Distance-Based Similarity Search in Multimedia Databases

Today, databases are frequently used to store and manage rapidly growing amounts of multimedia data produced ubiquitously in our daily lives by digital devices including mobile phones, digital cameras, or personal digital assistants. To counteract this ongoing data explosion, multimedia databases support users in searching and browsing these tremendous amounts of data via different search paradigms. The probably most important paradigm is similarity search, which aims at retrieving the most similar objects given a specific query object. Besides the classic database research area, the fields of content-based image/audio/video retrieval – in general multimedia retrieval [5] – are widespread research areas in which the search of similar multimedia documents is of major importance. In general, similarities among multimedia objects are frequently determined by computing distance values among automatically extracted feature representations. The underlying similarity model which specifies feature representation and distance function is of crucial importance for the systems’ overall performance. On the one hand, large multimedia databases have to be searched quickly, on the other hand the returned results have to match to the users idea. Thus, the success of similarity models is strictly related to their performances which can be measured in terms of effectiveness and efficiency. In the last decade, flexible types of feature representations, such as feature signatures [7], tend to outperform traditional types, such as feature histograms, as they are appropriate to reflect and adjust to the contents of diverse multimedia objects. Based on feature signatures, vector-based distance functions, such as the famous Minkowski Distances, are no longer applicable and setoriented distance functions including the Hausdorff Distance [3], Perceptually Modified Hausdorff Distance [6], Earth Mover’s Distance [7], or Weighted Correlation Distance [4] have to be applied. Inspired by the Quadratic Form Distance [2], we recently proposed the Signature Quadratic Form Distance [1] which generalizes the Quadratic Form Distance and bridges the gap between feature signatures and the cross-dimension concept. As a result, the Signature Quadratic Form Distance is able to compare feature signatures of different structure and size with each other. In our experiments, we show that the Signature Quadratic Form Distance competes with state-of-the-art distance functions in terms of effectiveness and, moreover, outperforms the Earth Mover’s Distance in terms of efficiency. We review adaptable distance functions on flexible feature representations which are applicable to similarity search in multimedia databases. We show the difference to traditional feature representations and introduce the aforementioned distance functions. Highlighting our Signature Quadratic Form Distance, we conclude with the experimental results.