Interactive Pattern Analysis for Searching Multimedia Databases

Relevance feedback is a mechanism to interactively learn the user’s query concept online, and has been extensively used to improve the performance of multimedia information retrieval. In this paper, we present a novel interactive pattern analysis method, which reduces the relevance feedback to a two-class classification problem, and classifies multimedia objects as relevant or irrelevant. From those objects that are classified as relevant, our approach returns the top nearest neighbors of the query to the user. To classify multimedia objects, we employ a pattern classification algorithm called random forests, which is a machine learning algorithm with proven good performance on many traditional classification problems. By using a novel two-level resampling method, we adapt the original random forests to an interactive algorithm, which achieves noticeable gains in efficiency over the original random forests. We empirically compare our method with previously well-known relevance feedback approaches on a large-scaled image database (with 31,438 COREL images), and report promising results of our method.