An Effective Similar Object Searching System for Complex Satellite Images

초록

As the amount of satellite images in the database increases, the demand for searching similar satellite images also increases. The content-based image retrieval (CBIR) is one of approaches for finding images containing similar objects from an image database. To apply CBIR to satellite images, image segmentation is necessary to separate the shape of an object from an image. However, due to several properties of satellite images, image segmentation is very difficult. Therefore, it is difficult to get a good result for satellite images by using CBIR. In this paper, we propose a new approach which doesn't require image segmentation to search similar images. We use the SIFT keypoint descriptor, which doesn't require image segmentation, as a shape descriptor. However, the basic SIFT matching needs to handle a large number of keypoint descriptors. In the proposed approach, we set a pruning circle to find similar objects from an image. Then, we prune some SIFT keypoint descriptors which are not located on similar objects. A similarity is assigned to an image based on the remaining SIFT keypoint descriptors. We also propose a way to use an index for scalability of the proposed approach. Experimental results show that the proposed approach searches images containing a similar object effectively in the case of satellite images.