Invariant, Intra-Class Retrieval in Digital Image Databases

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

Many image databases employ image features that describe the aggregate shape and color properties of a class of objects, and hence, are only capable of "inter-classes" discrimination (e.g., airplanes vs. butterflies). These features are sensitive to incidental changes in camera viewpoint and scene illumination, which severely constrain the appearance of query objects. The goal of this project is to design a set of new image-derived features which utilize detailed local image analysis to enable discriminating objects of very similar appearance within the same class (e.g., "intra-class" discrimination of different species of butterflies in a database of butterflies). These new features capture the essential traits of imaged objects in a way that is insensitive to incidental changes in both global environmental factors, such as viewpoint and illumination, and local configuration, such as shape deformation and articulated motion. The project addresses the efficiency and utility issues in deploying these invariant features in large, real-world image databases. In particular, flexible class templates are constructed for automating image segmentation and cataloging of database objects. Invariant image features are organized in a hierarchical manner, using both image-derived and domain-specific information, for efficient pruning of unlikely matches. A representation strategy, which combines global structure models with local invariant features, is used to achieve recognition insensitive to incidental articulated motion and deformation. As a result, the new image features are highly discriminative yet insensitive to incidental changes in shape, viewpoint, and illumination. In particular, they enable invariant, intra-class image retrieval, i.e., highly precise recognition and retrieval of images in large databases of similar objects (e.g., retrieving images of Old World Swallowtails from a database of butterflies) with few constraints on the appearance of query images. The results of the project will be disseminated by scientific papers, software that can be downloaded, and prototype image databases that can be queried on the Web. http://www.cs.ucsb.edu/~yfwang