Fast nearest neighbour searching algorithm using L1 norm

Simple but effective fast codebook searching algorithms for vector quantisation are presented. The Euclidean distance (L₂ norm) calculation requires a number of multiplications. The proposed algorithms use the L₁ norm between the input vector and codeword to discard many unlikely codewords. Since the proposed algorithms significantly reduce the number of multiplications, a considerable reduction in encoding time is achieved. Simulation results confirm the effectiveness of the proposed algorithms     

具有双重稀疏机制的在线学习算法

针对大数据背景下数据分类问题,已有的在线学习算法通常引入L₁范数正则化增强预测模型的稀疏性,但单一的正则化约束不能高效的获取稀疏模型.基于此,提出了一种具有双重稀疏机制的在线学习算法（an online learning algorithm with dual sparse mechanisms,DSOL）.在DSOL算法中,一方面利用L_1/2正则化项约束目标函数以增强预测模型的稀疏性,提高算法的泛化性能.另一方面用改进的梯度截取法对数据特征进行选择,有效稀疏化预测模型.通过L_1/2正则化与改进的梯度截取策略的有机融合,有效利用了历史数据信息,提高了算法分类数据的性能.通过与另4种代表性稀疏在线学习算法在9个公开数据集的实验对比表明DSOL算法对数据分类的准确性更高.     

一种基于L_1范数的非局部变分图像复原模型

针对L_2范数的非局部变分模型在迭代过程中未考虑图像局部梯度信息,模糊图像细节信息的缺点,提出了一种基于L_1范数的非局部变分模型。首先,对基于L_1范数的非局部变分模型的扩散性能进行了详细的分析。接着,将该模型应用于退化图像的复原中,并推导出该模型的Bregman交替迭代求解过程。最后,通过对比实验,证明本文提出的L_1范数的非局部变分复原模型能更好地重构图像的细节信息,相对于L_2范数的非局部变分模型峰值信噪比提高大于1dB,图像复原性能更优。     

一种宽带相干信源的无网格超分辨DOA估计方法

由于无网格（grid-less）稀疏重构方法的波达方向（direction of arrival,DOA）估计数学模型为单快拍形式,因此该方法只有在噪声电平趋近于零时才具有优越的性能.为了提高grid-less方法在信噪比（signal-to-noise ratio,SNR）较低时宽带相干信源的估计性能,提出了一种多快拍grid-less DOA估计方法.首先,对多快拍阵列观测矢量实施奇异值分解（singular value decomposition,SVD）获得观测矩阵的时域信号子空间,通过观测矩阵到时域信号子空间的投影实现观测矩阵的降噪;然后,为了不增加多快拍计算复杂度,将降噪后观测矩阵的列向量加权累加处理得到单快拍形式;最后,从理论上证明了本文提出的GL-SVD方法求解的模型是凸的,能够实现宽带信号DOA的精确重构.仿真结果表明,该方法在低SNR以及宽带相干信源情况下的估计精度都高于L₁范数最小化奇异值分解（L₁-norm minimum singular value decomposition,L₁-SVD）和离格稀疏贝叶斯推断奇异值分解（off-grid sparse Bayesian inference singular value decomposition,OGSBI-SVD）,且在较小角度间隔的情况下具有更高的估计概率和分辨率.     

An L1-Method for the Design of Linear-Phase FIR Digital Filters

This paper considers the design of linear-phase finite impulse response digital filters using an L₁ optimality criterion. The motivation for using such filters as well as a mathematical framework for their design is introduced. It is shown that L₁ filters possess flat passbands and stopbands while keeping the transition band comparable to that of least-squares filters. The uniqueness of L₁-based filters is explored, and an alternation type theorem for the optimal frequency response is derived. An efficient algorithm for calculating the optimal filter coefficients is proposed, which may be viewed as the analogue of the celebrated Remez exchange method. A comparison with other design techniques is made, demonstrating that the L₁ approach may be a good alternative in several applications.     

具有容错能力的L1最优化半自动2D转3D

半自动2D转3D的关键是将用户分配的稀疏深度转换为稠密深度.现有方法没有充分考虑纹理图像和深度图之间的结构差异,以及2D转3D对用户误标注的容错性.针对上述问题,借助L₁范数对异常数据的抵制,在一个统一框架下实现结构相关具有容错能力的稀疏深度稠密插值.首先,利用L₁范数表示估计深度和用户分配深度在标注位置的差异,建立数据项;其次,根据特征的相似性用L₁范数计算局部相邻像素点之间的深度差异,建立局部正则项;再次,对图像进行超像素分割,根据不同超像素内代表性像素点之间深度差异的L₁测度,建立全局正则项;最后,用上述数据项和正则项构建能量函数,并通过分裂Bregman算法予以求解.无误差和有误差情况下的实验结果表明,与边缘保持的最优化插值、随机游走、混合图割与随机游走、软分割约束的最优化插值和非局部化随机游走相比,本文估计深度图绘制的虚拟视点图像空洞和伪影损伤更小.在误操作情况下,本文比上述方法PSNR改善了0.9dB以上,且在视觉上屏蔽了用户误操作的影响.     

Robust linear prediction of speech signals based on orthogonalframework

A linear prediction is formulated via the orthogonal principle to facilitate the incorporation of various error minimisation criteria. A weighting function, which downplays extreme errors, is used to provide a robust estimate lying between the L₁ and L₂ criteria. Experiments based on synthetic vowels reveal that the proposed method outperforms the L₁ and L₂ estimates     

Design of QMF banks based on derivative information

New L₂ objective functions for the design of quadrature mirror filter (QMF) banks are proposed. They are based on the derivative information of the reconstruction error. Simple and explicit matrix-form formulas for the proposed objective functions are derived. Efficient design methods are proposed by incorporating a separability technique into the derived optimality conditions on prototype filters. The proposed design methods need only solve linear equation iteratively without nonlinear optimisation. Design examples demonstrate that good low-delay QMF banks and linear-phase QMF banks can be obtained in only a few iterations. Compared with the conventional approach, the new approach leads to QMF banks with larger stopband attenuation and smaller reconstruction errors     

利用深度全卷积编解码网络的单通道语音增强

针对传统的神经网络未能对时频域的相关性充分利用的问题，提出了一种利用深度全卷积编解码神经网络的单通道语音增强方法。在编码端，通过卷积层的卷积操作对带噪语音的时频表示逐级提取特征，在得到目标语音高级特征表示的同时逐层抑制背景噪声。解码端和编码端在结构上对称，在解码端，对编码端获得的高级特征表示进行反卷积、上采样操作，逐层恢复目标语音。跳跃连接可以很好地解决极深网络中训练时存在的梯度弥散问题，本文在编解码端的对应层之间引入跳跃连接，将编码端特征图信息传递到对应的解码端，有利于更好地恢复目标语音的细节特征。 对特征融合和特征拼接两种跳跃连接方式、基于L1和 L2两种训练损失函数对语音增强性能的影响进行了研究，通过实验验证所提方法的有效性。      

基于子空间结构正则化的L21非负矩阵分解高光谱解混

标准的非负矩阵分解(NMF)应用于高光谱解混时,容易受到噪声和异常值的干扰,解混效果较差。为了提高分解性能,该文将L₂₁范数引入标准的NMF算法中,对模型进行了改进,从而提高算法的鲁棒性。其次,为了提高分解后丰度矩阵的稀疏性,将双重加权稀疏约束引入L₂₁NMF模型中,使其中一个权值提高每个像元对应的丰度向量上的稀疏性,另一个权值提高每个端元对应的丰度向量上的稀疏性。同时,为了利用像元的全局空间分布信息,观察地物在不同图像中的真实分布情况,引入子空间结构正则项,提出了基于子空间结构正则化的L₂₁非负矩阵分解(L₂₁NMF-SSR)算法。通过在模拟数据集和真实数据集与其他经典算法的比较,验证了该算法具有更好的性能,同时具有去噪能力。     

Data fusion with minimal communication

Two sensors obtain data vectors x and y, respectively, and transmit real vectors m&oarr;₁(x) and m&oarr;₂(y), respectively, to a fusion center. The authors obtain tight lower bounds on the number of messages (the sum of the dimensions of m&oarr;₁ and m&oarr;₂) that have to be transmitted for the fusion center to be able to evaluate a given function f&oarr;(x,y). When the function f&oarr; is linear, they show that these bounds are effectively computable. Certain decentralized estimation problems can be cast in the framework and are discussed in some detail. In particular, the authors consider the case where x and y are random variables representing noisy measurements and f&oarr;(x,y)=E[z|x,y], where z is a random variable to be estimated. Furthermore, it is established that a standard method for combining decentralized estimates of Gaussian random variables has nearly optimal communication requirements     

基于双金属结构的光纤光栅温度不敏感滤波器

为提高光纤陀螺宽谱光源的平均波长稳定性,提出了一种用60μm超短光纤光栅制作带宽11.77 nm温度不敏感滤波器的方法。利用金属材料的热膨胀系数差,设计了双金属温度补偿结构,能够随温度的升高/降低对光纤光栅压缩/拉伸,有效地补偿光纤光栅由热光效应引起的波长变化。在30～60℃的温度范围内,光纤光栅的温度灵敏度系数为0.15 pm/℃,较未补偿前降低了60倍以上。该结构具有较好的温度不敏感性,可作为光源滤波器提高光源的平均波长稳定性并有望用于高精度光纤陀螺。     

Design of low-order linear-phase IIR filters via orthogonalprojection

This paper presents an indirect linear-phase IIR filter design technique based on a reduction of linear-phase FIR filters. The desired filter is obtained by minimizing the L₂ norm of the difference between the original FIR filter and the lower order IIR filter. We first establish a relationship between the Hankel singular values of the discarded part of the FIR filter and the L₂ norm of the corresponding filter approximation error based on model truncation. This result motivates us to propose a simple finite search method that will achieve better approximation results than commonly used truncation methods such as the balanced truncation (BT) and the impulse response gramian (IRG) methods. We then develop an iterative algorithm for finding an optimal IIR filter based on a matrix projection of the original FIR filter. The convergence of the proposed algorithm is established. Filters designed using the proposed algorithm are compared with those obtained by other techniques with respect to the amplitude response and group delay characteristics in the passband. Numerical examples show that the proposed algorithm offers the best performance     

Recovery of a sparse spike time series by L1 normdeconvolution

An L₁ norm minimization scheme is applied to the determination of the impulse response vector h of flaws detected in practical examples of ultrasonic nondestructive evaluation in CANDU nuclear reactors. For each problem, parametric programming is applied to find the optimum value of the damping parameter that will yield the best estimate of h according to a quantified performance factor. This performance factor is based on a quantified analysis of the transitions in estimates of h as the damping parameter is varied over a wide range of possible values. It is shown that for the examined cases in which the true impulse response is a sparsely filled spike strain, the L₁ norm provides significantly better results than the more commonly used L₂ norm minimization schemes. These results are shown to be consistent with theoretical predictions     

Selective minimum-norm solution of the biomagnetic inverse problem

A new multidipole estimation method which gives a sparse solution of the biomagnetic inverse problem is proposed. This solution is extracted from the basic feasible solutions of linearly independent data equations. These feasible solutions are obtained by selecting exactly as many dipole-moments as the number of magnetic sensors. By changing the selection, the authors search for the minimum-norm vector of selected moments. As a result, a practically sparse solution is obtained; computer-simulated solutions for L_p-norm (p=2, 1, 0.5, 0.2) have a small number of significant moments around the real source-dipoles. In particular, the solution for L₁-norm is equivalent to the minimum-L₁-norm solution of the original inverse problem. This solution can be uniquely computed by using linear programming     

变形L1正则化的高光谱图像稀疏解混

如何准确地刻画易于求解的稀疏正则化函数是高光谱图像稀疏解混的难点。变形L₁正则化函数是一类由绝对值函数组成的双线性变换的单参数族,类似于L_pp∈0,1范数,通过调整参数a∈0,可以准确表征L₀和L₁之间的任意范数,并具有无偏、稀疏和Lipschitz连续性。论文首先研究变形L₁正则化函数,然后提出变形L₁正则化的高光谱稀疏解混变分模型,最后提出变形L₁正则化高光谱稀疏解混模型的凸函数差分求解算法。通过模拟和真实的高光谱数据实验,与经典的SUnSAL算法相比,表明提出的算法能够更准确地刻画丰度系数的稀疏性,并获得更高的解混精度。     

Novel cooperative neural fusion algorithms for image restoration and image fusion.

To deal with the problem of restoring degraded images with non-Gaussian noise, this paper proposes a novel cooperative neural fusion regularization (CNFR) algorithm for image restoration. Compared with conventional regularization algorithms for image restoration, the proposed CNFR algorithm can relax need of the optimal regularization parameter to be estimated. Furthermore, to enhance the quality of restored images, this paper presents a cooperative neural fusion (CNF) algorithm for image fusion. Compared with existing signal-level image fusion algorithms, the proposed CNF algorithm can greatly reduce the loss of contrast information under blind Gaussian noise environments. The performance analysis shows that the proposed two neural fusion algorithms can converge globally to the robust and optimal image estimate. Simulation results confirm that in different noise environments, the proposed two neural fusion algorithms can obtain a better image estimate than several well known image restoration and image fusion methods.     

Novel approach to the design of I/Q demodulation filters

A novel filter design approach to digital I/Q demodulation is proposed. Two possible realisations are presented using this approach. The first one is based on the highpass filter method which is suitable for B⩽f₀ and while the other realisation is based on the lowpass filter method suitable for B⩽f₀, where B and f ₀ are the IF signal bandwidth and the IF frequency, respectively. Both new realisations maintain the advantages of an earlier lowpass approach such as zero DC offset, matched channel frequency responses, and good performance over a wide bandwidth. At the same time, the new highpass filter realisation method possesses higher computational efficiency than other wideband approaches reported in the literature     

Semiblind channel estimation for CDMA systems with parallel data and pilot signals

Semiblind channel estimation combines the methods of channel estimation based on a pilot signal and blind channel estimation based on a data-only conveying signal. Maximum-likelihood (ML)-based semiblind estimators with Gaussian assumptions can provide improvement in performance, compared with channel-estimation schemes using the pilot signal only. This improvement can be even larger when the pilot and the data signals are sent simultaneously, as is the case in the third-generation wideband code-division multiple-access standards. However, the Gaussian ML approach results in very large complexity. Previously proposed semiblind methods with low complexity have been derived for serial pilot and data transmission, and are not suitable for the parallel transmission case. In this paper, algorithms for semiblind channel estimation for the parallel data and training signal case are developed. Approximations which reduce the computational complexity of the Gaussian ML method significantly are proposed. Solutions with iterations with very low attendant complexity are provided. The mean squared error analysis of the proposed method is obtained and compared with that of a method with no approximations. The approximations are justified through simulations, and the performance improvement over estimation schemes using the pilot signal solely is verified.