Imgeun Lee

Lip localization based on active shape model and gaussian mixture model

This paper describes an efficient method for locating lip. Lip deformation is modeled by a statistically deformable model based on Active Shape Model(ASM). In ASM based methods, it is assumed that a training set forms a cluster in shape parameter space. However if there are some clusters in shape parameter space due to an incorrect position of landmark point, ASM may not be able to locate new examples accurately. In this paper, Gaussian mixture is used to characterize the distribution of shape parameter. The Expectation Maximization algorithm is used to determine the maximum likelihood parameters of Gaussian mixture. During search, we resolved the updated locations by projecting a shape into the shape parameter space by using Gaussian mixture. The experiment was performed on many images, and showed very encouraging result.

Another variant of robust fuzzy PCA with initial membership estimation

Principal component analysis (PCA) is a well-known method for dimensionality reduction and feature extraction. PCA has been applied in many areas successfully, however, one of its problems is noise sensitivity due to the use of sum-square-error. Several variants of PCA have been proposed to resolve the problem and, among the variants, improved robust fuzzy PCA (RF-PCA2) demonstrated promising results. RF-PCA2, however, still can be affected by noise due to equal initial membership values for all data points. The fact that RF-PCA2 is still based on sum-square-error is another reason for noise sensitivity. In this paper, a variant of RF-PCA2 called RF-PCA3 is proposed. The proposed algorithm modifies the objective function of RF-PCA2 to allow some increase of sum-square-error and calculates initial membership values using data distribution. RF-PCA3 outperforms RF-PCA2, which is supported by experimental results.

Luminance Stabilization of Image Sequence

Due to light condition or shadow around camera, acquired image sequence is often degraded by intensity fluctuation. This artifact is called luminance flicker. As the luminance flicker corrupts the performance of motion estimation or object detection, it should be corrected before further processing. In this paper, we analyze the flicker generation model and propose the new algorithm for flicker reduction. The proposed algorithm considers gain and offset parameter separately, and stabilizes the luminance fluctuation based on these parameters. We show the performance of the proposed method by testing on the sequence with artificially added luminance flicker and real sequence with object motion.

A Hybrid Genetic Algorithm with Simulated Annealing for Nonlinear Blind Equalization Using RBF Networks

In this study, a hybrid genetic algorithm, which merges a genetic algorithm with simulated annealing, is derived for nonlinear channel blind equalization using RBF networks. The proposed hybrid genetic algorithm is used to estimate the output states of a nonlinear channel, based on the Bayesian likelihood fitness function, instead of the channel parameters. From these estimated output states, the desired channel states of the nonlinear channel are derived and placed at the center of a RBF equalizer to reconstruct transmitted symbols. In the simulations, binary signals are generated at random with Gaussian noise. The performance of the proposed method is compared with those of a conventional genetic algorithm (GA) and a simplex GA. It is shown that the relatively high accuracy and fast convergence speed have been achieved.

A New Hybrid Genetic Algorithm for Nonlinear Channel Blind Equalization

In this study, a hybrid genetic algorithm merged with simulated annealing is presented to solve nonlinear channel blind equalization problems. The equalization of nonlinear channels is more complicated one, but it is of more practical use in real world environments. The proposed hybrid genetic algorithm with simulated annealing is used to estimate the output states of nonlinear channel, based on the Bayesian likelihood fitness function, instead of the channel parameters. By using the desired channel states derived from these estimated output states of the nonlinear channel, the Bayesian equalizer is implemented to reconstruct transmitted symbols. In the simulations, binary signals are generated at random with Gaussian noise. The performance of the proposed method is compared with those of a conventional genetic algorithm(GA) and a simplex GA. In particular, we observe a relatively high accuracy and fast convergence of the method.