Universal Multiscale Matching Pursuits 10.14209/jcis.2000.3

In this paper we describe the UMMP (Universal Multiscale Matching Pursuits), a practical universal algorithm for multi-dimensional data lossy compression. The method is based on approximate multiscale matching of recurrent patterns. It uses a dictionary of basis functions in which the data is decomposed in the spirit of Mallat's Matching Pursuits (MP). Unlike MP however, UMMP builds its own dictionary while encoding the data, instead of using a previously defined one. Also, all basis functions can be contracted or dilated to better match the input data during the expansion. This allows the algorithm to perform arbitrarily close to the source's D(R) function when the rate goes to zero. Simulation results show that it has good coding performance for a large class of data.