| Summary: | 碩士 === 國立成功大學 === 工程科學系碩博士班 === 98 === The task of image retrieval is to extract matched images from a large image dataset according to the user query. In this work, we aim at developing a landmark image retrieval scheme which supports multi-ROI queries. Specifically, a user can provide more than one example images and then optionally define a region of interest (ROI) within each image to form his or her query. Consequently, images which are more similar to those ROIs can be returned in the proposed scheme. On the other hand, although landmarks are usually easily recognizable by human beings, images containing landmarks tend to with variant illumination, view angles, scaling and so on. To ease this difficulty, we adopt the SIFT descriptor due to its robust invariance to common image transforms. Nevertheless, the SIFT descriptor is with a high computation complexity since a mass of keypoints which are high dimensional vectors can usually be generated from a single image. To improve the efficiency of our retrieval scheme, we propose to properly utilize the indexing structure and corresponding techniques. Empirical studies show that our scheme can handle the image retrieval task both efficiently and effectively.
|
|---|
