Automated prediction of protein active sites is essential for large-scale protein function prediction, classification, and druggability estimates. In this work, we present DoGSite, a new structure-based method to predict active sites in proteins based on a Difference of Gaussian (DoG) approach which originates from image processing. In contrast to existing methods, DoGSite splits predicted pockets into subpockets, revealing a refined description of the topology of active sites. DoGSite correctly predicts binding pockets for over 92\% of the PDBBind and the scPDB data set, being in line with the best-performing methods available. In 63\% of the PDBBind data set the detected pockets can be subdivided into smaller subpockets. The cocrystallized ligand is contained in exactly one subpocket in 87\% of the predictions. Furthermore, we introduce a more precise prediction performance measure by taking the pairwise ligand and pocket coverage into account. In 90\% of the cases DoGSite predicts a pocket that contains at least half of the ligand. In 70\% of the cases additionally more than a quarter of the respective pocket itself is covered by the cocrystallized ligand. Consideration of subpockets produces an increase in coverage yielding a success rate of 83\% for the latter measure.