G. Orlandi

Car plate recognition by neural networks and image processing

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.

A generalized learning paradigm exploiting the structure of feedforward neural networks

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.

Fast adaptive digital equalization by recurrent neural networks

Neural networks (NNs) have been extensively applied to many signal processing problems. In particular, due to their capacity to form complex decision regions, NNs 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 (RNNs) 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 RNNs a technique applied with success to fast learning of feedforward NNs 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 PAM signals in different transmission channels are described, which demonstrate the effectiveness of the proposed approach.

Fast combinatorial RNS processors for DSP applications

It is known that RNS VLSI processors can parallelize fixed-point addition and multiplication operations by the use of the Chinese remainder theorem (CRT). The required modular operations, however, must use specialized hardware whose design and implementation can create several problems. In this paper a modified residue arithmetic, called pseudo-RNS is introduced in order to alleviate some of the RNS problems when digital signal processing (DSP) structures are implemented. Pseudo-RNS requires only the use of modified binary processors and exhibits a speed performance comparable with other RNS traditional approaches. Some applications of the pseudo-RNS to common DSP architectures, such as multipliers and filters, are also presented in this paper. They are compared in terms of the area-time square product versus other RNS and weighted binary structures. It is proven that existing combinatorial or look-up table approaches for RNS are tailored to small designs or special applications, while the pseudo-RNS approach remains competitive also for complex systems. >

A systolic redundant residue arithmetic error correction circuit

In highly integrated processors, a concurrent fault tolerance capability is particularly important, especially for real-time applications. In fact, in these systems, transient errors are often present, but are difficult to correct online. Error recovery procedures applied for each processing or memory element require large amount of hardware and can reduce throughput. Residue arithmetic has intrinsic fault tolerance capability and requires less complex hardware. A single error correction procedure based on the use of a redundant residue number system (RRNS) and the base extension operation is proposed. The method uses a very small decision table and works in parallel mode; therefore it is suitable for high speed VLSI circuit realization. A parallel systolic architecture which realizes the algorithm is introduced. >

Multi-source localization in reverberant environments by ROOT-MUSIC and clustering

Localization of acoustic sources in reverberant environments by microphone arrays remains a challenging task in audio signal processing. As a matter of fact, most assumptions of commonly adopted models are not met in real applications. Moreover, in practical systems it is not convenient or possible to employ sophisticated and costly architectures, that require precise synchronization and fast data shuffling among sensors. In this paper, a new robust multi-step procedure for speaker localization in reverberant rooms is introduced and described. The new approach is based on a disturbed harmonics model of time delays in the frequency domain and employs the well-known ROOT-MUSIC algorithm, after a preliminary distributed processing of the received signals. Candidate source positions are then estimated by clustering of raw TDOA estimates. Main features of the proposed approach, compared to previous solutions, are the capability of tracking multiple speakers and the high accuracy of the closed form TDOA estimator.

Low-sensitivity recursive digital filters obtained via the delay replacement

A method is proposed for synthesizing very low-sensitivity recursive digital filters on the basis of three techniques: 1) delay replacement; 2) transformation of resistance-terminated reactive ladders into digital structures; 3) multiplier scaling. The method applies to the case where the poles of the transfer function of the filter are near z = 1 , as is the case when the sampling rate is increased. The low-sensitivity property of the resulting schemes is proved and illustrated by examples.

Low-sensitivity digital filters based on zero extraction

A method is proposed for synthetizing digital filters derived from the LBR two-pair extraction technique. The method forces the LBR two-pairs to have their transmission zeros coincident with those of the filter. The basic two-pair having transmission zeros on the unit circle is considered in detail with regard both to the realization problem and to the simplification of the synthesis procedure. In particular, explicit formulas are proposed for directly determining the representation of this two-pair. A simulation example illustrates the sensitivity performance of the resulting schemes.

ARMA estimation by the classical predictor

The relation between the parameters of an ARMA process and those of the equivalent AR process is investigated. On the basis of this relation an algorithm is proposed for determining the ARMA parameters from the values of the taps of the predictor obtained by applying any of the classical methods to the process under consideration.

Fast SVD-based algorithm for signal selective DOA estimation

The classical formulation of subspace direction of arrival (DOA) estimation algorithms, such as MUSIC, assumes temporally independent source signals and white noise background. If only a subset of source signals are of interest (SOI), a computationally expensive classification procedure should be performed after both DOA parameter estimation and signal extraction steps. In this paper, a simple and robust SVD-based estimation procedure is proposed, making use of an orthogonal two-channel decomposition of sensor signals in the time domain, optimized for the SOI. Simulation results are provided to support the capability of the proposed algorithm of discarding temporally uncorrelated interference and identifying the SOI DOA parameters only, without any need of post-processing.