Adaptable Distance Functions for Similarity-based Multimedia Retrieval

Today’s abundance of storage coupled with digital technologies in virtually all scientific or commercial applications such as medical and biological imaging or music archives deal with tremendous quantities of images, videos or audio files stored in large multimedia databases. For content-based data mining and multimedia retrieval purposes, suitable similarity models are crucial. Adaptable distance functions are particularly well-suited to match the human perception of similarity. Quadratic Forms (QF) were introduced to capture the notion of inter-feature similarity which sets them apart from the more traditional feature-by-feature measures from e.g. the Euclidean or Manhattan dissimilarity functions. The Earth Mover’s Distance (EMD) was adopted in Computer Vision to better approach human perceptual similarities by allowing feature transformation under a number of restrictions. After recapping the concepts of distancebased similarity search in databases, we familiarize the reader with the flexible building stones behind Quadratic Forms and the EMD. These enable their application to a large variety of multimedia retrieval problems. Unfortunately, the flexibility comes at a cost. Their computation is relatively time-consuming, which severely limits its adoption in interactive multimedia database scenarios. Therefore, we research methods to speed up the retrieval process and show some encouraging recent results to achieve just that via an index-supported multistep algorithm based on new lower bounding approximation techniques.