Study Of Spatial Scene Simialrity Measurement Model Based On Feature Matrix And The Relaxation Of Constraint Index

Similarity measurement is the key of geography, and widely applies to spatial retrieval, spatial information integration and spatial data mining. On the basis of content based image research, we study the similarity measurement of spatial scenes, and focus on some critical technologies, such as directional relations, topological relations, similarity measurement on spatial objects and spatial scenes, weights setting, the specific topics and the results of this thesis are as follows:(1) Mathematic Representation of Rectangle Algebra. In process of spatial calculation, we often use the MBR of spatial object, which makes rectangular constraint a key subset of spatial relationship. To reasoning about the rectangle algebra-the most expressive model that describes the constraint between rectangles, we propose using Feature-matrix to characterize RA relations. First, we characterize the 13 interval relations of Interval Algebra by Feature-value-tuple, to describe the constraints between projected intervals of rectangles, then we construct Feature-matrix with Feature-value-tuple pair to describe the rectangular constraints. Underlying properties, such as convex, pre-convex, transitive closure etc. are defined and illustrated.(2) The distance between rectangular constraints. In the research, based on the neighbor grid for rectangular algebra, we calculate the distance of rectangular constraints via shortest path between corresponding vertexes in the grid, and then, analyze how a rectangular constraint turn into another one due to the deformation of rectangles, such as scaling and translation. We use the Cartesian products of feature value tuple intervals to describe the new created feature matrixes during the deformation, and finally make a conclusion of the characteristics for the changes of feature matrix when rectangle deformation occur.(3) Matching of Multi-holed region. Multi-holed region (region with an arbitrary number of holes), mainly represents geographic objects having more than one interior boundaries, such as areas that contain a few lakes, or lakes with islands. To realize the matching between these spatial objects, we proposed a model of similarity measurement on multi-holed regions. In this model, we view multi-holed region as a micro-spatial-scene, where holes and direction between holes playing roles of spatial objects and spatial relations respectively. We use Fourier descriptor to describe shape of a hole, and extract Feature-matrix of direction to represent the directional relations between holes. Based on the assumption of micro-spatial-scene, we process similarity measurement as a constraint satisfaction problem (CSP).(4) Similarity of Region Topological Relations Based on Boundary contacts. Upon the common topological relations, it is difficult to distinguish different region compositions of which complicated boundary contacts occur. To implement similarity measurement on such region compositions, we propose a measuring model based on the boundary contact recording method. The model is structured by preliminary matching step and exact matching step. First, we recognize and filter dissimilar candidates comparing to the reference; meanwhile, obtain corresponding relations of regions between potential candidates and the reference. Then, we encode boundary contact records to binary sequence, and adopt sequence alignment method, an approach from bioinformatics to fulfill topological similarity measurement for two region compositions.(5) The Matching Method of Spatial Scenes Based on Relaxation Labeling Approach. In geo-database, since the object quantity and spatial relations between spatial entities are hardly equal, the result may be empty if we perform the precise matching. By considering the scale difference and study the spatial semantics of spatial scenes, we build a formalized description model and construct the feature matrix for multi-scale spatial scenes. After establishing the initial probability matrix for spatial scenes, we literately update the matrix by relaxation labeling approach until it convergences to a global minimum value. Then we determine the matched objects and calculate similarity of spatial scenes.(6) Weight setting based on Multi-level feedback of spatial scene matching. It is complicate to fulfill a precise spatial scene matching in geo-database, because a) spatial scenes interpretation has more uncertainties than text interpretation b) query modification in spatial scene retrieval is harder than that in text retrieval. Different person have different perception, e.g. for a spatial scene, one may focus on the shapes of objects, while others may concentrate on spatial relations. To introduce human perception into spatial scene retrieval, we perform relevance feedback for spatial scene matching, and update weights according to the user feedback automatically, as a result, the retrieval become more close to user requirement.