The goal of content-based image retrieval (CBIR) is to provide the user with a way to browse or retrieve images from large collections based on visual similarity. At the heart of any CBIR system are visual features that have been extracted from images and (dis)similarity functions that are used to quantify the similarity between these features. The combination of these two components will drive the overall performance of a system.

Two frequently studied research areas in CBIR are maximising retrieval performance using similarity measures and improving the efficiency and speed of search by applying indexing methods. Often these are mutually exclusive. The best performing similarity measures are usually computationally expensive and the optimal indexing approaches can place many restrictions on what features and similarity functions can be used.

In this talk we investigate how to localise the measurement of similarity. That is, emphasise points that are close to the query in some subspace of the full feature space. We show that this has dual benefits for CBIR, both improving retrieval performance and speeding up the search of high-dimensional features.

Watch the webcast replay >>