Optimal feature subset selection using differential evolution with Sequential Extreme Learning Machine for river ice images

Feature selection problem often occurs in pattern recognition and more specifically in classification. Feature set extracted from feature extraction methods could contain a large number of feature set. In this work, the features are extracted from Gray Level Co-occurrence Matrix (GLCM) in four different angles (0°, 45°, 90° and 135°) and feature subset selection is performed with Differential Evolution Feature Selection (DEFS) algorithm. In this paper, Sequential Extreme Learning Machine (SELM) will understand the input data one-by-one or portion-by-portion (block of data) with varying or fixed size is integrated with DEFS method. SELM-DEFS algorithm works for single hidden layer feed forward networks (SLFNs) with radial basis function (RBF) for hidden nodes. In SELM, the parameters of hidden nodes are arbitrarily selected and the output weights are analytically determined based on one after the other arriving data. Other than selecting the number of hidden nodes, no other parameters have to be manually selected. SELM-DEFS technique selects optimal feature subset from original feature set. The selected feature set will simplify the training data needed for the classifier. Features selected from the proposed method provide 97.78% accuracy for river ice images.