.25,相对基于原始倒谱特征空间的SGMM方法分别提高了0.8%和7.3%,而ABX和MOS测评表明转换性能非常接近于传统平行语料方法。这一结果说明采用倒谱本征空间结构化高斯混合模型进行非平行语料条件下的语音转换是有效的。
文摘标记分布学习是近年来提出的一种新的机器学习范式,它能很好地解决某些标记多义性的问题。现有的标记分布学习算法均利用条件概率建立参数模型,但未能充分利用特征和标记间的联系。本文考虑到特征相似的样本所对应的标记分布也应当相似,利用原型聚类的k均值算法(k-means),将训练集的样本进行聚类,提出基于kmeans算法的标记分布学习(label distribution learning based on k-means algorithm,LDLKM)。首先通过聚类算法kmeans求得每一个簇的均值向量,然后分别求得对应标记分布的均值向量。最后将测试集和训练集的均值向量间的距离作为权重,应用到对测试集标记分布的预测上。在6个公开的数据集上进行实验,并与3种已有的标记分布学习算法在5种评价指标上进行比较,实验结果表明提出的LDLKM算法是有效的。
文摘The texture vectors, as the fundatnental elements of the vector method (VM), are usually determined by the Durand's iterative method. In present paper, the texture vector is derived by two kinds of the maximum entropy method (MEM), which choose pole figure data (MEM(I)) and the series coefficients of pole figures (MEM(II)), respectively, as a constrained condition. The detailed comparisons, including the texture vector and residual vector in the pole figure and ODF, among the results obtained by different methods are given through the ideal fiber texture simulation with Gaussian distribution. It is demonstrated that, although both methods the good results in the ideal texture simulation, the solution on assumption of maximum entropy displays more attractive results. In order to compare the sensitivity of the different methods to the experimental errors, the stochastical errors in pole figures are introduced by the computer random processes (Monte-Carlo simulation). The Monte-Carlo simulation shows that the MEM with the series coefficients as a constrained condition is rather sensitive to the 'experimental' errors, however, inversely the conventional VM and MEM with pole figure data as a constrained condition.
基金supported by the 12th Five-Year Plan for the National Major Program(grant no.2012ZX10004219)the National S&T Major Program(grant no.2012ZX10004220).
文摘After the existence of phlebotomine sand flies was first reported in China in 1910,the distribution of different species and their role in the transmission of visceral leishmaniasis(VL)have been extensively studied.Up until 2008,four species have been verified as vectors of VL,namely,Phlebotomus chinensis(Ph.sichuanensis),Ph.longiductus(Ph.chinensis longiductus),Ph.wui(Ph.major wui),and Ph.alexandri.The sand fly species vary greatly depending on the natural environments in the different geographic areas where they are endemic.Ph.chinensis is euryecious and adaptable to different ecologies,and is thus distributed widely in the plain,mountainous,and Loess Plateau regions north of the Yangtze River.Ph.longiductus is mainly distributed in ancient oasis areas south of Mt.Tianshan in the Xinjiang Uygur autonomous region.Ph.wui is the predominant species in deserts with Populus diversifolia and Tamarix vegetation in Xinjiang and the western part of the Inner Mongolia autonomous region.Finally,Ph.alexandri is steroecious and found only in stony desert areas,such as at the foot of the mountains in Xinjiang and the western Hexi Corridor,in Gansu province.This review summarized the relationship between the geographic distribution pattern of the four sand fly species and their geographical landscape in order to foster research on disease distribution and sand fly control planning.Furthermore,some problems that remained to be solved about vectors of VL in China were discussed.
基金supported by the National Natural Science Foundation of China (Nos. 41877004 and 42130405)the China Scholarship Council (Nos. 201809040007 and 201808320124)。
文摘Soil bulk density(BD) is an important physical property and an essential factor for weight-to-volume conversion. However, BD is often missing from soil databases because its direct measurement is labor-intensive, time-consuming, and sometimes impractical, particularly on a large scale. Therefore, pedotransfer functions(PTFs) have been developed over several decades to predict BD. Here, six previously revised PTFs(including five basic functions and stepwise multiple linear regression(SMLR)) and two new PTFs, partial least squares regression(PLSR) and support vector machine regression(SVMR), were used to develop BD-predicting PTFs for coastal soils in East China. Predictor variables included soil organic carbon(SOC) and particle size distribution(PSD). To compare the robustness and reliability of the PTFs used, the calibration and prediction processes were performed 1 000 times using the calibration and validation sets divided by a random sampling algorithm. The results showed that SOC was the most important predictor, and the revised PTFs performed reasonably although only SOC was included. The PSD data were useful for a better prediction of BD, and sand and clay fractions were the second and third most important properties for predicting BD. Compared to the other PTFs, the PLSR was shown to be slightly better for the study area(the average adjusted coefficient of determination for prediction was 0.581). These results suggest that PLSR with SOC and PSD data can be used to fill in the missing BD data in coastal soil databases and provide important information to estimate coastal carbon storage, which will further improve our understanding of sea-land interactions under the conditions of ongoing global warming.
文摘提出了一种分层的和多分辨的镜头边界检测方法。该方法对各种不同的镜头间过渡类型给出了用不同方法进行联合检测的方案,该方案主要分为突变镜头检测(即视频切分)、淡化过渡检测、溶解过渡检测及划变过渡检测4个部分。检测方案并不是简单地将各种方法拼接在一起,而是通过小波变换的多分辨分析将它们有机地结合起来,相互关联,达到有效检测结果。首先用FCM聚类算法进行视频切分,然后根据聚类结果分别在整数小波分解后的高频部分用Gaussian加权Hausdorff距离结合边界改变率算法检测淡化过渡;对分解后的低频部分用所提出的SCD算法(Similarity of color distribution based method)检测溶解过渡,并通过自适应调节权系数(系数盲调节)使检测相异度函数更能适用于多种视频片段。最后根据切分以及前面两种过渡检测的结果,利用三维小波分解后高频成分中的运动部分所定义的运动矢量来检测划变过渡。用实际视频数据所做的仿真实验结果表明,该方法不但能同时检测突变过渡和渐变过渡,而且能准确地判断渐变过渡的类型及其位置。此外,还能有效地抑制闪光、运动等的影响,从而提高了检测精度。
