An investigation into weighted data fusion for content-based multimedia information retrieval

Content Based Multimedia Information Retrieval (CBMIR) is characterised by the combination of noisy sources of information which, in unison, are able to achieve strong performance. In this thesis we focus on the combination of ranked results from the independent retrieval experts which comprise a CBMIR system through linearly weighted data fusion. The independent retrieval experts are low-level multimedia features, each of which contains an indexing function and ranking algorithm. This thesis is comprised of two halves. In the first half, we perform a rigorous empirical investigation into the factors which impact upon performance in linearly weighted data fusion. In the second half, we leverage these finding to create a new class of weight generation algorithms for data fusion which are capable of determining weights at query-time, such that the weights are topic dependent.