On the design of reliable graph matching techniques

On the design of reliable graph matching techniques for change detection Sidharta Gautama, Werner Goeman, Johan D’Haeyer Department for Telecommunication and Information Processing St. Pietersnieuwstraat 41, B-9000 Gent sidharta. gautama@ugent. be Change detection in very-high-resolution images A major challenge in the production and use of geographic information is assessment and control of the quality of the spatial data. Content providers face the problem of continuously ensuring that the information they produce is reliable, accurate and up-to-date. The main issue is the consistency of the data with respect to the current "real-world" situation. Today the industry still relies on human operators, who collect and interpret data to check and correct the current state of the database. This is a very labour -intensive and costly process. In addition, human processing is a source of error and inconsistency. Automated detection of change and anomalies in the existing databases using image information can form an essential tool to support quality control and maintenance of spatial information. GIS vector data VHR imagery The proposed system for change detection is based on object based processing, aimed at deriving information from very-high-resolution images. The system compares detected object features in the image to corresponding features in the vector data. The system consists out of two stages: 1) a lowlevel feature detection process, which extracts image information in terms of lines, regions and junctions, and 2) a high-level matching process, which uses error-tolerant graph matching to find correspondences between the detected image information and external vector data. The graph matching process is driven by the spatial relations between the object features and takes into account different errors that can occur. The matched features can be used to calculate a rubbersheeting transformation between image and vector data. Additionally the object-to-object mapping is useful to define measures of change between datasets. spatial registration and change detection OK NOT report containing local measures of change detection of image objects System overview for automated registration and change detection using object description. Reliably comparing object descriptions In solving the matching problem which compares the image description with the GIS vector data, one should allow tolerance for imprecision and inconsistencies. Errors can occur on the location and shape of object features due to inaccurate detection, differences in spatial resolution of the data and data inconsistencies. In addition, false positives can be present in both datasets. Paramount in this work is the reliability of the system. The relevance of the reporting should be high, meaning that the correct region should be found and if change between the image and vector region is reported, it should reflect the real situation. This means we should be able to clearly define the meaning of relevant versus irrelevant results and the tolerance margins for inaccuracy. reference graph scene graph Using graph matching to compare different descriptions. Object features and their spatial relations are represented by an attributed graph. In this work, we concentrate on the correct modeling of a problem. If a problem is badly modeled, it will always lead to the wrong solution, regardless of the optimization technique that is used. The problem lies in a good definition of the compatibility coe�cients rij(λ, λ’) since they determine the quadratic problem of graph matching: To each object i a probability distribution pi(λ), λ Ωλ is associated expressing that object i has label λ. The value of the compatibility coe�cient of a constraint is used in two ways. Firstly, it encodes the level of violation of the constraint, i. e. more negative values mean a higher violation. Secondly, it encodes the relative ordering of the constraints. Some constraints are more important than others, which is reflected in the value of their compatibility coe�cient. Setting proper values to these coe�cients is not straightforward. For some constraints like the null assignment, it is however di�cult to determine a correct value. Since each object is a priori a possible null object, every assignment is consistent with the null assignment. The problem is to assess the relative importance of the null assignment with respect to the other constraints. It should be avoided that the null solution is the most consistent solution of the system. On the other hand, false correspondences of spurious points should be less consistent than the null assignment. We have derived a condition which controls the optimal solution of a graph matching problem. The condition relates the expected graph label error (quantified by f 0 and f-) with the compatibility coefficients assigned to the null solution and incompatible constraints (resp. w 0 and w-). > Illustration of a query q 1 and q 2 with a pattern pi in the database. The number of null labels is an important indicator whether or not the query result is relevant or not. . Results of a point matching problem with added spatial noise. Optimal measured w 0 for null label assignment compared to predicted optimal w 0, based on the distribution of the graph label error of relevant and irrelevant solutions.