基于卷积神经网络的图像数据增强算法

提升卷积神经网络的泛化能力和降低过拟合的风险是深度卷积神经网络的研究重点。遮挡是影响卷积神经网络泛化能力的关键因素之一,通常希望经过复杂训练得到的模型能够对遮挡图像有良好的泛化性。为了降低模型过拟合的风险和提升模型对随机遮挡图像识别的鲁棒性,提出了激活区域处理算法,在训练过程中对某一卷积层的最大激活特征图进行处理后对输入图像进行遮挡,然后将被遮挡的新图像作为网络的新输入并继续训练模型。实验结果表明,提出的算法能够提高多种卷积神经网络模型在不同数据集上的分类性能,并且训练好的模型对随机遮挡图像的识别具有非常好的鲁棒性。     

实时目标检测算法YOLO的批再规范化处理

针对实时目标检测YOLO(You Look Only Once)算法中存在的检测精度低和网络模型训练速度慢等问题,结合批再规范化算法处理小批样本以及非独立同分布数据的优势,提出了在YOLO网络结构加入批再规范化处理的改进算法。该YOLO改进算法把卷积层经过卷积运算产生的特征图看作一个个神经元,然后对这些神经元进行规范化处理。同时,在网络结构中移除了Dropout,并增大了网络训练的学习率。实验结果表明,该改进算法相对于原YOLO算法具有更高的检测精度、更快的实时检测速度以及通过适当设置批样本大小可使网络模型在训练时间和硬件设备方面成本有一定的降低。     

基于ROI-KNN卷积神经网络的面部表情识别

深度神经网络已经被证明在图像、语音、文本领域具有挖掘数据深层潜在的分布式表达特征的能力. 通过在多个面部情感数据集上训练深度卷积神经网络和深度稀疏校正神经网络两种深度学习模型, 对深度神经网络在面部情感分类领域的应用作了对比评估. 进而, 引入了面部结构先验知识, 结合感兴趣区域(Region of interest, ROI)和K最近邻算法(K-nearest neighbors, KNN), 提出一种快速、简易的针对面部表情分类的深度学习训练改进方案——ROI-KNN, 该训练方案降低了由于面部表情训练数据过少而导致深度神经网络模型泛化能力不佳的问题, 提高了深度学习在面部表情分类中的鲁棒性, 同时, 显著地降低了测试错误率.     

基于面向对象自适应粒子群算法的神经网络训练*

针对传统的神经网络训练算法收敛速度慢和泛化性能低的缺陷,提出一种新的基于面向对象的自适应粒子群优化算法(OAPSO)用于神经网络的训练。该算法通过改进PSO的编码方式和自适应搜索策略以提高网络的训练速度与泛化性能,并结合Iris和Ionosphere分类数据集进行测试。实验结果表明：基于OAPSO算法训练的神经网络在分类准确率上明显优于BP算法及标准PSO算法,极大地提高了网络泛化能力和优化效果,具有快速全局收敛的性能。     

基于GR-CNN算法的网络入侵检测模型设计与实现

针对现有网络入侵检测系统对网络行为检测准确率较低、实时性较差、泛化性能较低的问题,利用深度学习具有良好分类性能及强泛化能力等优点,设计基于增益率算法和卷积神经网络算法的网络入侵检测模型。采用增益率筛选数据集数据特征,在保证入侵检测准确率的同时,缩短卷积神经网络训练时间。实验结果表明,该模型相比其他基于机器学习的入侵检测模型具有较高的准确率和较强的泛化能力,同时优化卷积神经网络训练方式,保证准确率的同时使神经网络训练时间减少了77%。     

单纯形微粒群组合算法优化神经网络及在数据挖掘中的应用

神经网络作为数据挖掘的重要工具,如果提高神经网络训练过程的达优率与泛化能力,就能够改善其在分类预测任务中的预测准确度和噪声承受力。 该文提出以复合适应度代替传统的均方误差作为神经网络的训练指标,从而改善神经网络泛化能力;并将用于训练神经网络微粒群算法通过单纯形算法改进,从而提高了全局搜索能力与达优率。最终产生一种基于复合适应度的单纯微粒群神经网络训练算法模型。 将这样的模型应用于数据挖掘的预测任务中,实验结果表明,预测准确度和噪声承受力均有所提高。     

基于Adaboost算法的交通事件检测

提出一种Adaboost BP神经网络的交通事件检测方法:以上下游的流量和占有率作为特征,用BP神经网络作为分类器进行交通事件的自动分类与检测.在BP神经网络的训练过程中,提出一种新的训练算法,提高了神经网络的分类能力.为了进一步提高神经网络的泛化能力,采用改进的Adaboost算法,进行网络集成.运用Matlab进行仿真分析.结果表明所提出的交通事件检测算法具有较好的检测性能.     

基于模型的深度学习通信信号鲁棒识别算法

深度学习现在是计算机视觉和自然语言处理的热门话题。在许多应用中,深度神经网络（DNN）的性能都优于 传统的方法,并且已经成功应用于调制分类和无线电信号表示等任务的学习。近几年研究发现深度神经网络极易受到对抗性 攻击,对“对抗性示例”缺乏鲁棒性。笔者就神经网络的通信信号识别算法的鲁棒性问题,将经过PGD攻击的数据看作基于模 型的数据,将该数据输入神经网络,使得信号识别分类结果错误;然后借助基于模型的防御算法,即鲁棒训练算法和对抗训练 算法,进行训练后实验结果表明,两种方法都具有较好的防御效果。     

基于本地蒸馏联邦学习的鲁棒语音识别技术

当联邦学习（FL）算法应用于鲁棒语音识别任务时,为解决训练数据非独立同分布（Non-IID）与客户端模型缺乏个性化问题,提出基于个性化本地蒸馏的联邦学习（PLD-FLD）算法。客户端通过上行链路上传本地Logits并在中心服务器聚合后下传参数,当边缘端模型测试性能优于本地模型时,利用下载链路接收中心服务器参数,确保了本地模型的个性化与泛化性,同时将模型参数与全局Logits通过下行链路下传至客户端,实现本地蒸馏学习,解决了训练数据的Non-IID问题。在AISHELL与PERSONAL数据集上的实验结果表明,PLD-FLD算法能在模型性能与通信成本之间取得较好的平衡,面向军事装备控制任务的语音识别准确率高达91%,相比于分布式训练的FL和FLD算法具有更快的收敛速度和更强的鲁棒性。     

受损图片分类的双并行交叉降噪卷积神经网络

应对大量受干扰图像的分类问题,提出了一种双并行交叉降噪卷积模型,该模型由两部分并行交叉网络结构组成,分别对应改进的自编码方式和并行交叉卷积神经网络,同时在该模型训练的过程中,使用批量正则化和改进激活函数的方法。经实验验证,与同类模型相比,该模型首先具有降噪能力强、鲁棒性好、泛化能力强和准确率高的特点,其次避免过拟合,加快收敛速度。在图片不同程度受损的情况下,它也可较好地完成图像目标识别分类任务。     

A fast method for choosing the optimum extrapolation parameter for an e.a.d.i. scheme

This paper describes a fast method for choosing the optimum extrapolation parameter ω in an Extrapolated Alternating Direction Implicit (E.A.D.I.) Scheme, for the numerical solution of elliptic partial differential equations. An analysis is presented to show that in the case that the acceleration parameter r is varied, a set of values for the extrapolation parameter ω could be defined. The optimum value of ω is easily obtained from this set. A numerical example demonstrates the applicability of the proposed method.     

基于CompactLogix和现场总线的过程控制系统

对过程控制中的集散控制系统DCS与现场总线控制系统FCS进行了比较，详细阐述了基于Rockwell Automation的现场总线技术DeviceNet的网络控制方法，并通过PID算法编程与监控软件开发，实现了基于CompactLogix控制器与DeviceNet现场总线的过程控制系统。     

COMPUTER-MOVIE的设计与实现技术

COMPUTER MOVIE是一个适用于PC系列微机的计算机教学动画演示系统,可应用于计算机原理、计算机系统结构、微型计算机等课程的教学。本文论述了该系统的构成以及在该系统中采用的一系列作图技术,包括:实现结构化图形实体移动的方法、“面向图形实体/事件驱动/自然遮挡”动画作图法、菜单、背景色调整、相对坐标和大小的采用、多条同时动态延伸线的画法等。这些技术适用于各种教学动画软件的设计与实现。     

基于嵌入式系统的网络化自动抄表系统

设计基于嵌入式系统的网络化自动抄表系统,并着重阐述系统的硬件、软件构成和设计特点。该系统可用于建筑中电、水、气、热等仪表数据的采集和住户耗能设备的控制。自动抄表有利于能量管理部门实现能源的监控管理,提高能源使用效益,对于中国的可持续发展具有非常重要的意义。     

KANTBP: A program for computing energy levels, reaction matrix and radial wave functions in the coupled-channel hyperspherical adiabatic approach

A FORTRAN 77 program is presented which calculates energy values, reaction matrix and corresponding radial wave functions in a coupled-channel approximation of the hyperspherical adiabatic approach. In this approach, a multi-dimensional Schrödinger equation is reduced to a system of the coupled second-order ordinary differential equations on the finite interval with homogeneous boundary conditions of the third type. The resulting system of radial equations which contains the potential matrix elements and first-derivative coupling terms is solved using high-order accuracy approximations of the finite-element method. As a test desk, the program is applied to the calculation of the energy values and reaction matrix for an exactly solvable 2D-model of three identical particles on a line with pair zero-range potentials.

Program summary

Program title: KANTBPCatalogue identifier: ADZH_v1_0Program summary URL:http://cpc.cs.qub.ac.uk/summaries/ADZH_v1_0.htmlProgram obtainable from: CPC Program Library, Queen's University, Belfast, N. IrelandLicensing provisions: Standard CPC licence, http://cpc.cs.qub.ac.uk/licence/licence.htmlNo. of lines in distributed program, including test data, etc.: 4224No. of bytes in distributed program, including test data, etc.: 31 232Distribution format: tar.gzProgramming language: FORTRAN 77Computer: Intel Xeon EM64T, Alpha 21264A, AMD Athlon MP, Pentium IV Xeon, Opteron 248, Intel Pentium IVOperating system: OC Linux, Unix AIX 5.3, SunOS 5.8, Solaris, Windows XPRAM: depends on (a) the number of differential equations; (b) the number and order of finite-elements; (c) the number of hyperradial points; and (d) the number of eigensolutions required. Test run requires 30 MBClassification: 2.1, 2.4External routines: GAULEG and GAUSSJ [W.H. Press, B.F. Flanery, S.A. Teukolsky, W.T. Vetterley, Numerical Recipes: The Art of Scientific Computing, Cambridge University Press, Cambridge, 1986]Nature of problem: In the hyperspherical adiabatic approach [J. Macek, J. Phys. B 1 (1968) 831-843; U. Fano, Rep. Progr. Phys. 46 (1983) 97-165; C.D. Lin, Adv. Atom. Mol. Phys. 22 (1986) 77-142], a multi-dimensional Schrödinger equation for a two-electron system [A.G. Abrashkevich, D.G. Abrashkevich, M. Shapiro, Comput. Phys. Comm. 90 (1995) 311-339] or a hydrogen atom in magnetic field [M.G. Dimova, M.S. Kaschiev, S.I. Vinitsky, J. Phys. B 38 (2005) 2337-2352] is reduced by separating the radial coordinate ρ from the angular variables to a system of second-order ordinary differential equations which contain potential matrix elements and first-derivative coupling terms. The purpose of this paper is to present the finite-element method procedure based on the use of high-order accuracy approximations for calculating approximate eigensolutions for such systems of coupled differential equations.Solution method: The boundary problems for coupled differential equations are solved by the finite-element method using high-order accuracy approximations [A.G. Abrashkevich, D.G. Abrashkevich, M.S. Kaschiev, I.V. Puzynin, Comput. Phys. Comm. 85 (1995) 40-64]. The generalized algebraic eigenvalue problem AF=EBF with respect to pair unknowns (E,F) arising after the replacement of the differential problem by the finite-element approximation is solved by the subspace iteration method using the SSPACE program [K.J. Bathe, Finite Element Procedures in Engineering Analysis, Englewood Cliffs, Prentice-Hall, New York, 1982]. The generalized algebraic eigenvalue problem (A−EB)F=λDF with respect to pair unknowns (λ,F) arising after the corresponding replacement of the scattering boundary problem in open channels at fixed energy value, E, is solved by the LDLT factorization of symmetric matrix and back-substitution methods using the DECOMP and REDBAK programs, respectively [K.J. Bathe, Finite Element Procedures in Engineering Analysis, Englewood Cliffs, Prentice-Hall, New York, 1982]. As a test desk, the program is applied to the calculation of the energy values and reaction matrix for an exactly solvable 2D-model of three identical particles on a line with pair zero-range potentials described in [Yu. A. Kuperin, P.B. Kurasov, Yu.B. Melnikov, S.P. Merkuriev, Ann. Phys. 205 (1991) 330-361; O. Chuluunbaatar, A.A. Gusev, S.Y. Larsen, S.I. Vinitsky, J. Phys. A 35 (2002) L513-L525; N.P. Mehta, J.R. Shepard, Phys. Rev. A 72 (2005) 032728-1-11; O. Chuluunbaatar, A.A. Gusev, M.S. Kaschiev, V.A. Kaschieva, A. Amaya-Tapia, S.Y. Larsen, S.I. Vinitsky, J. Phys. B 39 (2006) 243-269]. For this benchmark model the needed analytical expressions for the potential matrix elements and first-derivative coupling terms, their asymptotics and asymptotics of radial solutions of the boundary problems for coupled differential equations have been produced with help of a MAPLE computer algebra system.Restrictions: The computer memory requirements depend on:

•

(a) the number of differential equations;

•

(b) the number and order of finite-elements;

•

(c) the total number of hyperradial points; and

•

(d) the number of eigensolutions required.

Restrictions due to dimension sizes may be easily alleviated by altering PARAMETER statements (see Long Write-Up and listing for details). The user must also supply subroutine POTCAL for evaluating potential matrix elements. The user should supply subroutines ASYMEV (when solving the eigenvalue problem) or ASYMSC (when solving the scattering problem) that evaluate the asymptotics of the radial wave functions at the right boundary point in case of a boundary condition of the third type, respectively.Running time: The running time depends critically upon:

•

(a) the number of differential equations;

•

(b) the number and order of finite-elements;

•

(c) the total number of hyperradial points on interval [0,ρmax]; and

•

(d) the number of eigensolutions required.

The test run which accompanies this paper took 28.48 s without calculation of matrix potentials on the Intel Pentium IV 2.4 GHz.     

离散系统的一种从模型取状态鲁棒MRAC方案

本文对具有可测扰动的离散系统提出了一种从模型取状态的鲁棒模型参考自适应控制ＭＲＡＣ（模型参考自适应控制）新算法。     

Circular Separability of Polygons

Two planar sets are circularly separable if there exists a circle enclosing one of the sets and whose open interior disk does not intersect the other set. This paper studies two problems related to circular separability. A linear-time algorithm is proposed to decide if two polygons are circularly separable. The algorithm outputs the smallest separating circle. The second problem asks for the largest circle included in a preprocessed, convex polygon, under some point and/ or line constraints. The resulting circle must contain the query points and it must lie in the halfplanes delimited by the query lines. Received October 25, 1998; revised April 21, 1999.     

利用Remoting技术实现分布式系统的开发

运用.NET Remoting进行分布式应用程序开发不仅提高开发效率,而且具有开发的简单性和系统的灵活性,特别是在局域网内运用TCP进行数据传输可以提高数据传输效率.本文介绍了.NET Remoting的体系结构,提出基于.NET Remoting的分布式应用程序的开发方法,并详细论述了在Visual Studio .NET中实现该方法的过程.     

Classification and Uncertainty Visualization of Dendritic Spines from Optical Microscopy Imaging

Neuronal dendrites and their spines affect the connectivity of neural networks, and play a significant role in many neurological conditions. Neuronal function is observed to be closely correlated with the appearance, disappearance and morphology of the spines. Automatic 3‐D reconstruction of neurons from light microscopy images, followed by the identification, classification and visualization of dendritic spines is therefore essential for studying neuronal physiology and biophysical properties. In this paper, we present a method to reconstruct dendrites using a surface representation of the dendrite. The 1‐D skeleton of the dendritic surface is then extracted by a medial geodesic function that is robust and topologically correct. This is followed by a Bayesian identification and classification of the spines. The dendrite and spines are visualized in a manner that displays the spines' types and the inherent uncertainty in identification and classification. We also describe a user study conducted to validate the accuracy of the classification and the efficacy of the visualization.