Operational feature selection in gaussian mixture models
Adrien Lagrange · HAL (Le Centre pour la Communication Scientifique Directe) · 2016
This report presents a forward feature selection algorithm based on Gaussian mixture model (GMM) classifiers. The algorithm selects iteratively features that maximize a criterion function which can be either a classification rate or a measure of divergence. We explore several variations of this algorithm by changing the criterion function and also by testing a floating forward variation allowing backward step to discard already selected features. An important effort is made in exploiting GMM properties to implement a fast algorithm. In particular, update rules of the GMM model are used to compute the criterion function with various sets of features. The result is a C++ remote module for the remote sensing processing toolbox Orfeo (OTB) developed by CNES. Finally, the method is tested and also compared to other classifiers using two different datasets, one of hyperspectral images with a lot of spectral variables and one with heterogeneous spatial features. The results validate the fact that the method performs well in terms of processing time and classification accuracy in comparison to the standard classifiers available in OTB.