| Classic content-based image retrieval (CBIR) takes a single non-annotated query image, and retrieves similar images from an image repository. Such a search must rely upon a holistic (or global) view of the image. Yet often the desired content of an image is not holistic, but is localized. We define localized content-based image retrieval as a CBIR task where the user is only interested in a portion of the image, and the rest of the image is irrelevant. We presents three new localized CBIR systems, each of which deals with a different task and retrieval scenario.;Image representation is a critical element of localized CBIR. A sufficient image representation must capture the content of complex objects of interest which are composed of many parts. It must be able to do so in a way that allows it to distinguish the object of interest from the background in which it occurs. We present two new image representations, one based on segmentation and one based on salient points, as necessary complements to our proposed CBIR systems.;Our first localized CBIR system uses labeled images in conjunction with a new multiple instance (MI) learning algorithm to rank images in the database using a similarity measure that is based upon individual regions within the image. This CBIR system is able to retrieve more robust sets of images and achieve better performance than previous system on our datasets.;Our second localized CBIR system, presents a different learning approach that combines MI learning with semi-supervised learning (SSL), where the image repository is available during training. CBIR tasks can often benefit from SSL. Our system is the first to combine MI with SSL and uses a novel new approach.;Finally, in our third system, we address the task of searching within an image to extract a desired object from its background in complex scenes. We present an object extraction method that uses image search and requires no human-segmented training images or object model. Our system is unique in that it can be applied in scenarios where very few training images are available. |
