A NEW CHARACTERIZATION OF THE CLOSEDNESS OF THE SUM OF TWO SUBSPACES

In this article, a new characterization of the closedness of the sum of two closed subspaces of a Hilbert space is established.

A Note on the Structure of Affine Subspaces of <i>L</i>²(R<i>^d</i>)

This paper investigates the structure of general affine subspaces of L2(Rd) . For a d × d expansive matrix A, it shows that every affine subspace can be decomposed as an orthogonal sum of spaces each of which is generated by dilating some shift invariant space in this affine subspace, and every non-zero and non-reducing affine subspace is the orthogonal direct sum of a reducing subspace and a purely non-reducing subspace, and every affine subspace is the orthogonal direct sum of at most three purely non-reducing subspaces when |detA| = 2.

Coupled Cross-correlation Neural Network Algorithm for Principal Singular Triplet Extraction of a Cross-covariance Matrix

This paper proposes a novel coupled neural network learning algorithm to extract the principal singular triplet (PST) of a cross-correlation matrix between two high-dimensional data streams. We firstly introduce a novel information criterion (NIC), in which the stationary points are singular triplet of the crosscorrelation matrix. Then, based on Newton's method, we obtain a coupled system of ordinary differential equations (ODEs) from the NIC. The ODEs have the same equilibria as the gradient of NIC, however, only the first PST of the system is stable (which is also the desired solution), and all others are (unstable) saddle points. Based on the system, we finally obtain a fast and stable algorithm for PST extraction. The proposed algorithm can solve the speed-stability problem that plagues most noncoupled learning rules. Moreover, the proposed algorithm can also be used to extract multiple PSTs effectively by using sequential method. © 2014 Chinese Association of Automation.

Interference Alignment Based on Subspace Tracking in MIMO Cognitive Networks with Multiple Primary Users

The interference alignment （IA） algorithm based on FDPM subspace tracking （FDPM-ST IA） is proposed for MIMO cognitive network （CRN） with multiple primary users in this paper. The feasibility conditions of FDPM-ST IA is also got. Futherly, IA scheme of secondary network and IA scheme of primary network are given respectively without assuming a priori knowledge of interference covariance matrices. Moreover, the paper analyses the computational complexity of FDPM-ST IA. Simulation results and theoretical calculations show that the proposed algorithm can achieve higher sum rate with lower computational complexity.

Global Avalanche Characteristics of Boolean Functions by Concatenation

In order to measure the correlation propeties of two Boolean functions,the global avalanche characteristics of Boolean functions constructed by concatenation are discussed,i.e.,f_1‖f_2and f_1‖f_2‖f_3‖f_4.Firstly,for the function f = f_1‖f_2,the cross-correlation function of f_1,f_2 in the special condition are studied.In this case,f,f_1,f_2 must be in desired form.By computing their sum-of-squares indicators,the crosscorrelation function between f_1,f_2 is obtained.Secondly,for the function g = f_1‖f_2‖f_3‖f_4,by analyzing the relation among their auto-correlation functions,their sum-of-squares indicators are investigated.Based on them,the sum-of-squares indicators of functions obtained by Canteaut et al.are investigated.The results show that the correlation property of g is good when the correlation properties of Boolean functions f_1,f_2,f_3,f_4 are good.

Cross-spectral root-min-norm algorithm for harmonics analysis in electric power system

To avoid drawbacks of classic discrete Fourier transform(DFT)method,modern spectral estimation theory was introduced into harmonics and inter-harmonics analysis in electric power system.Idea of the subspace-based root-min-norm algorithm was described,but it is susceptive to noises with unstable performance in different SNRs.So the modified root-min-norm algorithm based on cross-spectral estimation was proposed,utilizing cross-correlation matrix and independence of different Gaussian noise series.Lots of simulation experiments were carried out to test performance of the algorithm in different conditions,and its statistical characteristics was presented.Simulation results show that the modified algorithm can efficiently suppress influence of the noises,and has high frequency resolution,high precision and high stability,and it is much superior to the classic DFT method.

融合全局语义的CLIP-GPT图像描述模型

图像描述是指为图像自动生成与其内容相符的语言描述。桥接计算机视觉和自然语言处理两个领域的预训练模型构建图像描述模型时,跨模态语义一致性是共享子空间嵌入的核心问题。本文将图像拆分成若干片作为视觉语义单元与语言特征进行自由的跨模态关联,突破了有限视觉特征分类的限制;联合运用掩码学习和图文特征匹配两个损失函数,挑选高难度负样本训练跨模态跳接网络提取一致性全局语义,提高了子空间邻域内高相似度图文特征点匹配的准确度。在MS COCO和Flickr30k两个数据集上的实验结果表明,与同样采用CLIP+GPT生成图像描述的模型及其他主流模型相比,性能均有提升,证明了所提出模型的有效性。

Rank-One Cross Commutators on Backward Shift Invariant Subspaces on the Bidisk

For a backward shift invariant subspace N in H^2（Г^2）, the operators Sz and Sw on N are defined by Sz = PNTz｜N and Sw, = PNTw｜N, where PN is the orthogonal projection from L^2（Г^2） onto N. We give a characterization of N satisfying rank [Sz, Sw^＊] = 1.

基于子空间学习的快速自适应局部比值和判别分析

降维是处理高维数据的一项关键技术,其中线性判别分析及其变体算法均为有效的监督算法。然而大多数判别分析算法存在以下缺点:a)无法选择更具判别性的特征;b)忽略原始空间中噪声和冗余特征的干扰;c)更新邻接图的计算复杂度高。为了克服以上缺点,提出了基于子空间学习的快速自适应局部比值和判别分析算法。首先,提出了统一比值和准则及子空间学习的模型,以在子空间中探索数据的潜在结构,选择出更具判别信息的特征,避免受原始空间中噪声的影响;其次,采用基于锚点的策略构造邻接图来表征数据的局部结构,加速邻接图学习;然后,引入香农熵正则化,以避免平凡解;最后,在多个数据集上进行了对比实验,验证了算法的有效性。

双模态跨语料库语音情感识别

语音情感识别(SER)在双模态的跨数据库语音情感识别研究较少,跨数据库情感识别过度减少数据集之间差异的同时,会忽视情感判别能力的特征的问题。YouTube数据集为源数据,互动情感二元动作捕捉数据库(IEMOCAP)为目标数据。在源数据和目标数据中,Opensmile工具箱用来提取语音特征,将提取的语音特征输入到CNN和双向长短期记忆网络(BLSTM),来提取更高层次的特征,文本模态为语音信号的翻译稿。首先双向编码器表示转换器(Bert)把文本信息向量化,BLSTM提取文本特征,然后设计模态不变损失来形成2种模态的公共表示空间。为了解决跨语料库的SER问题,通过联合优化线性判别分析(LDA)、最大平均差异(MMD)、图嵌入(GE)和标签回归(LSR),学习源数据和目标数据的公共子空间。为了保留情绪辨别特征,情感判别损失与MMD+GE+LDA+LSR相结合。SVM分类器作为迁移公共子空间的最终情感分类,IEMOCAP上的实验结果表明,此方法优于其他先进的跨语料库和双模态SER.

Theoretical Study on One- and Two-Photon Absorption Properties of PPV Derivative with Electron-Donor Phenylanthracene as Pendent Group

Based on the equilibrium structures from quantum mechanics AM1 method, employing INDO/CI method and the sum-over-state （SOS） formula, the one-photon absorption （OPA） and two-photon absorption （TPA） properties as well as the second hyperpolarizabilities were discussed in detail for a kind of PPV derivative poly 2-（9-phenylanthracen-10-yl）-1,4-phenylenevinylene （P1）. The results indicate that the two-photon cross section （3） increases with the increasing of the number of repeating segment （n）, however only a slight increase corresponds to the increasing of molecular weight when the repeating unit number arrives at a certain number. From this point of view, the TPA cross section values of P1 were extrapolated through the linear fit of 3 value vs. 1/n. The δ value of P1 is as high as 181 GM （1 GM=10^-50 cm^4·s·photon^-1）. Concerning the influence of pendent group and extension of π-conjugation on 3 values, another four oligomers P2, PPV, P1-1, and P2-1 were investigated for comparison. The calculation results reveal that the position and property of pendent group have great influence on the OPA and TPA properties. A most crucial role for relatively larger 3 value was played by the extension of π-conjugation.

一种基于内容相关性的跨媒体检索方法

针对传统基于内容的多媒体检索对单一模态的限制,提出一种新的跨媒体检索方法.分析了不同模态的内容特征之间在统计意义上的典型相关性,并通过子空间映射解决了特征向量的异构性问题,同时结合相关反馈中的先验知识,修正不同模态多媒体数据集在子空间中的拓扑结构,实现跨媒体相关性的准确度量.实验以图像和音频数据为例验证了基于相关性学习的跨媒体检索方法的有效性.

基于多子空间直和特征融合的人脸识别算法

多个子空间直和能保证多个子空间数据融合时多个子空间得到的特征向量相互两两正交,融合数据采用该特征表示时冗余最小,更有利于分类识别。本文基于多子空间直和进行特征融合,提出了一种新的人脸识别算法。通过2DPCA算法,首先分别计算所有训练样本归一化后正脸、左侧脸及右侧脸图像的协方差矩阵的各P个最大特征值对应的P个相互正交的特征向量,然后通过选取3个子空间的部分满足直和条件的特征向量组成各自的特征空间(投影空间),再将样本正脸、左侧脸及右侧脸图像分别向各自特征空间投影得到3个特征矩阵,最后将此3个特征矩阵融合为该样本的特征矩阵用于最近邻分类器进行分类识别。最终通过本文3组实验数据的对比说明了该算法能减少计算量并且提高了识别率。

快速归一化互相关算法及DSP优化实现

归一化互相关匹配算法具有较好的鲁棒性和很高的匹配精度,但是计算量很大,不能满足实时性要求较高的系统。为此本文提出了一种快速的归一化互相关算法,通过和表法来减少分母的计算量,通过多模板方案来减少分子的计算时间,并对算法和结构进行优化,移植到DSP平台中,充分发挥了ADSP TS201的优良性能,实验结果表明经优化后算法运行速度得到明显提高但匹配精度不变。

一种用于电力系统谐波与间谐波分析的超分辨率算法

文中总结了经典DFT方法的缺陷,将现代谱估计理论引入电力系统谐波、间谐波分析领域,详细讨论了基于子空间分解的Min-Norm算法原理,针对其估计性能受噪声影响大的缺点,提出了基于互相关的互谱Min-Norm算法,利用白噪声序列的独立性及互相关运算有效地抑制了噪声对算法的影响,并将其用于电力系统谐波、间谐波分析。仿真结果表明该算法不仅具有超分辨率特性,而且可在低信噪比环境中获得高分辨率、高精度、高稳定性的谐波、间谐波分析结果,具有经典DFT方法无可比拟的优越性,可推广应用于电力系统谐波、间谐波测量领域。

模板图像匹配中互相关的一种快速算法

基于归一化互相关系数的算法在模板匹配和特征跟踪中运用十分广泛,但缺点是其计算量很大.为此提出了一种在空间域利用盒形基简化互相关的快速算法,在不修改归一化互相关匹配原理的前提下,用原模板图像在一组正交盒形基张成的子空间上的投影取代原图像来进行互相关计算,以降低图像精度来缩减计算复杂度.实验说明,当搜索窗口大小较小时,此快速算法计算量明显小于传统的频域快速归一化互相关算法.

基于抽样技术的ITS数据存档策略

介绍了ITS数据存储的最佳抽样方法,即对原始数据进行质量控制之后,运用误差平方和法(SSE)或互验法(CV)确定出原始ITS数据中最具有代表性的样本,从而降低所需的数据存储空间。通过开发相应的数据处理软件,针对美国得克萨斯州圣安东尼奥TransGuide交通管理中心的数据进行了测试,结果表明:对数据进行处理以后,仅需存储十分之一的原始数据,所得到的最佳样本则包含了最多的原始数据信息,且该样本数据能够满足潜在的不同交通需求。

基于互相关的印花织物疵点检测

为实现印花织物中疵点的自动检测,以互相关理论为基础,结合图像处理技术,以MatLab7.0构建了一套印花织物疵点自动检测系统。在疵点检测过程中,提出以加和表理论为基础实现互相关系数的快速计算。通过对软件模拟的印花花纹疵点的识别,说明这个系统能够实现印花过程中常见的花纹偏移、颜色色差等疵点的自动检测。实际印花织物疵点的检测实验表明,所提出的算法具有有效性、鲁棒性等优点。通过比较不同子窗口大小的检测结果,选定25像素×25像素作为最终检测系统中子窗口的大小。

跨媒体语义共享子空间学习研究进展

随着信息技术的飞速发展,现实世界中涌现出大量的跨媒体数据.所谓跨媒体数据是指那些表达的内容相似,但以不同模态、不同来源、不同背景等形式出现的数据.比如,一张描述花豹的网页通常采用共生的图片和文本等不同的模态刻画花豹的外形和习性.这些跨媒体数据通常呈现出底层特征异构、高层语义相关的特性.传统的单媒体学习方法已无法适应跨媒体数据呈现出的特征异构性.因此,跨媒体学习相关理论与方法的研究是当前数字媒体分析领域的热点研究课题之一.该文主要介绍了跨媒体学习的研究背景和应用价值,概括介绍了各类跨媒体学习相关方法的数学原理和基本特性,并重点介绍了跨媒体共享子空间学习的研究进展,比较了基于投影、矩阵分解、任务和度量等四大类子空间学习方法的优缺点,分析了未来的发展方向.

基于特征子空间直和的跨年龄人脸识别方法

针对跨年龄人脸识别任务,在同时进行人脸身份识别和年龄分类这两个任务的多任务卷积神经网络的基础上加入直和模块,提出了一种基于特征子空间直和的多任务卷积神经网络(FSDS-CNN)。该网络利用2个并行子网分别从深度特征中提取出身份相关特征和年龄相关特征,并对这2个相关特征所对应的特征子空间施加直和约束,使得身份相关特征与年龄相关特征尽可能无关。通过多损失的联合监督学习,该网络可以获得随年龄变化鲁棒的年龄无关人脸身份特征。分别在Morph Album 2、CACD-VS和Cross-Age LFW数据集上进行实验,其中在CACD-VS数据集中,所提方法的AUC最优值为99.7%;在Cross-Age LFW数据集中,所提方法在等错误率(EER)和错误匹配率(FMR)为0.1时的错误非匹配率(FNMR)上分别取得了最优值10.1%和10.2%。同时在3个数据集上的实验均进行了消融对比实验以验证直和模块的有效性。实验结果表明,身份特征与年龄特征的相关性被FSDS-CNN中的直和模块有效地降低,从而有效提升了模型跨年龄人脸识别的性能。