Pattern recognition in the differentiated image for the powder and granulated materials particle size classification

Abstract

Abstract The paper shows and investigates the technique of classifying the particle size of powder and granulated materials. The objects of the study are industrially produced mineral fertilizers. Samples with different composition (about five types of mineral fertilizers) and various degrees of particle size (less than 100 µm, less than 500 µm and granules of 2-5 mm) were examined. The samples particles have an irregular shape, close to spherical (in the case of granules) or cubic (in the case of powders). To improve the accuracy and eliminate the particle shape influence on analysis, the preliminary pressing of samples on a boric acid substrate was used. The keynote of the proposed technique is to obtain an optoelectronic image of an object with a resolution of at least 640x480 pixels (a three-dimensional matrix of pixel intensity in the Red-Green-Blue (RGB) system). Next, the area of analysis is separated from the obtained image and transformed into grayscale (a two-dimensional matrix of pixel intensities with a resolution of at least 200x200 pixels). The influence of external illumination (gradient, temperature and brightness) is eliminated by the grayscale image differentiation. The result is the “surface map” of the sample, which reflects defects in the pressed structure (patterns, which are responsible for the size of the particles). According to the found patterns, the samples are classified according to their particle size. Four classification algorithms were investigated (linear, linear with L1 and L2 regularization, and a nonlinear “random forest”). All proposed approaches are automated and implemented in the Python 3.6 programming language. There is provided the selection of the operating parameters of all the described algorithms.