The segmentation of a volumetric dataset, as they are commonly generated in computed tomography, into disjoint semantically understandable parts poses an ongoing problem. This is especially the case in the industrial domain of image processing for computed tomography, where the main problems lie in the wide range of applications together with the size of the generated datasets (up to Terabytes for a single scan). Both problems severely prohibit the processing of such data with established purely data-driven methods. This work proposes the OntoSeg system, which consists of a combination of semantic knowledge (in form of an ontology) with image processing methods. The latter are implemented in individual libraries, which may be attributed to any semantic component contained in the dataset. Thus, each library can be exchanged or enhanced flexibly and can incorporate arbitrary additional knowledge, which is hardly ever possible in classical monolithical image analysis systems. The proposed system incrementally restricts the dataset to the relevant regions, which allows for the segmentation of the final components with existing methods. An evaluation on a small ontology specific to the structure of cars enabled the efficient localization and segmentation of certain components, without having to segment every single of its billions of voxels.