State-of-the-art image analysis systems often feature limited robustness under varying environment conditions. A significant improvement in this research area can only be achieved based on better understanding of how humans interpret digital images and by modeling this behavior. Moreover, human knowledge about a particular application domain should also be taken into consideration. Integration of the high-level background knowledge modeled as an ontology into the low-level image processing in a cognitive and mathematically consistent way is the only chance to achieve improvements in this research field. The talk will comprehensively discuss this hypothesis. At the beginning, an introduction and an overview of relevant research topics the speaker has been working on over the last few years will be given. Further, three of them will be presented in detail. The first one addresses the problem of appearances-based classification and localization of 3D objects in 2D digital images without using any domain specific knowledge. The second one combines low-level image classification with high-level spatial reasoning. The third one, submitted to the DFG as an Emmy Noether Proposal, proposes a general approach for integration of the background knowledge stored in an ontology into the process of medical image analysis. The talk will end with some final conclusions and general statements about the future research direction of the speaker.