Proc. in Springer Verlag P. Perner, Different Learning Strategies in a Case-Based Reasoning System for Image Interpretation, Advances in Case-Based Reasoning, B. Smith and P. Cunningham (Eds.), LNAI 1488, Springer Verlag 1998, S. 251-261. Different Learning Strategies in a Case-Based Reasoning System for Image Interpretation Petra Perner Institute of Computer Vision and Applied Computer Sciences Arno-Nitzsche-Str. 45, 04277 Leipzig Email: ibaiperner@aol.com Abstract. In our previous work, we introduced the basic structure of a case- based reasoning system for image interpretation, a structural similarity measure, and some fundamental learning techniques. In this paper, we describe more so- phisticated learning techniques that are different in abstraction level. We evalu- ate our method on a set of images from the non-destructive testing domain and show the feasibility of the approach. As result, we can show that conventional image processing methods combined with machine learning techniques form a powerful tool for image interpretation. Keywords: CBR Learning, Incremental Learning, Image Interpretation 1 Introduction We introduced in [1] the basic structure of a case-based reasoning system for image interpretation and a structural similarity measure. The fundamental incremental learn- ing method for such a CBR system was introduced in [2]. In this paper, we describe some improved incremental learning capabilities of a CBR system. In Section 2, we give a brief introduction to CBR and review some fundamental topics introduced in [2][3]. The cases are represented as attributed image graphs where the nodes are the objects in the image, the node attributes are the attributes of the objects, and the arcs are the spatial relations between the object. Such kinds of representations are widely used in image interpretation systems like e.g. scene analysis and interpretation of technical drawings. The main problem with such kind of applications is that it is hard to build a vision model because of the complexity of the domain [5]. However, often a large number of cases are available, which favors the use of a CBR image interpretation system equipped with proper learning tech- niques. Section 3 describes the learning strategies in the CBR system, which were already introduced in [2][4]. Different learning strategies are necessary to get automatically a more compact and error-tolerant representation of the case base. The used learning strategies are different in abstraction ability. The simplest one is just learning of new