| 英文摘要 | Recently, the successful launch of several commercial optical satellites, such as Ikonos and QuickBird, has greatly enriched the resources of high resolution satellite images, which enables the high clarity, multi-angle and even real time observation of the earth. Consequently, the high resolution satellite images are drawning more and more attenstion from multiple fields, such as the creation and update of GIS data, the cartography, urban planning, disaster monitoring, environment monitoring and digital earch, etc. However, the current development of the corresponding theories and techniques, especially in the field of object extraction from urban images, is too slow to handle the increasingly improved data quality, which has obviously weakened the automation and intelligence of the spacial information collection. Hence, the main work of our thesis is focused on the building extraction from densely built-up urban images, and the major achievements and contributions are summarized as following: 1.Propose a 2D building extraction algorithm based on region analysis.In our algorithm, we first recognize the building regions by SVM and texture features, and then we propose a seed-based region grouping method to estimate the location and cover areas of all buildings, after that, building hypotheses are generated based on these estimated cover areas, and lines are used to refine some error edge locations, finally the shadow and geometry constraints are used to delete the noises. 2.Propose a 2D building extraction algorithm based on edge verification. The whole algorithm follows the hypothesis-verification strategy. The key contribution of this approach is the edge verification method during the hypothesis verification process, which greatly improves the accuracy of the building extraction results from complex scenes. To extract more accurate edges, we construct a probabilistic model and an optimization framework: first, a probability being the optimal edge is given to each possible edge, then the constraints of any two possible edges are estimated based on a machine learning method incorporating the image evidences, finally, these constraints and other prior knowledge are integrated into an optimization problem, by solving which these probabilities can be computed and the optimal edge can be selected. 3.Propose a method to evaluate the regularity of man-made object contours and construct an evaluation system for the consistency of different parts of buildings.The regularity evaluation method constructs four discrete regular edge descriptors, based on which a statistic method is used to describe the regularity of contours. To analyze the consistency of different parts of buildings, we construct an image feature system from the interior, the outer supporting regions and the shadows of target building. This is a completed and reliable building extraction feature model, which is very helpful for objects extraction from complex scenes. |
修改评论