INTRODUCTION
THE GAUSSIAN mixture models (GMM) classifier has gained increasing attention in the pattern recognition community. GMM can be classified as a semi-parametric density estimation method since it defines a very general class of functional forms for the density model. In this mixture model, a probability density function is expressed as a linear combination of basis functions. Improved classification performances have been demonstrated in many pattern recognition applications . Performance figures of more than 95% have already been reported for applications such as electronic nose and gas identification. Another interesting property of GMM is that the training procedure is done independently for each class in turn by constructing a Gaussian mixture of a given class. Adding a new class to a classification problem does not require retraining the whole system and does not affect the topology of the classifier making it attractive for pattern recognition applications. While GMM provides very good performances and interesting properties as a classifier, it presents some problems that may limit its practical use in real-time applications. One problem is that GMM can require large amounts of memory to store various coefficients and can require complex computations mainly involving exponential calculations. Thus, this scheme can be put to efficient practical use only if good hardware implementation strategies are developed.
In this project, we propose an efficient digital VLSI implementation that we believe can meet the computational requirement of GMM-based classifiers. First, after analyzing the complexity of the GMM classifier it was found that the vector-matrix multiplication and the exponential calculations are the most critical operations in the classifier. A good tradeoff between real-time processing and hardware resources requirements is obtained using a serial-parallel architecture and an efficient pipelining strategy. Second, a linear piecewise function (LPF) is proposed to replace the exponential calculation. Implementing LPF-based GMM, also permits to avoid the need for using area consuming look-up table (generally used in digital implementation) to implement the exponential function. The effect of both limited precision and the mixture models approximation using LPF on the classification performance is investigated using seven different data-sets. These data-sets are also used to compare the performance of GMM with other benchmark classifiers.
THE GAUSSIAN mixture models (GMM) classifier has gained increasing attention in the pattern recognition community. GMM can be classified as a semi-parametric density estimation method since it defines a very general class of functional forms for the density model. In this mixture model, a probability density function is expressed as a linear combination of basis functions. Improved classification performances have been demonstrated in many pattern recognition applications . Performance figures of more than 95% have already been reported for applications such as electronic nose and gas identification. Another interesting property of GMM is that the training procedure is done independently for each class in turn by constructing a Gaussian mixture of a given class. Adding a new class to a classification problem does not require retraining the whole system and does not affect the topology of the classifier making it attractive for pattern recognition applications. While GMM provides very good performances and interesting properties as a classifier, it presents some problems that may limit its practical use in real-time applications. One problem is that GMM can require large amounts of memory to store various coefficients and can require complex computations mainly involving exponential calculations. Thus, this scheme can be put to efficient practical use only if good hardware implementation strategies are developed.
In this project, we propose an efficient digital VLSI implementation that we believe can meet the computational requirement of GMM-based classifiers. First, after analyzing the complexity of the GMM classifier it was found that the vector-matrix multiplication and the exponential calculations are the most critical operations in the classifier. A good tradeoff between real-time processing and hardware resources requirements is obtained using a serial-parallel architecture and an efficient pipelining strategy. Second, a linear piecewise function (LPF) is proposed to replace the exponential calculation. Implementing LPF-based GMM, also permits to avoid the need for using area consuming look-up table (generally used in digital implementation) to implement the exponential function. The effect of both limited precision and the mixture models approximation using LPF on the classification performance is investigated using seven different data-sets. These data-sets are also used to compare the performance of GMM with other benchmark classifiers.
No comments:
Post a Comment