基于GPU平台的联合迭代重构算法加速

电子断层三维重构技术(Electron Tomography,ET)是在纳米尺度下研究不具有全同性的细胞或大分子三维结构的重要方法。迭代重构法是ET中重构效果最好的方法,但是其性能较差,重构大尺寸图像时需要数天的时间甚至更长,使其应用受到限制。迭代重构法中经典的方法是代数重构法(Algebraic Reconstruction Technique,ART)和联合迭代重构法(Simultaneous Iterative Reconstruction Technique,SIRT),SIRT算法总是收敛的并且比ART重构的效果更好。利用CUDA语言设计和实现了基于Tesla C1060GPU平台上的并行SIRT重构算法,并利用存储器合并访问、常量存储器、共享存储器等优化技术对并行算法进行优化,优化后的SIRT并行算法在Tesla C1060GPU平台上的最大加速比是Intel i7 920CPU上的串行算法的47倍,并且重构的质量没有任何下降。     

The iterative flux-profile method for remote sensing applications

Iterative flux-profile methods allow calculation of sensible heat and momentum fluxes using temperature gradients and wind speed data obtained in the lower atmosphere. Flux-profile methods are also applied to the problem of regional evapotranspiration evaluation by the use of satellite imagery. The large number of pixels in these images and the long image time series that have become available make it necessary to develop reliable, fast and accurate algorithms to process the data. In this paper an analysis is made of the equations underlying the flux-profile method with the practical objective to improve the existing procedures for application with large datasets. After a review of the iterative methods to arrive at the solution, simple changes are suggested to improve estimates of the initial value and to speed up function evaluation. As an illustration of the suggested changes, a NOAA Advanced Very High Resolution Radiometer (AVHRR) case study from western Turkey is discussed. It is shown that the improved iterative algorithm is in this case a factor of 4-5 faster than the standard method. The results are also compared with those obtained by a recent analytical approximation.     

面向超大数据集的SVM近似训练算法

标准SVM学习算法运行所需的时间和空间复杂度分别为O(l~3)和O(l~2),l为训练样本的数量,因此不适用于对超大数据集进行训练.提出一种基于近似解的SVM训练算法:Approximate Vector Machine(AVM).AVM采用增量学习的策略来寻找近似最优分类超平面,并且在迭代过程中采用热启动及抽样技巧来加快训练速度.理论分析表明,该算法的计算复杂度与训练样本的数量无关,因此具有良好的时间与空间扩展性.在超大数据集上的实验结果表明,该算法在极大提高训练速度的同时,仍然保持了原始分类器的泛化性能,并且训练完毕具有较少的支持向量,因此结果分类器具有更快的分类速度.     

Convergence of the iterative Hammerstein system identification algorithm

The convergence of the iterative identification algorithm for the Hammerstein system has been an open problem for a long time. In this paper, a detailed study is carried out and various convergence properties of the iterative algorithm are derived. It is shown that the iterative algorithm with normalization is convergent in general. Moreover, it is shown that convergence takes place in one step (two least squares iterations) for finite-impulse response Hammerstein models with i.i.d. inputs.     

Localized relaxation theory of circuits and its applications in electro-thermal analyses

In the high-performance IC design with increasing design complexity,it is a very important design content to efficiently analyze IC parameters.Thus,the electro-thermal (ET) analyses including power/ground (P/G) analysis and thermal analysis are hot topics in today’s IC research.Since ET analysis equation has a sparse,positive definite and strictly diagonally dominant coefficient-matrix,we prove that the ET analysis has the advantage of locality.Owing to this advantage,localized relaxation method is formally proposed,which has the same accuracy as the global relaxation done with the constraint of the same truncation error limitation.Based on the localized relaxation theory,an efficient and practical localized successive over-relaxation algorithm (LSOR2) is introduced and applied to solve the following three ET analysis problems.(1) Single-node statistical voltage analysis for over-IR-drop nodes in P/G networks;(2) single-node statistical temperature analysis for hot spots in 3D thermal analysis;(3) fast single open-defect analysis for P/G networks.A large amount of experimental data demonstrates that compared with the global successive over-relaxation (SOR) algorithm,LSOR2 can speed up 1-2 orders of magnitudes with the same accuracy in ET analyses.     

On the P-type and Newton-type ILC schemes for dynamic systems with non-affine-in-input factors

In this paper, P-type learning scheme and Newton-type learning scheme are proposed for quite general nonlinear dynamic systems with non-affine-in-input factors. Using the contraction mapping method, it is shown that both schemes can achieve asymptotic convergence along learning repetition horizon. In order to quantify and evaluate the learning performance, new indices—Q-factor and Q-order—are introduced in particular to evaluate the learning convergence speed. It is shown that the P-type iterative learning scheme has a linear convergence order with limited learning convergence speed under system uncertainties. On the other hand, if more of system information such as the input Jacobian is available, Newton-type iterative learning scheme, which is originated from numerical analysis, can greatly speed up the learning convergence speed. The effectiveness of the two learning control methods are demonstrated through a switched reluctance motor system.     

基于强化学习的无人坦克对战仿真研究

对标准的强化学习进行改进,通过引入动机层,来引入先验知识,加快学习速度。策略迭代选择上,通过采用“同策略”迭代的Sarsa学习算法,代替传统的“异策略”Q学习算法。提出了基于多动机引导的Sarsa学习（MMSarsa）算法,分别和Q学习算法、Sarsa学习算法在坦克对战仿真问题上进行了三种算法的对比实验。实验结果表明,基于多动机引导的Sarsa学习算法收敛速度快且学习效率高。     

基于结构简化的BP网络的多类形状识别

BP算法是目前应用极为广泛的神经网络算法,但它也存在一些不足。文中提出了一种使BP网络结构简化的改进的BP算法,它使得网络的速度得到了提高,减少了迭代次数。最后将传统的BP算法和改进的BP算法进行了比较,仿真结果表明改进的算法在学习次数和收敛速度上比传统的算法有很大的改善。     

安检中X光断层扫描加速迭代算法研究

随着安检要求的日益提高,计算机断层扫描的安检方式正成为X光安检中越来越重要的部分.为了保证行李扫描的检测效率,加快断层扫描图像的迭代计算速度,提出了一种采用小波变换域取代图像空间来测算图像小波系数的估值方法.该方法以交替最小化算法为框架,在逼近系数时,创建一个替代函数,随着迭代次数的增加,基于同一标准的小波树得以更新扩展,每层的细节系数均更新,小波树也依此扩展,以此达到加速迭代的效果.实验结果表明,该方法比原交替最小化算法最大能减少66.7％的迭代运算次数.     

有序子集最小二乘OS—LS图像重建迭代算法

为推导一种新的快速图像迭代重建方法，首先将有序子集(ordered subsets，OS)技术应用到最小二乘图像重建迭代算法(least square reconstruction，LS)；然后对仿真Phantom模型数据和实际医用正电子发射断层成像仪(PET)数据进行重建，并研究了在不同子集划分下的重建结果，同时分析比较了不同子集的选取对OS—LS重建罔像质量以及重建收敛速度的影响。重建结果表明，这种基于有序子集的最小二乘图像重建迭代算法(OS—LS)具有较高的重建图像质量和较短的计算时间，相对于传统LS算法的重建，OS—LS的收敛速度加速了约L倍(L为子集个数)．其重建图像质量也好于传统的滤波反投影(FBP)方法的重建．町应用在PET图像重建中。     

一种新算法在ECT图像重建中的应用研究

为了解决大残量、小残量或零残量问题和普通迭代方法收敛速度慢的问题,引入NL2SOL优化算法,并与GaussNewton法相结合,提出了一种多项式加速的正则化迭代电容层析成像算法.在介绍12电极电容层析成像系统基本原理的基础上,给出了该算法的数学模型,并利用谱分析对该算法的收敛性进行了证明.仿真实验验证了该算法满足收敛条件且重建图像误差小,在电容层析成像应用中具有可行性.     

Quarter-sweep iterative alternating decomposition explicit algorithm applied to diffusion equations

The aim of this article is to describe the formulation of the quarter-sweep iterative alternating decomposition explicit (QSIADE) method using the finite difference approach for solving one-dimensional diffusion equations. The concept of the QSIADE method is inspired via combination between the quarter-sweep iterative and the iterative alternating decomposition explicit (IADE) methods known as one of the technique in two-step iterative methods. The QSIADE method has been shown to be very fast as compared with the standard IADE method. Some numerical tests were included to support our statement.     

基于改进图像重建算法的工业CT系统

目前工业CT扫描只能得到某一断面上的二维信息,不能得到被检测物的整体描述,人们希望获得具有高真实感的三维图像;Smith和Grangeat分别提出了三维投影重建算法,H.Kudo在对两种算法分别进行研究的基础上,提出了Smith-Grangeat方法,同时继承了两种算法的优点,但存在图像明显失真和衰减问题;通过对Smith-Grangeat方法进行改进,使图像失真减小;仿真结果验证了算法的有效性和工程实用性.     

基于音频指纹的两步固定音频检索

提出了一种基于过零率和音频指纹的两步固定音频检索算法.在基于过零率直方图的初步检索中,采用直方图的迭代计算和动态的观测窗滑动步长来减少计算量并加快搜索速度,快速筛选出相似度较高的候选音频片段;接着基于降维Philips音频指纹对候选音频进行精检索,进一步提高检索精度.实验结果表明,该音频检索算法在保证较好的检索准确性基础上,大幅度提高了检索速度,且具有较好的鲁棒性.     

Sequential weighted least squares algorithm for PET image reconstruction

The weighted least squares (WLS) algorithm has proven useful for modern positron emission tomography (PET) scanners to approach reconstructions with non-Poisson precorrected measurement data. In this paper, we propose a new time recursive sequential WLS algorithm whose derivation uses the time-varying property of data acquisition of PET scanning. It ties close relationship with the time-varying Kalman filtering and can be extended appropriately to an iteration fashion as the absence of proper a priori initializations. The performance of sequential WLS is evaluated experimentally. The results show its fast convergence over both the multiplicative and coordinate-based iterative WLS methods. It also produces relative uniform estimate variances that makes it more suitable for routine applications.     

Registration of CAD mesh models with CT volumetric model of assembly of machine parts

X-ray CT (computed tomography) has experienced tremendous growth in industrial application in recent years, and acquiring information about the mechanical parts from CT data has been a great challenge for researchers. This paper presents a new method for the registration of a CT volumetric model of an assembly of parts with a CAD mesh model of a part of the assembly using ICP (iterative closest point) registration method. A few steps to extract feature points should be done as a preprocessing step of the volumetric model and the mesh model before applying the ICP registration algorithm, since the volumetric model and the mesh model are different in their data representation. This preprocessing step is important in order to unify the input of the ICP algorithm, and contributes to the robustness and the speed of the registration process.     

An efficient search direction for linear programming problems

In this paper, we present an auxiliary algorithm, in terms of the speed of obtaining the optimal solution, that is effective in helping the simplex method for commencing a better initial basic feasible solution. The idea of choosing a direction towards an optimal point presented in this paper is new and easily implemented. From our experiments, the algorithm will release a corner point of the feasible region within few iterative steps, independent of the starting point. The computational results show that after the auxiliary algorithm is adopted as phase I process, the simplex method consistently reduce the number of required iterations by about 40%.Scope and purposeRecent progress in the implementations of simplex and interior point methods as well as advances in computer hardware has extended the capability of linear programming with today's computing technology. It is well known that the solution times for the interior point method improve with problem size. But, experimental evidence suggests that interior point methods dominate simplex-based methods only in the solution of very large scale linear programs. If the problem size is medium, how to combine the best features of these two methods to produce an effective algorithm for solving linear programming problems is still an interesting problem. In this research we present a new effective ε-optimality search direction based on the interior point method to start an initial basic feasible solution near the optimal point for the simplex method.     

A (/spl mu/ + /spl lambda/) evolutionary algorithm for reconstruction of SPECT data

Algorithms used to reconstruct single photon emission computed tomography (SPECT) data are based on one of two principles: filtered back projection or iterative methods. In this paper, an evolution strategy (ES) was applied to reconstruct transaxial slices of SPECT data. Evolutionary algorithms are stochastic global search methods that have been used successfully for many kinds of optimization problems. The newly developed reconstruction algorithm consisting of /spl mu/ parents and /spl lambda/ children uses a random principle to readjust the voxel values, whereas other iterative reconstruction methods use the difference between measured and simulated projection data. The (/spl mu/ + /spl lambda/)-ES was validated against a test image, a heart, and a Jaszczak phantom. The resulting transaxial slices show an improvement in image quality, in comparison to both the filtered back projection method and a standard iterative reconstruction algorithm.     

一种通过反馈提高神经网络学习性能的新算法

为了有效提高前向神经网络的学习性能，需要从一个新的角度考虑神经网络的学习训练．基于此，提出了一种基于结果反馈的新算法——FBBP算法．将神经网络输入调整与通常的权值调整的反向传播算法结合起来，通过调整权值和输入矢量值的双重作用来最小化神经网络的误差函数．并通过几个函数逼近和模式分类问题的实例仿真，将FBBP算法与加动量项BP算法、最新的一种加快收敛的权值更新的算法进行了比较，来验证所提出的算法的有效性．实验结果表明，所提出的算法具有训练速度快和泛化能力高的双重优点，是一种非常有效的学习方法．     

Combining GA and iterative searching DOA estimation for CDMA signals

This paper deals with direction-of-arrival (DOA) estimation based on iterative searching technique for code-division multiple access signals. It has been shown that the iterative searching technique is more likely to converge to a local maximum, causing errors in DOA estimation. In conjunction with a genetic algorithm for selecting initial search angle, we present an efficient approach to achieve the advantages of iterative DOA estimation with fast convergence and more accuracy estimate over existing conventional spectral searching methods. Finally, several computer simulation examples are provided for illustration and comparison.