Multispectral classification of Landsat-images using neuralnetworks

The authors report the application of three-layer back-propagation networks for classification of Landsat TM data on a pixel-by-pixel basis. The results are compared to Gaussian maximum likelihood classification. First, it is shown that the neural network is able to perform better than the maximum likelihood classifier. Secondly, in an extension of the basic network architecture it is shown that textural information can be integrated into the neural network classifier without the explicit definition of a texture measure. Finally, the use of neural networks for postclassification smoothing is examined     

Learning algorithm for pattern classification using cellular neuralnetworks

A new learning algorithm for pattern classification using cellular neural networks is described. The authors show that patterns belonging to the training set as well as patterns outside it can be classified reliably using the proposed algorithm. Comparisons with well established classification techniques clearly highlight the performances of the approach developed herein     

Information rates of non-Gaussian processes

The rate distortion functionR(D)of an information source was introduced by Shannon to specify the channel capacity required in transmitting information from the source with an average distortion not exceedingD. Exact rates have been calculated for Gaussian sources under a mean-square error criterion. For non-Gaussian continuous sources, Shannon has given upper and lower bounds onR(D). In specific cases, the difference between these two bounds may not be sufficiently small to provide a useful estimate ofR(D). The present paper is concerned with improving estimates of information rates of non-Gaussian sources under a mean-square error criterion. The sources considered are ergodic, and their statistical properties are characterized by a bounded and continuousn-dimensional probability density function. The paper gives a set of necessary and sufficient conditions forR(D)to equal Shannon's lower bound. For sources satisfying these conditions, exact rate calculations are possible. For sources that do not satisfy the required conditions, an improved upper bound is obtained that never exceeds Shannon's upper bound. Under rather general conditions, the new upper bound approaches Shannon's lower bound for small values of distortion, so that the true value ofR(D)can be estimated very accurately for smallD.     

Decision-theoretic approach to visual inspection using neuralnetworks

The Bayesian real-time network (BARTIN) is applied to solving a visual-inspection problem requiring translation, rotation and scale (TRS) invariance. The system is capable of classifying n-fold symmetric engineering parts from near-axial views which may contain more than one part. It is evaluated and compared with other approaches using real visual-inspection data. A novel TRS-invariant preprocessor, the polygon transform, which is optimised for near-circular objects, provides information about the line and circle structure in two-dimensional images. An integral part of the polygon transform is a new Hough transform for circle radii used for both scale invariance and image characterisation. The BARTIN formalism is presented from the viewpoint of subjective Bayesian analysis, and this approach demonstrates how the personal probabilities and utilities of BARTIN can be used to optimise an externally provided reward function. A method is given for adjusting the global level of caution. To handle sparse training data, parameter parsimony in the observer was achieved using a structure comprising a stripped-out Parzen-windows classifier followed by a softmax perceptron trim. For real-time operation, the system is initialised by pretraining it using data extracted from design drawings     

Noncausal nonminimum phase ARMA modeling of non-Gaussian processes

A method is presented for the estimation of the parameters of a noncausal nonminimum phase ARMA model for non-Gaussian random processes. Using certain higher order cepstra slices, the Fourier phases of two intermediate sequences (h_min(n) and h_max(n)) can be computed, where h_min(n) is composed of the minimum phase parts of the AR and MA models, and h_max(n) of the corresponding maximum phase parts. Under the condition that there are no zero-pole cancellations in the ARMA model, these two sequences can be estimated from their phases only, and lead to the reconstruction of the AR and MA parameters, within a scalar and a time shift. The AR and MA orders do not have to be estimated separately, but they are by product of the parameter estimation procedure. Through simulations it is shown that, unlike existing methods, the estimation procedure is fairly robust if a small order mismatch occurs. Since the robustness of the method in the presence of additive noise depends on the accuracy of the estimated phases of h_min(n) and h_max(n), the phase errors due to finite length data are studied and their statistics are derived     

High-order image subsampling using feedforward artificial neuralnetworks

We propose a method for high-order image subsampling using feedforward artificial neural networks (FANNs). In our method, the high-order subsampling process is decomposed into a sequence of first-order subsampling stages. The first stage employs a tridiagonally symmetrical FANN, which is obtained by applying the design algorithm introduced by Dumitras and Kossentini (see IEEE Trans. Signal Processing, vol.48, p.1446-55, 2000). The second stage employs a small fully connected FANN. The algorithm used to train both FANNs employs information about local edges (extracted using pattern matching) to perform effective subsampling of both high detail and smooth image areas. We show that our multistage first-order subsampling method achieves excellent speed-performance tradeoffs, and it consistently outperforms traditional lowpass filtering and subsampling methods both subjectively and objectively.     

Trajectory prediction for moving objects using artificial neuralnetworks

A method to predict the trajectory of moving objects in a robotic environment in real-time is proposed and evaluated. The position, velocity, and acceleration of the object are estimated by several neural networks using the six most recent measurements of the object coordinates as inputs. The architecture of the neural nets and the training algorithm are presented and discussed. Simulation results obtained for both 2D and 3D cases are presented to illustrate the performance of the prediction algorithm. Real-time implementation of the neural networks is considered. Finally, the potential of the proposed trajectory prediction method in various applications is discussed     

Adaptive speed control for induction motor drives using neuralnetworks

In this paper, the adaptive speed control of induction motor drives using neural networks is presented. To obtain good tracking and regulating control characteristics, a digital two-degree-of-freedom (2DOF) controller is adopted and a design procedure is developed for systematically finding its parameters according to prescribed specifications. The parameters of the controller corresponding to various drive parameter sets are found off-line and used as the training patterns to estimate the connection weights of neural networks, Under normal operation, the true drive parameters are real-time identified and they are converted into the controller parameters through multilayer forward computation by neural networks. The parameters of the 2DOF controller can be adapted to match the desired specifications under various operating conditions     

Integration of acoustic and visual speech signals using neuralnetworks

Results from a series of experiments that use neural networks to process the visual speech signals of a male talker are presented. In these preliminary experiments, the results are limited to static images of vowels. It is demonstrated that these networks are able to extract speech information from the visual images and that this information can be used to improve automatic vowel recognition. The structure of speech and its corresponding acoustic and visual signals are reviewed. The specific data that was used in the experiments along with the network architectures and algorithms are described. The results of integrating the visual and auditory signals for vowel recognition in the presence of acoustic noise are presented     

Modeling non-Gaussian time-varying vector autoregressive processes by particle filtering

We present a novel and general methodology for modeling time-varying vector autoregressive processes which are widely used in many areas such as modeling of chemical processes, mobile communication channels and biomedical signals. In the literature, most work utilize multivariate Gaussian models for the mentioned applications, mainly due to the lack of efficient analytical tools for modeling with non-Gaussian distributions. In this paper, we propose a particle filtering approach which can model non-Gaussian autoregressive processes having cross-correlations among them. Moreover, time-varying parameters of the process can be modeled as the most general case by using this sequential Bayesian estimation method. Simulation results justify the performance of the proposed technique, which potentially can model also Gaussian processes as a sub-case.     

应用非高斯重尾分布模型的子带滤波

图像经小波处理后,其细节分量具有明显高尖峰重拖尾分布的非高斯特性.据此,提出一种广义高斯α稳定分布的双参数柯西混合描述模型,并给出了相应的基于贝叶斯收缩和最大后验概率收缩的非线性滤波器的设计技术.通过与小波软门限滤波器方法比较,并对含加性噪声的标准图像进行抑制实验,结果表明,设计的滤波器具有较好的子带滤波效果,达到了有...     

Cluster analysis preprocessor for path loss modelling using neuralnetworks

The neural networks technique is applied to model path loss of indoor radio propagation. Cluster analysis is employed as a preprocessor to simplify the characterisation of the complicated indoor environment. Simulation results demonstrate that this method is feasible, resulting in a substantial reduction of data input requirement     

Recursive estimation of images using non-Gaussian autoregressivemodels

We consider recursive estimation of images modeled by non-Gaussian autoregressive (AR) models and corrupted by spatially white Gaussian noise. The goal is to find a recursive algorithm to compute a near minimum mean square error (MMSE) estimate of each pixel of the scene using a fixed lookahead of D rows and D columns of the observations. Our method is based on a simple approximation that makes possible the development of a useful suboptimal nonlinear estimator. The algorithm is first developed for a non-Gaussian AR time-series and then generalized to two dimensions. In the process, we draw on the well-known reduced update Kalman filter (KF) technique of Woods and Radewan (1977) to circumvent computational load problems. Several examples demonstrate the non-Gaussian nature of residuals for AR image models and that our algorithm compares favorably with the Kalman filtering techniques in such cases.     

Detection of non-Gaussian signals using integrated polyspectrum

We consider the problem of detecting an unknown, random, stationary, non-Gaussian signal in Gaussian noise of unknown correlation structure. The same framework applies if one desires to determine whether the given random signal is non-Gaussian. The most commonly used method for detection of random signals is the so-called energy detector, which cannot distinguish between Gaussian and non-Gaussian signals and requires the knowledge of the noise power. Recently, the use of bispectrum and/or trispectrum of the signal has been suggested for detection of non-Gaussian signals. The higher order spectra-based detectors do not require the knowledge of the noise statistics if the noise is Gaussian. In this paper, we suggest the use of an integrated polyspectrum (bispectrum of trispectrum) to improve computational efficiency of the detectors based on polyspectrum and to possibly further enhance their detection performance. We investigate conditions under which use of the integrated polyspectrum is appropriate. The detector structure is derived, acid its performance is evaluated via simulations and comparisons with several other existing approaches     

On theepsilon-entropy and the rate-distortion function of certain non-Gaussian processes

Letxi = {xi(t), 0 leq t leq T}be a process with covariance functionK(s,t)andE int_0^T xi^2(t) dt < infty. It is proved that for everyvarepsilon > 0thevarepsilon-entropyH_{varepsilon}(xi)satisfies begin{equation} H_{varepsilon}(xi_g) - mathcal{H}_{xi_g} (xi) leq H_{varepsilon}(xi) leq H_{varepsilon}(xi_g) end{equation} wherexi_gis a Gaussian process with the covarianeeK(s,t)andmathcal{H}_{xi_g}(xi)is the entropy of the measure induced byxi(in function space) with respect to that induced byxi_g. It is also shown that ifmathcal{H}_{xi_g}(xi) < inftythen, asvarepsilon rightarrow 0begin{equation} H_{varepsilon}(xi) = H_{varepsilon}(xi_g) - mathcal{H}_{xi_g}(xi) + o(1). end{equation} Furthermore, ff there exists a Gaussian processg = { g(t); 0 leq t leq T }such thatmathcal{H}_g(xi) < infty, then the ratio betweenH_{varepsilon}(xi)andH_{varepsilon}(g)goes to one asvarepsilongoes to zero. Similar results are given for the rate-distortion function, and some particular examples are worked out in detail. Some cases for whichmathcal_{xi_g}(xi) = inftyare discussed, and asymptotic bounds onH_{varepsilon}(xi), expressed in terms ofH_{varepsilon}(xi_g), are derived.     

Harmonic retrieval in colored non-Gaussian noise using cumulants

This paper addresses the harmonic retrieval problem in colored linear non-Gaussian noise of unknown covariance and unknown distribution. The assumptions made in the reported studies that the non-Gaussian noise is asymmetrically distributed and no quadratic phase coupling occurs are released. Using the elaborately defined fourth-order cumulants of the complex noisy observations, which are obtained by Hilbert transform, we can estimate either the correlation or the AR polynomial of the non-Gaussian noise via cumulant projections or ARMA modeling; then, the prewhitening or prefiltering techniques can be employed to retrieve harmonics, respectively. Simulation results are presented to demonstrate the effectiveness of the proposed algorithms     

非高斯双变量模型contourlet图像去噪

Contourlet变换是继小波变换之后的又一新变换.由于contourlet变换的多尺度和多方向特性,能有效地捕获到自然图像中的轮廓,并对其进行稀疏表示.详细分析了图像contourlet系数的统计特性,并利用非高斯双变量分布对系数层间相关性进行建模.最后,将此分布应用于图像去噪,就PSNR、NMSE和视觉质量这三方面的评价指标与contourlet HMT和小波阈值法进行了比较.实验结果表明：算法能获得较好的结果,尤其是对于含有丰富纹理的图像.     

Blind identification of quadratic nonlinear models using neuralnetworks with higher order cumulants

A novel approach to blindly estimate kernels of any discrete- and finite-extent quadratic models in higher order cumulants domain based on artificial neural networks is proposed in this paper. The input signal is assumed an unobservable independently identically, distributed random sequence which is viable for engineering practice. Because of the properties of the third-order cumulant functions, identifiability of the nonlinear model holds, even when the model output measurement is corrupted by a Gaussian random disturbance. The proposed approach enables a nonlinear relationship between model kernels and model output cumulants to be established by means of neural networks. The approximation ability of the neural network with the weights-decoupled extended Kalman filter training algorithm is then used to estimate the model parameters. Theoretical statements and simulation examples together with practical application to the train vibration signals modeling corroborate that the developed methodology is capable of providing a very promising way to identify truncated Volterra models blindly     

Stochastic differential equations: an approach to the generation ofcontinuous non-Gaussian processes

The generation of continuous random processes with jointly specified probability density and covariation functions is considered. The proposed approach is based on the interpretation of the simulated process as a stationary output of a nonlinear dynamic system, excited by white Gaussian noise and described by a system of a first-order stochastic differential equations (SDE). The authors explore how the statistical characteristics of the equation's solution depends on the form of its operator and on the intensity of the input noise. Some aspects of the approximate synthesis of stochastic differential equations and examples of their application to the generation of non-Gaussian continuous processes are considered. The approach should be useful in signal processing when it is necessary to translate the available a priori information on the real random process into the language of its Markov model as well as in simulation of continuous correlated processes with the known probability density function     

Using physics-based modeler outputs to train probabilistic neuralnetworks for unexploded ordnance (UXO) classification in magnetometrysurveys

The outputs from a physics-based modeler of magnetometry data have been successfully used with a probabilistic neural network (PNN) to discriminate unexploded ordnance (UXO) from ordnance-related scrap. Cross-validation predictions were performed on three data sets to determine which modeler parameters were most valuable for UXO classification. The best performing parameter combination consisted of the modeler outputs depth, size, and inclination. The cross-validation results also indicated that good prediction performance could be expected. Model outputs from one location at a site were used to train a PNN model, which could correctly discriminate UXO from scrap at a different location of the same site. In addition, data from one site, the former Buckley Field, Arapahoe County, CO, was used to predict targets detected at an entirely different training range. The Badlands Bombing Range, Bull's Eye 2 (BBR 2), Cuny Table, SD. Through careful selection of the probability threshold cutoff, the UXO detection rate obtained was 95% with a false alarm rate of only 37%. The ability to distinguish individual UXO types has been demonstrated with correct classifications between 71% and 95%