Image Retrieval In Multipoint Queries

Traditional content-based image retrieval (CBIR) systems find relevant images close to an example image. This single-point model has been shown inadequate for complex queries built on high-level concepts. Recent CBIR systems allow users to use multiple examples to compose their queries. The multipoint model provides extended flexibility in identifying relevant sets of arbitrary shape that the previous approach is unable to formulate. However, the continuing use of conventional measures (e.g., LP norms) to evaluate these queries undermines the advantages of the new system. From the results of recent studies, we show that two important inferences can be made. Specifically, they are the continuity of image representation and the nonhomogeneity of the feature space. These characteristics enable the precise identification of points that satisfy the constraints established in multipoint queries and for any orthogonal feature representations. Generally, the sets are convex hulls and can be described by a linear system of equations with constraints. We discuss how to solve the system and propose an indexing procedure to efficiently determine the exact sets. We evaluated the performance of the proposed technique against state-of-the-art methods on large sets of images. The results indicate that the new measure captures semantic image classes very well, and the superiority of our approach over the recent techniques is evident in simulated and realistic environments. © 2008 Wiley Periodicals, Inc