FEATURE-SELECTION BY A GENETIC ALGORITHM - APPLICATION TO SEED DISCRIMINATION BY ARTIFICIAL VISION

Genetic algorithms (GAs) are efficient search methods based on the par adigm of natural selection and population genetics. A simple GA was ap plied for selecting the optimal feature subset among an initial featur e set of larger size. The performances were tested on a practical patt ern recognition problem, which consisted on the discrimination between four seed species (two cultivated and two adventitious seed species) by artificial vision. A set of 73 features, describing size, shape and texture, were extracted from colour images in order to characterise e ach seed. The goal of the GA was to select the best subset of features which gave the highest classification rates when using the nearest ne ighbour as a classification method. The selected features were represe nted by binary chromosomes which had 73 elements. The number of select ed features was directly related to the probability of initialisation of the population at the first generation of the GA. When this probabi lity was fixed to 0.1, the GA selected about five features. The classi fication performances increased with the number of generations. For ex ample, 6.25% of the seeds were misclassified by using five features at generation 140, whereas another subset of the same size led to 3% mis classification at generation 400. The present work shows the great pot ential of GAs for feature selection (dimensionality reduction) problem s. (C) 1998 SCI.