計算幾何を用いた高品質イメージ検索システムに関する研究

点击翻译按钮获取中文摘要

The objective of this research is to improve the quality of image retrieval in a real-world application to be as high as successful a text retrieval method. We focus on applying shape feature of main objects, which are extracted from a query image, for identifying similarity among images. Two algorithms are proposed. The first one is for improving the time complexity when applying the shape similarity measure. The second algorithm is for improving the quality of the image segmentation when using base-monotone regions. Also to allow the system to be able to automatically locate the important objects.In this research, the shape similarity measure called Modified Hausdorff Distance is applied. Given two set of boundary points P and Q, to compare the two shapes using the Modified Hausdorff Distance, one shape needs to be aligned on the other. The Modified Hausdorff Distance is the average distance of the closest points between the points on the two shape boundary. To obtain an optimal simi … More larity measure, the shapes must be aligned to the most similar part of each other. In a naive method, all pairs of points are applied for finding the optimal transformation. Therefore, the time complexity is cubic in the size of the boundary points.Instead of applying all possible transformation, we proposed a method which applies a pair of correspondence points for mapping the two shapes to the similar part of each other. We also proposed a shape descriptor called a Local Distance Interior Ratio (LDIR) for describing the shape between a feature point and every other boundary points. A pair of points such that the LDIR are similar is called a correspondence. By using the correspondence points, the time complexity for computing the Modified Hausdorff Distance is improved. Moreover, the quality of the retrieved result is as good as applying the naive method.To deal with the large size of image database, it is important for the system to be able to locate and extract the shape contour of the important objects in an image automatically. We proposed a semi-automatic image segmentation algorithm. We employ an algorithm called a room-edge region for removing background region. In order to segment an image containing multiple objects, it can be segmented by decomposing the given pixel grid into small subgrids and apply the room-edge region for each subgrids. One limitation is'the quality of the segmented result depends on decomposition of the subgrids.We present two algorithms for decomposing an image optimally. The first one is called a quadtree decomposition, which an image is optimally decomposed using the quadtree structure. The second on is called an optimal baseline location, which optimally placed a partition lines.In the future, we plan to apply machine learning methods and other features such as color for improving the quality of the retrieved result. Moreover, the label attached to the image is taken into consideration in order to widening the scope of the search. Less