Patricia A. Nava

Off-line recognition of isolated Persian handwritten characters using multiple hidden Markov models

In this paper a new method for off-line recognition of isolated handwritten Persian characters based on hidden Markov models (HMMs) is proposed. In the proposed system, document images are acquired in 300-dpi resolution. Multiple filters such as median and morphologal filters are utilized for noise removal. The features used in this process are methods based on regional projection contour transformation (RPCT). In this stage, two types of feature vectors, based on this technique, are extracted. The recognition system consists of two stages. For each character in the training phase, multiple HMMs corresponding to different feature vectors are built. In the classification phase, the results of the individual classifiers are integrated to produce the final recognition.

A Comparison of Various Respiratory System Models Based on Parameter Estimates From Impulse Oscillometry Data

Impulse oscillometry offers an advantage over spirometry when conducting pulmonary function tests. Not only does it require minimal patient cooperation, it provides useful data in a form amenable to engineering methods. In particular, the data can be used to obtain parameter estimates for electric circuit-based models of the respiratory system, which can in turn aid the detection and diagnosis of various diseases/pathologies. Of the six models analyzed during this study, the DuBois model and a newly proposed extended RIC model seem to provide the most robust parameter estimates for our entire data set of 106 subjects with various respiratory ailments such as asthma and chronic obstructive pulmonary disease. Such a diagnostic approach, relying on estimated parameter values, may require additional measures to ensure proper identification of diseases/pathologies but the preliminary results are promising.

Neural network algorithms for tuning of fuzzy certainty factor expert systems

This paper presents a description of an algorithm that is used for tuning fuzzy certainty factors in a fuzzy expert system. Many expert system building tools have been developed, but only a few can produce systems that deal with inexact knowledge. One of the models that can deal with inexact knowledge is the fuzzy certainty factor expert system. Given a fuzzy certainty factor expert system, one of the very difficult tasks in its development is the tuning of the fuzzy certainty factors. This paper presents a method for using backpropagation, a well-known neural network training algorithm, for tuning of certainty factors. The fuzzy certainty factor expert system is defined, and then it is tuned by translating or mapping the fuzzy logic system to a feedforward neural network framework. The tuning then takes place as a session analogous to neural network training. This method is shown to be much more efficient than previous tuning methods. The application area of dental diagnostics is used to demonstrate the method. The system is reviewed and results that show its efficacy are discussed.

A fuzzy method for automatic generation of membership function using fuzzy relations from training examples

Fuzzy systems rely on membership functions to represent input values for problem presentation and eventual problem solution. These can be generated in different ways, one of which is obtaining an expert to define the functions. This method is not always cost effective or available, so automatic membership function definition is extremely desirable Many methods for constructing membership functions based on knowledge engineering have been developed. Previous work has shown that statistical methods can be used to generate these membership functions. The quality of the result, however, is very application dependent. This study focuses on a method of automatic membership function generation that relies on the use of fuzzy relations. This paper describes the implementation of one such method, and examines its application to several data sets, including the identification of vowel sounds in spoken English.

A user-friendly integrated software environment for heart rate variability analysis of sleep disordered breathing

We present an integrated software environment that enables sleep specialists to analyze, quantity and diagnose sleep disordered breathing based upon time-domain, frequency-domain, and nonlinear dynamics measurement measures of heart rate variability (HRV) signal. The integrated graphical user interface (GUT) and the signal processing algorithms were developed and implemented in MATLAB, This environment provides the facility to import or read sleep data (ECG, EEG, blood pressure, respiratory, oxygen saturation, etc.) under user control and displays them individually or collectively in a data window for visual inspection. It then enables the user to clip the length of data to be used in carrying out the analysis. Raw ECG data is preprocessed for reliable QRS detection and the HRV signal is derived following the guidelines of the Task Force of the European Society of Cardiology and the North American Society for Pacing and Electrophysiology, which is then displayed along with respiration signal in the analysis window. After this step time-domain, frequency- domain, and nonlinear dynamics analyses of the HRV signal are performed to extract sensitive measures used in detecting and diagnosing sleep disordered breathing. The computer analysis can then generate a complete report for the specialist and further statistical and/or automatic analyses. The system was developed and validated using data from the MIT-BIH Polysomnographic Database. After validation and reliability testing it was used to analyze sleep data for detection of sleep disordered breathing (SDB) in children. Data from normal and children diagnosed with SDB showed that the system could potentially distinguish between normal children and children suffering from sleep disordered breathing.

Interval-based neural networks for soft decisions

The performance of neural networks, when the training data is limited, can be improved by incorporation of interval techniques. These techniques improve performance by introducing the ability to classify imprecise data. Performance can be further improved by incorporating the ability to make soft decisions. Soft decisions differ from hard decisions by allowing the decision-making system the option of deferring to a human. This decision-rejection option has the effect of reducing the error rate of the decision-making system. The paper discusses three distinct techniques for making soft decisions and the performance of the interval-based neural network that utilizes these techniques.

Reconfigurable data path processor: implementation and application for signal processing algorithms

The paper briefly describes reconfigurable architecture classifications and comparisons, as well as an FPGA implementation of a simplified version of the reconfigurable data path processor (RDPP) using Xilinx design tools. The FPGA configured as RDPP can be employed for a wide variety of signal processing applications. That is, the inherent parallel nature of the architecture lends itself to signal processing algorithms, which are discussed. Simulations of operations on the input data stream, which include multiply-and-accumulate (MAC) operations used in filters and matrix manipulation for image processing applications, are presented.

A neural network implementation of dual-frequency output control for multilevel inverters

This paper proposes the use of a multilayer feedforward neural network to control a multilevel inverter-based dual-frequency induction heating power supply, plus the use of distributed computing to train that network. The motivation is because the control function mappings from the desired waveforms two modulation indices to its step-angles are not simple closed-form expressions, so using look-up tables to implement the mappings accurately and comprehensively would require a significant amount of memory. The neural network was first trained using a single central processing unit, and the results were then compared to similar training using distributed computing (with multiple local area network-connected central processing units). It was found that using distributed computing reduced significantly the training time needed to achieve the desired level of accuracy.

Parametric sensitivity analysis of human respiratory impedance

Measurements of the respiratory impedance in humans are a means to help diagnose underdevelopment and/or pathological conditions. Various models of the human respiratory system of varying complexity and fidelity have been developed for study to enable greater specificity of the diagnosis. In this paper, a parametric sensitivity analysis that was performed on the respiratory impedance of four different respiratory system models is described. The results of this analysis can be used to obtain valuable information on how certain respiratory system parameter variations affect the respiratory impedance at different frequencies. Such an analysis can help determine the optimal input excitation characteristics to more accurately extract particular model component values. Then calculated reference values can be more precise and form a more reliable baseline for diagnosing specific physiologic and pathologic changes in the respiratory system. The sensitivities derived for each parameter of the four models, and their plots using normal values for the lung mechanics of low birth weight newborns are provided. An example of the interpretation and usefulness of sensitivity functions is also presented.

Use of intervals for soft classification in fuzzy neural networks

Neural networks can be used to classify input data into one of a given set of categories. With limited training sets, crisp neural network results are predictably poor. Incorporation of fuzzy techniques improves performance in these cases. Even though fuzzy neural networks classify imprecise data quite well, the incorporation of a soft decision classification lowers the error rate substantially. This paper discusses methods for soft decision making, including a method that uses intervals. A neuro-fuzzy system that classifies input vectors is examined. This neuro-fuzzy system not only uses intervals in a fuzzy neural network, but also employs a method of utilizing intervals in a soft decision for classification. This neuro-fuzzy systems performance in computer simulations is examined and compared with crisp neural networks performance.