Advanced Signal Processing Technology by Soft Computing

The following sections are included:

• Series Editor's Preface

• Volume Editor's Preface

• Contents

One of the problems of 5% error rate encountered in Continuous speech recognition is partly due to the difficulty in the identification of a mixed up of two phonemes in a close concatenation. For instance, one speaks of “Let's go” instead of “Let us go”. There are two kinds of speech segmentations: the linguistic segmentation and the acoustic segmentation. The linguistic segmentation relies on a combination of acoustic, lexical, semantic, and statistical knowledge sources, which has been studied in [1,2,3,4]. Daily spoken conversations are usually abbreviated for speakers' convenience. The acoustic segmentation is to separate the mixed sounds such as /ts/ into /t/ & /s/ for automatically finding linguistic units. Adaptive Wavelet Transform (AWT) developed by Szu [6,9] is a linear superposition of banks of constant-Q zero-mean mother wavelets implemented by an Artificial Neural Network (ANN) called a “WaveNet”. Each neuron is represented by a daughter wavelet, which can be an affined scale change of identical or different mother wavelet for a continuous AWT. AWT was designed for the cocktail party effect and to solve the acoustic segmentation of phonemes using a supervised learning ANN architecture. In this paper, we reviewed AWT from Independent Component Analysis (ICA) viewpoint, and then applied Blind Source Separation (BSS) to the acoustic de-mixing and segmentation.

This article proposes a supervised Fourier demixing filter based on the statistical independence between an original signal and an observation noise. The estimation of the original signal is obtained as a weighted sum of each harmonic of the noisy observation signal. The filter is constructed by the network, and its weighting parameters are adjusted by learning with use of the higher order moments of an observation noise even though no information can be obtained on a dynamics of a target system. The proposed filter is widely applicable to an elimination of an observation noise with an arbitrary distribution. The effectiveness and the validity of the proposed filter have been verified by applying it to the filtering of a noisy speech signal.

In this contribution a novel tool for browsing and analysing large acoustic databases in sound engineering is presented, denoted as acoustic navigator. This tool exploits dimensionality reducing mappings and advanced interactive visualization techniques for navigation and visual exploratory data analysis in acoustic databases. In addition to database projection and display, complementing interactive access to individual sounds and respective comparison facility is provided. Thus, the analysis of key features for certain acoustical impressions, e.g. richness or pleasance of a sound, can be efficiently identified and consequently used in sound engineering. Further, the available and proven design methods and tools for sound engineering are complemented by interactive facilities in the acoustic navigator.

A hybrid multilayer perceptron neural network is described and discussed. The considered hybrid is based on utilizing an adaptive centroid transformation which adds an extra hidden layer to the network structure. Effect of this centroid transformation is discussed, and basic training methods for learning the transformation are described. In addition, a method for determining the number of units in the extra layer is outlined, and network weight initialization is considered. Empirical simulations demonstrate the features of the hybrid method. The method is expected to provide substantial gains especially in complex classification problems.

Transiently Evoked Otoacoustic Emission(TEOAE) is low-level sound emitted by cochlea when induced by a brief acoustic stimulus such as click. As a quick, non-invasive, and objective method to test the cochlear function, TEOAE has been investigated and applied widely. However, the traditional methods are not well suited for its discrimination and analysis. Aiming at the limitations, this research put forward new identification criterion and analysis methods. Short-time Spectrum(STFT) has been demonstrated an effective method to identify between TEOAE and stimulus artifact, and analyze qualitatively TEOAE time-frequency properties. Average Energy Distribution(AED) constructed by Wavelet Transform provided a precise way to analyze quantitatively TEOAE frequency information.

In practice, it is not likely that enough data about the failure behavior can be collected to use the probability for reliability analyzing. In some cases, the information we have about the functioning of components and systems is not based on statistics, but is of a linguistic nature. Very often the reliability of the system is presented in a form of the mean time to failure (MTTF) or comparative MTTFs. In order to use this reliability measure for further computations the new models have to be developed. A theory of imprecise probabilities might be a basis for reliability analyzing when we have such the incomplete information about reliability behavior of systems. The basic tool for computing new reliability measures is the natural extension which can be regarded as a linear optimization problem. The reliability of series, parallel, cold standby, m-out-of-n and simple repairable systems is provided by using the natural extension. The chapter may be viewed as an attempt to generalize the classical reliability theory based on the probabilistic models and to study the basic properties and advantages of the new reliability models.

Breast cancer most commonly presents as a mass. In this chapter, we describe an efficient technique for mass detection. This technique utilizes morphological filtering, a finite generalized Gaussian mixture (FGGM) modeling, and a contextual Bayesian relaxation labeling technique (CBRL) to enhance and extract suspicious masses. A feature space is constructed based on multiple feature extraction from the regions of interest (ROIs). Finally, a multi-modular neural network (MMNN) is employed to distinguish true masses from non-masses. We have applied these methods to test our mammogram database. The experimental results indicate that morphological filtering combined with FGGM model-based segmentation is an effective way to extract mammographic suspicious mass patterns. Compared with conventional neural networks, the probabilistic MMNN can lead to a more efficient learning algorithm and can provide more understanding in the analysis of the distribution patterns of multiple features extracted from the suspicious masses.

Imagery edges occur naturally in human visual systems as a consequence of redundancy reduction towards “sparse & orthogonality feature maps,” which have been recently derived from the maximum entropy information- theoretical first principle of artificial neural networks. After a brief math review of such an Independent Component Analysis (ICA) or Blind Source Separation (BSS) of edge maps, we explore the de-mixing condition for more than two imagery objects recognizable by an intelligent pair of cameras with memory in a time-multiplex fashion.

We review two smart cameras with short term working memory to do better image de-mixing. We consider channel communication application that we can efficiently mix four images using matrices [A₀] and [A₁] to send through two channels.

In this paper, two wire routing algorithms called LMR and LMR-FSC are proposed. Both of the algorithms are based on the Lagrangian method for the continuous valued constrained optimization problem. In LMR (Lagrangian method for routing) all the nets are routed simultaneously. The problem of deciding the net-ordering and rip-up reroute scoring metrics doesn't exist, rather nets are routed by a natural continuous valued shove aside resulting from the mutual competition of the nets. The characteristics of massive parallelism of LMR makes it suitable for neurocomputing. LMR-FSC (Lagrangian method for routing with forced satisfaction of consistency) is a cost-based iterative algorithm. Initially illegal condition of overlapping of different nets is allowed. The overlapping of different nets is accounted as cost. The wires placed on the higher cost parts are gradually moved to the lower cost parts. Each net is routed for a minimum cost. Effectiveness of both the approaches is shown by some experimental results.

Hebbian learning of synaptic weight W_ij between the i-th and the j-th neurons lies in the heart of Artificial Neural Networks (ANN) whose collective dynamics is re-derived herewith in terms of ANN'S effect upon a single net input of the i-th neuron, u_i(t_{n + 1}) = K(t_n) σ(u_i(t_n)) - θ_i, where we have defined u_i(t_n) = Σ_j W_ij(t_n) v_j(t_n), and threshold θ_i. This single neuron dynamics is similar to the classical technique used in solving the many-body problem with a single-body Greens function. However, the proposed model is nonlinear, exact and separable because of the product nature of Hebbian law W_ij ≈ v_i v_j, where v_i; is the output firing rate, v_i = σ(u_i), determined by the net input u_i and a nonlinear sigmoidal function. Moreover, the derived equation is identical to an iterative map generalized the population logistic map. In case of ANN, however, every neuron map has a different slope function, K(t_n)=Σ_j≠iσ^±(u_i(t_n))v_j(t_n), reflecting the dynamic learning process. Such an ANN Green's function provides the exact diagnosis need, e.g. checking into the contrast reversal, limited cycles, chaos etc., by changing the (habituation) threshold value θ_i, when the N-shaped sigmoidal function [1,2,6] is used to generate chaotic ANN (CNN). We postulate that such a chaotic set of innumerable realization forms a fuzzy membership function (FMF) exemplified by fuzzy feature maps of eyes, nose, etc., for the invariant face classification. In addition, CNN has an interesting technology application. The image compression block-artifact can be quickly softened with a negative threshold value, viewed as a reward for higher firing rates.

The retrieval of digital imagery is an active area of research in computer science. Several products that provide information based on the stored contents are already available in the market place. However, most of the proposed methods emphasize on the points that are far from human intuition or emotion. This paper aims to explore image retrieval based on human intuition. To achieve this goal, we have developed a system that combines content based image retrieval technique with interactive genetic algorithm. Especially wavelet transform is used to extract features from images. Experimental results with subjective test and statistical t-test show the usefulness of this approach.

To use ATM network resources effectively, call control in ATM networks must be able to adapt gracefully to the dynamic behavior of traffic loads and network conditions during the network operation. In this paper, we implement a call admission control with interval arithmetic Coulomb energy network, a generalization of Coulomb energy network, which can handle interval data and point data for the learning efficiency. As a modular neural network, each single layer of interval arithmetic Coulomb energy network is trained sequentially to make a multi-layered network. Because each layer can be trained separately, the error does not propagate through the layers and the network becomes modular. Simulation results demonstrate the advantage of modularity of Coulomb energy network and flexibility of interval arithmetic algorithm.

The following sections are included:

• About the Authors

• Keyword Index