Facial Feature Land-marking writh Optimized Gabor Parameters based on Maximization of Separation between Features

摘要

Elastic Bunch Graph Matching(EBGM) is a popular method in automatic localization of facial feature points, where the selection of optimal Gabor parameters plays a key role in the extraction of Gabor jets with a high degree of discrimination. We propose a method for the selection of parameters by minimizing the energy function consisting of within-class and between-class scatters using the gradient descendent method. We formulate the learning rule and design the algorithm for the learning of parameters. Numerical experiments have been performed to investigate the performance of estimated parameters in improving the precision of automatic localization.