Existing pattern recognition approaches can be roughly divided into two categories. The first ones (scientific methods) are very generic and methodologically advanced, but they feature limited robustness in real-world environments and, therefore, cannot be applied in concrete application domains without significant and expensive modifications. The second ones (industrial systems) solve very concrete problems from concrete application domains taking into consideration problem-specific features and, therefore, cannot be applied to other pattern recognition tasks without significant and expensive re-engineering.
In this talk the concept of pattern recognition approaches will be presented that, on the one hand, remain generic and methodologically universal, but on the other hand, make use of domain-specific knowledge modelled as ontologies and can adapt to specific application domains. It will be explained for Marcin Grzegorzek`s current research fields, namely cognitive object recognition, semantic multimedia retrieval, and ontology-based image processing, as his developments in this areas share exactly these two commonalities. Firstly, they show how generic pattern recognition systems can benefit from context information (domain knowledge), if it is available. Secondly, they adaptively optimise the data representation in runtime in terms of its discriminative properties.