To optimize output streams in mechanical waste treatment plants dynamic grain size control is a promising approach. In addition to relevant actuators - such as a adjustable shredder gap width - this also requires technology for real time measurement of the grain size distribution, which is the focus of this work. For this purpose, an approach which classifies individual particles based on information extracted from a two-dimensional image into categories of their real grain size according to the drum screen is followed. Therefore, samples of shredded, solid commercial waste were assigned to defined grain classes using the screen cuts 80mm, 60mm, 40mm, 20mm and 10mm. Then the fractions lightweight materials, metals, plastics, wood, paper-paperboard-cardboard-packaging were sorted out in order to be able to consider each material class separately. Photos of the individual particles were taken using RGB cameras. These were used as the basis for the calculation of intensive (for example: roundness, elliptical Fourier descriptors) as well as extensive (for example: area, diameter, circumference) descriptors. Univariate and multivariate regression methods were used to find correlations among descriptors themselves or to the respective sieve class and to find redundant information within the descriptors. Ultimately, a model that predicts the grain class according to a drum screen using the multivariate regression method Partial Least Squares Regression was chosen. The base for the model are particle describing descriptors which are extracted from a two-dimensional image file. For example in the case of the wood fraction an assignment reliability of over 85% was achieved.