Fusion Coding of Correlated Sources for Storage and Selective Retrieval

We focus on a new, potentially important application of source coding directed toward storage and retrieval, termed fusion coding of correlated sources. The task at hand is to efficiently store multiple correlated sources in a database so that, at any point of time in the future, data from a selective subset of sources specified by user can be efficiently retrieved. Only statistical information about future queries is available in advance. A typical application scenario would be in storage of correlated data generated by dense sensor networks, where information from specific regions is requested in the future. We propose a fusion coder (FC) for lossy storage and retrieval, wherein different queries are handled by allowing for selective (compressed) bit retrieval. We derive the properties of an optimal FC and present an iterative algorithm for its design. Since iterative design is initialization-dependent, we present initialization heuristics that help avoid poor local optima. An analysis of design complexity reveals complexity growth with query-set size. We first tackle this problem by exploiting optimality properties of FCs. We also consider quantization of the query-space with decision trees in order to adapt to new queries, unseen during FC design. Experiments conducted on real and synthetic data-sets demonstrate that the proposed FC is able to achieve significantly better tradeoffs than joint compression by vector quantization (VQ), with retrieval speedups reaching 3 ?? and distortion gains of up to 3.5 dB possible.