基于PSO和改进神经网络的图像滤波方法的研究

提出了一种基于改进BP神经网络和粒子群优化算法（PSO）的图像滤波方法。该方法利用对数最小均方误差函数（LNLS）代替BP神经网络传统的最小均方误差函数（LMS）,用来减小图像噪声对神经网络精度的影响;并将改进后的BP神经网络利用PSO算法优化,从而避免神经网络陷入局部极小值点,进一步提高神经网络滤波能力。实验结果表明,与传统滤波方法相比,该方法不仅能有效地滤除图像中的高斯噪声而且能很好地保护图像细节。     

基于改进BP神经网络和粒子群优化算法的图像滤波方法的研究

提出了一种基于改进BP神经网络和粒子群优化算法(PSO)的图像滤波方法.该方法利用双曲正切形式的误差函数代替BP神经网络传统的最小均方误差函数(LMS),并将改进后的BP神经网络利用PSO算法优化,用来减小图像噪声对神经网络精度的影响以及避免神经网络陷入局部极小值点,从而提高神经网络去噪能力.实验结果表明,与传统滤波方法相比,该方法不仅能有效地滤除图像中的高斯噪声而且能很好地保护图像细节.     

基本电路

0600695 用进化规划思想优化径向基函数神经网络结构的均衡器[刊,中]／田俊霞／／北京理工大学学报．-2005,25 (9)．-819-822(L) 针对训练径向基函数(RBF)神经网络均衡器的随机梯度算法(SG)中,神经网络的结构是指定的并且所用训练样本较长的问题,引入进化规划思想,用进化规划方法确定径向基函数神经网络的结构,用基于最小     

多准则图象重建的神经网络模型及实现

本文提出一种采用Ｈｏｐｆｉｅｌｅ神经网络（Ｈｏｐｆｉｅｌｄ Ｎｅｉｒａｌ Ｎｅｔｗｏｒｋ简称ＨＮＮ）优化的图象重建算法。将图象重建问题转化为ＨＮＮ优化问题,取重建图象的峰值函数最小以及原始投影与再投影之间的误差平方和最小作为图象重建的优化目标,作为能量函数构造连续型ＨＮＮ模型,由ＨＮＮ能量函数极小化可得到重建问题的优化解。这种方法具有简单、计算量小、收敛快、便于并行计算等特点。对照ＡＲＴ算法,用计算机模拟产生的无噪声投影数据检验新算法,验证了新算法的优越性。     

基于RBF神经网络的压电执行器迟滞建模

针对压电陶瓷执行器的迟滞非线性特性问题,该文提出了一种最小二乘法与径向神经网络相结合的建模方法。首先,通过搭建压电执行器位移测试系统,得到执行器输出位移与输入电压的对应曲线关系,然后用最小二乘法对该曲线进行多项式拟合,得到压电执行器的迟滞数学模型,在此基础上再用径向基函数神经网络方法对该模型进行优化。最后对建立的模型进行分析发现,用最小二乘法拟合的多项式数学模型,其最大误差Emax=0.244 7 μm,标准方差δ=0.059 02 μm,而利用径向基函数(RBF)神经网络优化建模后,Emax=0.079 89 μm,δ=0.016 04 μm。实验证明该模型有较高的准确性, 该文为压电执行器迟滞建模提供了一种新的方法。     

HHGA-RBFNN在DGS设计中的应用

将递阶遗传算法和最小二乘法相结合, 构成一种混合递阶遗传算法,用于优化径向基函数神经网络的结构和参数,通过MATLAB仿真实现该算法,并将该神经网络运用于微带缺陷接地结构的神经网络建模之中,实现对微带缺陷接地结构传输系数的仿真.实验结果表明,应用混合递阶遗传算法优化的径向基函数神经网络,具有很高的拟合精度和很好的泛化能力,可以与时域有限差分分析方法结合,辅助微带缺陷接地结构的设计.     

深度学习算法中不同优化器的性能分析

针对神经网络学习使其找到使得损失函数的值最小的参数,寻找最优参数的问题,在TensorFlow的深度学习算法的框架下,利用Kersa搭建神经网络,运用import、train&test、sequential、compile、fi t和summary六步法搭建出相关神经网络,对不同优化器的性能进行了分析比较,包括SGD、...     

基于优化计算的神经网络能量函数分析

Hopfield神经网络是求解组合优化问题的一种有效方法。将所求问题转化为能量函数是神经网络求解组合优化问题的难点。针对此问题给出了能量函数设计的一般方法和步骤,证明了网络稳定的充分必要条件是网络的解有界。讨论了网络的平衡态与能量函数的极小点的关系,进一步完善了能量函数的方法。作为应用,严格证明了Hopfield神经网络的收敛性,所获结果不仅推广了一些已有的结论,而且为该网络的应用提供了一定的理论基础。     

约束非线性规划问题的L1精确罚函数神经网络方法

 优化计算是神经网络的一个重要应用领域.针对已有神经网络求解约束非线性规划问题时,不能兼顾网络规模、计算效率、精确性的问题,本文提出了一种基于精确罚函数的约束非线性规划问题的神经网络计算方法.将约束非线性规划问题的一种L₁精确罚函数作为神经网络的能量函数,利用该能量函数的最速下降原理构造了神经网络的动力学方程并给出了其稳定收敛性说明.理论分析及算例仿真表明,所提出神经网络动力学方程能够全局、精确收敛于原规划问题的一个局部最优解.特别是,该神经网络动力学方程易于映射为动态电路,是一种工程优化问题的实时计算方法.     

一种新的量子神经网络训练算法

量子神经网络是一种借鉴量子理论中的态叠加思想而设计的单隐层前馈神经网络,其主要用于数据分类。由于采用多层激励函数神经元,并且在量子间隔训练中采用了新的目标函数,即同类输入数据的隐层节点输出方差最小,从而使量子神经网络具备了发掘不同类别数据间模糊性的能力。但由于训练时对量子神经网络权值和量子间隔使用了不同的目标函数,使迭代过程中两者不可避免的会出现相互冲突,从而导致训练迭代次数的增加和网络性能的下降。本文借鉴约束优化理论,在两个目标函数的梯度下降求解中引入了惩罚函数,提出了一种新的量子神经网络训练算法,消除了两个目标函数间的冲突。实验结果表明,本文提出的训练算法可以显著提升训练的速度和网络的性能。      

基于Gabor变换的高鲁棒汉字识别新方法

本文提出了针对字符图像的基于Gabor变换的汉字识别新方法.在对Gabor变换深入分析的基础上,本文针对汉字图像的统计信息,提出了一种有效的Gabor滤波器组参数优化方法;同时,对Gabor滤波器组的输出进行非线性变换,使其适应不同亮度和低质量灰度字符图像的识别.本文还改进了分块特征的抽取算法,提高了对字符细节的分辨能力.实验表明,这种特征抽取方法大大加强了识别系统抵御图像噪声、干扰、亮度变化、笔画模糊、笔画断裂以及字符形变的能力,在应用于各种低质量的二值或者灰度的印刷和脱机手写字符图像识别时,能获得较其他算法更良好的识别性能.     

Least-squares design of IIR filters with prescribed magnitude andphase responses and a pole radius constraint

This paper presents a method for the frequency domain design of infinite impulse response (IIR) digital filters. The proposed method designs filters approximating prescribed magnitude and phase responses. IIR filters of this kind can have approximately linear-phase responses in their passbands, or they can equalize magnitude and phase responses of given systems. In many cases, these filters can be implemented with less memory and with fewer computations per output sample than equivalent finite impulse response (FIR) digital filters. An important feature of the proposed method is the possibility to specify a maximum radius for the poles of the designed rational transfer function. Consequently, stability can be guaranteed, and undesired effects of implementations using fixed-point arithmetic can be alleviated by restricting the poles to keep a prescribed distance from the unit circle. This is achieved by applying Rouche's theorem in the proposed design algorithm. We motivate the use of IIR filters with an unequal number of poles and zeros outside the origin of the complex plane. In order to satisfy simultaneous specifications on magnitude and phase responses, it is advantageous to use IIR filters with only a few poles outside the origin of the z-plane and an arbitrary number of zeros. Filters of this type are a compromise between IIR filters with optimum magnitude responses and phase-approximating FIR filters. We use design examples to compare filters designed by the proposed method to those obtained by other methods. In addition, we compare the proposed general IIR filters with other popular more specialized structures such as FIR filters and cascaded systems consisting of frequency-selective IIR filters and phase-equalizing allpass filters     

A design of linear-phased IIR Nyquist filters

This paper discusses a new method of designing linear-phased IIR Nyquist filters with zero intersymbol interference. The filters designed by this method possess linear-phase characteristics and are lower in order than other Nyquist filters designed by existing methods. Expressions are derived for zero-phased IIR Nyquist filters and efficient design methods are examined for them. The opted design method is based on an iteration process, and in each iteration step a modified version of the Remez exchange algorithm is used. In addition, the implementation of the designed zero-phased IIR filters is considered. Finally, the proposed design method is demonstrated through various design examples     

The Application of Particle Swarm Optimization to Passive and Hybrid Active Power Filter Design

This paper proposes an optimal design method for passive power filters (PPFs) and hybrid active power filters (HAPFs) set at high voltage levels to satisfy the requirements of harmonic filtering and reactive power compensation. Multiobjective optimization models for PPF and HAPF were constructed. Detuning effects and faults were also considered by constructing constraints during the optimal process, which improved the reliability and practicability of the designed filters. An effective strategy was adopted to solve the multiobjective optimization problems for the designs of PPF and HAPF. Furthermore, the particle swarm optimization algorithm was developed for searching an optimal solution of planning of filters. An application of the method to an industrial case involving harmonic and reactive power problems indicated the superiority and practicality of the proposed design methods.      

Design of optimal minimum phase FIR filters by direct factorization

An improved method of designing optimal minimum phase FIR filters by directly finding zeros is proposed. The zeros off the unit circle are found by an efficient special purpose root-finding algorithm without deflation. The proposed algorithm utilizes the passband minimum ripple frequencies to establish the initial points, and employs a modified Newton's iteration to find the accurate initial points for a standard Newton's iteration. We show, with examples, that the proposed algorithm can be used to design very long filters (L = 325) with very high stopband attenuations.     

Design of Exactly Linear Phase $K$-Regular IIR Half-Band Filter

This paper proposes a novel method to design exactly linear phase infinite impulse response half-band filters with arbitrary regularity. Broadly speaking, the design problem is formulated as a semi-infinite program, which is then turned into a semidefinite program of minimal order via a new linear matrix inequality characterization of convex hulls of trigonometric polynomials. In contrast to maximally flat approach, the proposed method allows direct control of various design parameters, which in turn enables the synthesis of filters with better transition response. The viability of the proposed method is demonstrated through several numerical examples.      

基于张量滤波器的多波段红外目标辨识

红外多光谱成像在面对红外干扰和红外隐身时具有全色红外成像手段无法比拟的优势,通过目标在不同波段响应的差异,能有效地克服干扰和检测隐身目标.但传统基于向量的方法在处理多波段图像时没有有效地利用光谱和空间之间的相关性,通过在张量框架下设计辨识方法时,可以综合利用多光谱图像的光谱和空间特性.在设计张量辨识方法时,多光谱图像被...     

Generalized Chebyshev Filters for the Design of IIR Filters and Filter Banks

This paper introduces the generalized IIR Chebyshev filters. The proposed filters are obtained by applying bilinear transformation to the corresponding analog filters. The novelty of the method is the introduction of a new rational Chebyshev function, which includes Chebyshev Type I and Chebyshev Type II IIR filters as special cases. The application of the proposed digital filters to design perfect reconstruction two-channel filter banks is described. The proposed filters can be applied in orthogonal discrete wavelet transform.     

Novel droop-compensated comb decimation filter with improved alias rejections

This paper presents a novel two-stage comb decimator with the improved magnitude characteristic. Simple multiplierless corrector filters, which are designed using the frequency sampling and IFIR methods, are introduced. The proposed filters compensate the comb passband droop in the wideband passband region and increase the attenuations in the folding bands. Using the multirate identity the filters may be moved to a lower rate. The filter design depends only on the number of the cascaded comb filters and do not depend on the decimation factor M.     

Least-Squares Design of FIR Filters Based on a Compacted Feedback Neural Network

The design of finite-impulse response (FIR) filters can be performed by using neural networks by formulating the objective function to a Lyapunov energy function. Focusing on this goal, the authors present an improved structure of a feedback neural network to implement the least-squares design of FIR filters. In addition to using the closed-form expressions for the synaptic weight matrix and the bias parameter of the Hopfield neural network (HNN), the proposed approach can achieve a notable reduction both in the amount of computation required and hardware complexity compared to the previous neural-based method. Simulation results indicate the effectiveness of the proposed approach