| Popis: |
This PhD thesis of the Delft University of Technology in 1995 describes various aspects of model-based image processing. One aspect is a model-based interpretation system for two types of maps, presented in chapter 5. There, and in chapter 1, a model-based interpretation methodology is proposed which consists of three different models. The model of the interpretation flow describes the flow of the interpretation process, the functionality of the modules, and their connections. The model of the map describes map specific information: which objects can be found, what are the relations between the various objects, where can they be found, what can go wrong, and how to correct errors. The third model, the model of the object, describes how the object can be found. This methodology is used for interpreting two types of maps. The approach is reasonably flexible, as can be seen from the transition from an interpretation system for cadastral maps, to an interpretation system for large-scale base maps. The basic interpretation flow and some modules remain the same, a few modules have to be removed and some others have to be added, but it is not necessary to redo the whole interpretation system. The performance of the model-based cadastral map interpretation system is compared to a bottom-up interpretation system, described in chapter 2. The first performs better than the latter, and that is caused by the model knowledge which allows us to correct for the errors made. The bottom-up system was the base line system for investigating a map interpretation system. It also served to investigate the vectorization problem. It appears that during vectorization, problems may arise because some lines are broken or because too many nodes and edges are generated. This was not only caused by the vectorization algorithm used (the Douglas-Peucker algorithm), but also by the preprocessing necessary to obtain one pixel thick lines, or by distortions already present in the (paper) input image. In the base line system, rules were devised to correct for this, and indeed the number of nodes and edges decreased considerably, while the error only increased slightly. Although the rules worked well, it was not a satisfactory solution because a slight increase in error is not always acceptable. Therefore, in chapter 4 an adaptive vectorization algorithm was developed which used the input line drawing to improve and to correct the vectorization. This adaptive vectorization algorithm performed better than the standard Douglas-Peucker algorithm, and it also performed better than the Douglas-Peucker algorithm with the 'clean up' rules from chapter 2. This adaptive vectorization algorithm is the second aspect in model-based image processing in this thesis. It was called 'adaptive' to avoid confusion with the model-based notion in chapters 1 and 5. In chapter 3 a system to compile mosaics from separately scanned line drawings is presented. This system automatically finds matching edges, and from these corresponding points in two separately scanned line drawings. These corresponding points are used in a geometric transformation to transform the coordinate system from one line drawing to the other. Thus, the model used is on a low level. Such a system can be used as a submodule before the actual interpretation algorithm: it permits us to present a larger line drawing to the algorithm instead of several smaller ones, and thus avoid boundary problems. For instance, an interpretation system for cadastral maps cannot find parcels on the boundary of the image, because objects necessary for the interpretation (e.g. parcel numbers) may be on the other -invisible or unknown- side of the boundary. The notion 'model-based' is often only used for high level systems. However, we believe that the principle of model-based (expectations of reality used to guide the interpretation, and if necessary, to correct it) can be found on all levels, as illustrated in this thesis. |
