A setup for online measurement of crystallization processes was developed, where particle images were taken and in combination with artificial neural networks (ANNs) were used to convert the particle size distribution into populations of single crystals and agglomerates. For generation of an ANN that discriminates all particle sizes equally well, the training set composition was investigated while proportional similarity was applied for image descriptor selection. Adipic acid/water served as a model system. The effect of the measurement setup on experiments performed for normal cooling crystallization was evaluated. It was found that the major challenge of measuring crystal agglomerates lies in superimposing aggregates that falsify the online measurement results, as offline measurements demonstrated.