To optimize output streams in mechanical waste treatment plants dynamic particle size control is a promising approach. In addition to relevant actuators – such as an adjustable shredder gap width – this also requires technology for online and real-time measurements of the particle size distribution. The paper at hand presents a model in MATLAB® which extracts information about several geometric descriptors – such as diameters, lengths, areas, shape factors – from 2D images of individual particles taken by RGB cameras of pre-shredded, solid, mixed commercial waste and processes this data in a multivariate regression model using the Partial Least Squares Regression (PLSR) to predict the particle size class of each particle according to a drum screen. The investigated materials in this work are lightweight fraction, plastics, wood, paper-cardboard and residual fraction. The particle sizes are divided into classes defined by the screen cuts (in mm) 80, 60, 40, 20 and 10. The results show assignment reliability for certain materials of over 80%. Furthermore, when considering the results for determining a complete particle size distribution – for an exemplary real waste – the accuracy of the model is as good as 99% for the materials wood, 3D-plastics and residual fraction for each particle size class respectively as assignment errors partially compensate each other.