| Popis: |
The content of an image is often associated with the main object(s) present in an image. Therefore, for effective content-based retrieval, the database images need to be indexed by features extracted from the object of interest, ignoring any irrelevant image background. In this work, we propose content-based retrieval strategies focusing on the use of color-based features for specialized image domains where the performance of general-purpose color image retrieval techniques is poor. The retrieval performance is improved by taking the special characteristics of the domain into account to extract the object of interest when possible, or capture the properties of the important objects present in an image when it is not possible to extract an object of interest a priori. Three test domains are selected which have very different characteristics requiring different retrieval strategies. These domains are representative of a larger class of specialized image databases which have similar characteristics. A two-phase image retrieval engine which is robust in the presence of interfering backgrounds and large variations in the size of the query object in the target images, is proposed for an advertisement images domain where there are extreme variations in backgrounds and the size of the object of interest. An iterative segmentation algorithm for extracting the object of interest is proposed when there is useful domain knowledge available about the subject of the database images, as in the flower images domain tested in this dissertation work. Automatic segmentation of the object of interest is extended to a database of bird images where there is no subject-specific domain knowledge available, using general observations true for any image where the object of interest is prominent in the image. |
