基于多谱特征生成对抗网络的语音转换算法

语音转换在教育、娱乐、医疗等各个领域都有广泛的应用,为了得到高质量的转换语音,提出了基于多谱特征生成对抗网络的语音转换算法。利用生成对抗网络对由谱特征参数生成的声纹图进行转换,利用特征级多模态融合技术使网络学习来自不同特征域的多种信息,以提高网络对语音信号的感知能力,从而得到具有良好清晰度和可懂度的高质量转换语音。实验结果表明,在主、客观评价指标上,本文算法较传统算法均有明显提升。

基于注意力机制的多任务3D CNN-BLSTM情感语音识别

语音情感识别广泛应用于车载驾驶系统、服务行业、教育以及医疗等各个领域。为了使计算机能更准确地识别出说话人的情感,提出了一种基于注意力机制的多任务三维卷积神经网络(ConvolutionNeuralNetwork,CNN)和双向长短期记忆网络(BidirectionalLong-Short Term Memory,BLSTM)相结合的情感语音识别方法(3D CNN-BLSTM)。基于多谱特征融合组图,利用三维卷积神经网络提取深层语音情感特征,结合性别分类的多任务学习机制提升语音情感识别准确率。在CASIA汉语情感语料库上的实验结果表明,该方法获得了较高的准确率。

Long memory of price-volume correlation in metal futures market based on fractal features

An empirical test on long memory between price and trading volume of China metals futures market was given with MF-DCCA method. The empirical results show that long memory feature with a certain period exists in price-volume correlation and a fittther proof was given by analyzing the source of multifractal feature. The empirical results suggest that it is of important practical significance to bring the fractal market theory and other nonlinear theory into the analysis and explanation of the behavior in metal futures market.

Extracting Feature Bands for Damaged Rice Leaves by Planthoppers Using Multi-spectral Imaging Technology

[Objective] The aim of this study was to extract effective feature bands of damaged rice leaves by planthoppers to make identification and classification rapidly from great amounts of imaging spectral data. [Method] The experiment, using multi-spectral imaging system, acquired the multi-spectral images of damaged rice leaves from band 400 to 720 nm by interval of 5 nm. [Result] According to the principle of band index, it was calculated that the bands at 515, 510, 710, 555, 630, 535, 505, 530 and 595 nm were having high band index value with rich information and little correlation. Furthermore, the experiment used two classification methods and calcu-lated the classification accuracy higher than 90.00% for feature bands and ful bands of damaged rice leaves by planthoppers respectively. [Conclusion] It can be con-cluded that these bands can be considered as effective feature bands to identify damaged rice leaves by planthoppers quickly from a large scale of crops.

Analysis of rice paper's morphological features based on multispectral imaging technology

Computer forensics and identification for traditional Chinese painting arts have caught the attention of various fields. Rice paper＇s feature extraction and analysis methods are of high significance for the rice paper is an important carrier of traditional Chinese painting arts. In this paper, rice paper＇s morphological feature analysis is done using multi spectral imaging technology. The multispectral imaging system is utilized to acquire rice paper＇s spectral images in different wave- length channels, and then those spectral images are measured using texture parameter statistics to acquire sensitive bands for rice paper＇s feature. The mathematical morphology and grayscale statistical principle are utilized to establish a rice paper＇s morphological feature analytical model which is used to acquire rice paper＇ s one-dimensional vector. For the purpose of eval- uating these feature vectors＇ accuracy, they are entered into the support vector machine（SVM） classifier for detection and classification. The results show that the rice paper＇s feature is out loud in the spectral band 550 nm, and the average classifi- cation accuracy of feature vectors output from the analytical model is 96 %. The results indicate that the rice paper＇s feature analytical model can extract most of rice paper＇s features with accuracy and efficiency.

PROJECTION BASED STATISTICAL FEATURE EXTRACTION WITH MULTISPECTRAL IMAGES AND ITS APPLICATIONS ON THE YELLOW RIVER MAINSTREAM LINE DETECTION

Mainstream line is significant for the Yellow River situation forecasting and flood control.An effective statistical feature extraction method is proposed in this paper.In this method, a between-class scattering matrix based projection algorithm is performed to maximize between-class differences, obtaining effective component for classification;then high-order statistics are utilized as the features to describe the mainstream line in the principal component obtained.Experiments are performed to verify the applicability of the algorithm.The results both on synthesized and real scenes indicate that this approach could extract the mainstream line of the Yellow River automatically, and has a high precision in mainstream line detection.

Feature spectrum extraction of human fingernails based on LCTF multispectral imaging

A multispectral imaging system consisting of liquid crystal tunable filter(LCTF)and charge coupled device(CCD)camera was used to collect the images of fingernail samples at intervals of 10 nm during the spectral range of 450-1 000 nm,and a multispectral image of human fingernails containing 56 bands was obtained.The accurate reflectivity information of fingernails was obtained through referring whiteboard comparative measurement method.Principal component analysis(PCA)and band index method were used to reduce the dimension of the sample images respectively and two feature spaces were obtained.Spectral angle mapping(SAM)was used to classify human fingernails in these two feature spaces.The classification accuracy were above 92.5%and 82.9%respectively.Therefore,the feature space obtained by the PCA can be used as the characteristic spectrum of human fingernails,which provides a reliable basis for the analysis of multispectral spectrum of fingernails and human health assessment in the future.

Spectral Theorem of Many-Body Green's Functions When Complex Eigenvalues Appear

In this paper, an extended spectral theorem is given, which enables one to calculate the correlation functions when complex eigenvalues appear. To do so, a Fourier transformation with a complex argument is utilized. We treat all the Matsbara frequencies, including Fermionic and Bosonic frequencies, on an equal footing. It is pointed out that when complex eigenvalues appear, the dissipation of a system cannot simply be ascribed to the pure imaginary part of the Green function. Therefore, the use of the name fluctuation-dissipation theorem should be careful.

On the Spectral Radii of Bicyclic Graphs

A graph G of order n is called a bicyclic graph if G is connected and the number of edges of G is n＋1. Let B（n） be the set of all bicyclic graphs on n vertices. In this paper, the first three largest spectral radii in the class B（n） （n ≥9） together with the corresponding graphs are given.

An H^m-conforming spectral element method on multi-dimensional domain and its application to transmission eigenvalues

We develop an Hm-conforming(m 1) spectral element method on multi-dimensional domain associated with the partition into multi-dimensional rectangles. We construct a set of basis functions on the interval [-1, 1] that are made up of the generalized Jacobi polynomials(GJPs) and the nodal basis functions.So the basis functions on multi-dimensional rectangles consist of the tensorial product of the basis functions on the interval [-1, 1]. Then we construct the spectral element interpolation operator and prove the associated interpolation error estimates. Finally, we apply the H2-conforming spectral element method to the Helmholtz transmission eigenvalues that is a hot problem in the field of engineering and mathematics.