Dante Augusto Couto Barone

Graphical Models and Point Pattern Matching

This paper describes a novel solution to the rigid point pattern matching problem in Euclidean spaces of any dimension. Although we assume rigid motion, jitter is allowed. We present a noniterative, polynomial time algorithm that is guaranteed to find an optimal solution for the noiseless case. First, we model point pattern matching as a weighted graph matching problem, where weights correspond to Euclidean distances between nodes. We then formulate graph matching as a problem of finding a maximum probability configuration in a graphical model. By using graph rigidity arguments, we prove that a sparse graphical model yields equivalent results to the fully connected model in the noiseless case. This allows us to obtain an algorithm that runs in polynomial time and is provably optimal for exact matching between noiseless point sets. For inexact matching, we can still apply the same algorithm to find approximately optimal solutions. Experimental results obtained by our approach show improvements in accuracy over current methods, particularly when matching patterns of different sizes

A probabilistic model for the human skin color

We present a multivariate statistical model to represent the human skin color. There are no limitations regarding whether the person is white or black, once the model is able to learn automatically the ethnicity of the person involved. We propose to model the skin color in the chromatic subspace, which is by default normalized with respect to illumination. First, skin samples from both white and black people are collected. These samples are then used to estimate a parametric statistical model, which consists of a mixture of Gaussian probability density functions (pdfs). Estimation is performed by a learning process based on the expectation-maximization (EM) algorithm. Experiments are carried out and receiver operating characteristics (ROC curves) are obtained to analyse the performance of the estimated model. The results are compared to those of models that use a single Gaussian density.

Do mixture models in chromaticity space improve skin detection

This note reports an experiment where a single Gaussian model and several Gaussian mixture models were used to model skin color in the rg chromaticity space. By using training and test databases containing millions of skin pixels, we show that mixture models can improve skin detection, but not always. There is a relevant operating region where no performance gain is observed.

Speaker identification using nonlinear dynamical features

Abstract This paper reports results obtained in a speaker identification system which combines commonly used feature parameters, such as LP-derived cepstral coefficients and pitch, with nonlinear dynamic features, namely, fractal dimension, entropy and largest Lyapunov exponent, extracted from every window of 30 ms of speech, applied every 10 ms. The corpus used in the tests is composed of 37 different speakers, and the best results are obtained when the nonlinear dynamic features are included, suggesting that the information added with these features was not present in the others so far.

Performance evaluation of single and multiple-Gaussian models for skin color modeling

We present an experimental setup to evaluate the relative performance of single Gaussian models and Gaussian mixture models for skin color modeling. Firstly, a sample set of 1,120,000 skin pixels from a number of ethnic groups is selected and represented in the chromaticity space. Parameter estimation for both the single Gaussian and seven (with 2 to 8 Gaussian components) Gaussian mixture models is performed. For the mixture models, learning is carried out via the expectation-maximisation (EM) algorithm. In order to compare performances achieved by the 8 different models, we apply to each model a test set of 800 images - none from the training set. True skin regions, representing ground truth, are manually selected, and false positive and true positive rates are computed for each value of a specific threshold. Finally, receiver operating characteristics (ROC) curves are plotted for each model, making it possible to analyze and compare their relative performances. Results obtained show that, for medium to high true positive rates, mixture models (with 2 to 8 components) outperform the single Gaussian model. Nevertheless, for low false positive rates, all the models behave similarly.

Hardware implementation of RAM neural networks

Abstract This work describes an alternative technique for hardware and software implementation of RAM-based Boolean neural networks, which describes neurons using the VHDL language. An example of application consisting of the classification problem of the British mail scanned address is attended with a RAM architecture presenting (340 x 12)-input neurons. The weights of each neuron are represented by its truth table and described using simple logic gates (AND, OR, and NOT), aiming to make possible the network logic minimisation and its hardware implementation by the ALTERA MAX + PLUS II fast prototyping package (Altera, 1992). The developed software tool allows the specification and training of the network. Then, its VHDL description is generated to be interpreted and minimised by the ALTERA EPLD design system. If it is not necessary to have high-speed processing or if pre-processing phases are needed, the ANN can be implemented in software. The software strategy makes use of the direct translation of the VHDL description into a simplified C language code. Once the user has specified and taught the network, this approach makes possible automatic prototyping of RAM neural networks in software and hardware.

A speaker identification system using a model of artificial neural networks for an elevator application

Abstract This paper presents a comparison of some features for speaker identification applied to a building security system. The features used in this paper are pitch, frequency formants, linear predictive coding (LPC) coefficients and cepstral coefficients computed from LPC. The comparison was based on a system for building security that uses the voice of the residents to control the access to the building. The system uses a model of artificial neural network called multi-layer perceptron (MLP) as a classifier. This paper shows that cepstral coefficients are more efficient than LPC coefficients for the security system.

High speed neural control for robot navigation

This paper addresses the real time control of the Khepera mobile robot navigation in a maze with reflector walls. Boolean neural networks such as RAM and GSN models are applied to drive the vehicle, following a light source, while avoiding obstacles. Both neural networks are implemented with simple logic and arithmetic functions (NOT, AND, OR, Addition, and Comparison), aiming to improve the system speed. The results obtained are compared with two other control strategies: multilayer perceptron and fuzzy logic.

An Optimal Probabilistic Graphical Model for Point Set Matching

We present a probabilistic graphical model for point set matching. By using a result about the redundancy of the pairwise distances in a point set, we represent the binary relations over a simple triangulated graph that retains the same informational content as the complete graph. The maximal clique size of this resultant graph is independent of the point set sizes, what enables us to perform exact inference in polynomial time with a Junction Tree algorithm. The resulting technique is optimal in the Maximum a Posteriori sense. Experiments show that the algorithm significantly outperforms standard probabilistic relaxation labeling.

Fractal dimension applied to speaker identification

Reports the results obtained in a speaker identification system based on Bhattacharrya distance, which combines LP-derived cepstral coefficients, with a nonlinear dynamic feature namely fractal dimension. The nonlinear dynamic analysis starts with the phase space reconstruction, and the fractal dimension of the correspondent attractor trajectory is estimated. This analysis is performed in every speech window, providing a measure of a time-dependent fractal dimension. The corpus used in the tests is composed by 37 different speakers, and the best results are obtained when the fractal dimension is included, suggesting that the information added with this feature was not present so far.