Polynomial feature engineering for classification of textural images

Abstract

I consider a number of methods of automatic quadratic features adjustment for digital textural images of biological tissues in order to improve the quality of classification. The proposed approaches are based on optimization procedures that use various quality criteria of a feature space as target functions. I investigate the methods based on random search, genetic algorithm, simulation of annealing, as well as the original hybrid algorithm. I presented results of experimental studies of the proposed algorithms on sets of real X-ray images of bone tissue and the lung CT images. We show that the hybrid algorithm provides more stable results regardless of the chosen quality criterion of the feature space, which is expressed in decreasing of the average percentage of incorrectly recognized images in comparison with the use of the specific optimization methods individually.