Existing algorithms for wideband direction finding are mainly based on local approximations of the Gaussian log-likelihood around the true directions of arrival (DOAs), assuming negligible array calibration errors. Suboptimal and costly algorithms, such as classical or sequential beamforming, are required to initialize a local search that eventually furnishes DOA estimates. This multistage process may be nonrobust in the presence of even small errors in prior guesses about angles and number of sources generated by inherent limitations of the preprocessing and may lead to catastrophic errors in practical applications. In this paper, a new approach to wideband direction finding is introduced and described. The proposed strategy combines a robust near-optimal data-adaptive statistic, called the weighted average of signal subspaces (WAVES), with an enhanced design of focusing matrices to ensure a statistically robust preprocessing of wideband data. The overall sensitivity of WAVES to various error sources, such as imperfect array focusing, is also reduced with respect to traditional CSSM algorithms, as demonstrated by extensive Monte Carlo simulations.

No abstract

No abstract

In this paper we describe an experimental system for the recognition of Italian-style car license plates. Images are usually taken from a camera at a toll gate and preprocessed by a fast and robust 1-D DFT scheme to find the plate and character positions. Characters are classified by a multilayer neural network trained by the recently developed BRLS learning algorithm. The same neural network replaces both the traditional feature extractor and the classifier. The percentage of correctly recognized characters reaches the best scores obtained in literature, being highly insensitive to the environment variability, while the architecture appears best suited for parallel implementation on programmable DSP processors

In recent years, neural networks (NN's) have been extensively applied to many signal processing problems. In particular, due to their capacity to form complex decision regions, NN's have been successfully used in adaptive equalization of digital communication channels. The mean square error (MSE) criterion, which is usually adopted in neural learning, is not directly related to the minimization of the classification error, i.e., bit error rate (BER), which is of interest in channel equalization. Moreover, common gradient-based learning techniques are often characterized by slow speed of convergence and numerical ill conditioning, In this paper, we introduce a novel approach to learning in recurrent neural networks (RNN's) that exploits the principle of discriminative learning, minimizing an error functional that is a direct measure of the classification error. The proposed method extends to RNN's a technique applied with success to fast learning of feedforward NN's and is based on the descent of the error functional in the space of the linear combinations of the neurons (the neuron space); its main features are higher speed of convergence and better numerical conditioning w.r.t. gradient-based approaches. whereas numerical stability is assured by the use of robust least squares solvers. Experiments regarding the equalization of PARI signals in different transmission channels are described, which demonstrate the effectiveness of the proposed approach

Wide-band Direction of Arrival (DOA) estimation with sensor arrays is an essential task in sonar, radar, acoustics, biomedical and multimedia applications. Many state of the art wide-band DOA estimators coherently process frequency binned array outputs by approximate Maximum Likelihood, Weighted Subspace Fitting or focusing techniques. This paper shows that bin signals obtained by filter-bank approaches do not obey the finite rank narrow-band array model, because spectral leakage and the change of the array response with frequency within the bin create ghost sources dependent on the particular realization of the source process. Therefore, existing DOA estimators based on binning cannot claim consistency even with the perfect knowledge of the array response. In this work, a more realistic array model with a finite length of the sensor impulse responses is assumed, which still has finite rank under a space-time formulation. It is shown that signal subspaces at arbitrary frequencies can be consistently recovered under mild conditions by applying MUSIC-type (ST-MUSIC) estimators to the dominant eigenvectors of the wide-band space-time sensor cross-correlation matrix. A novel Maximum Likelihood based ST-MUSIC subspace estimate is developed in order to recover consistency. The number of sources active at each frequency are estimated by Information Theoretic Criteria. The sample ST-MUSIC subspaces can be fed to any subspace fitting DOA estimator at single or multiple frequencies. Simulations confirm that the new technique clearly outperforms binning approaches at sufficiently high signal to noise ratio, when model mismatches exceed the noise floor.

In this paper a general class of fast learning algorithms for feedforward neural networks is introduced and described. The approach exploits the separability of each layer into linear and nonlinear blocks and consists of two steps. The first step is the descent of the error functional in the space of the outputs of the linear blocks (descent in the neuron space), which can be performed using any preferred optimization strategy. In the second step, each linear block is optimized separately by using a least squares (LS) criterion. To demonstrate the effectiveness of the new approach, a detailed treatment of a gradient descent in the neuron space is conducted. The main properties of this approach are the higher speed of convergence with respect to methods that employ an ordinary gradient descent in the weight space backpropagation (BP), better numerical conditioning, and lower computational cost compared to techniques based on the Hessian matrix. The numerical stability is assured by the use of robust LS linear system solvers, operating directly on the input data of each layer. Experimental results obtained in three problems are described, which confirm the effectiveness of the new method.

In this paper two novel nonlinear cascade adaptive architectures, here called sandwich models, suitable for the identification of general nonlinear systems are presented. The proposed architectures rely on the combination of structural blocks, each one implementing a linear filter or a memoryless nonlinear function. All the nonlinear functions involved in the adaptation process are based on spline functions and can be easily modified during learning using gradient-based techniques.\ud In particular, a simple form of the on-line adaptation algorithms for the two architectures is derived. In addition, we analytically obtain a bound for the selection of the learning rates involved in the learning algorithms, in order to guarantee a convergence towards a minimum of the cost function. Finally, some experimental results demonstrate the effectiveness of the proposed method