THE SpEnt METHOD FOR LOSSY SOURCE CODING †

At present, the most successful methods for lossy source compression are sample-function adaptive coders. Prominent examples of these techniques are the still image compression methods utilizing wavelet expansions and tree structures, such as the zero-tree method or the SPIHT algorithm, and variable rate speech coders that allocate bits to parameters within a frame based upon the classification of the current frame. All of these techniques can be classified as non-linear approximation methods. In this work, we use Campbell’s coefficient rate, and the spectral entropy of the source random process, as a guide to formulate a new non-linear approximation method to lossy source compression. We call this new approach, the spectral entropy (SpEnt) method, and we develop and report on the promise of SpEnt based coders for the lossy compression of still images and wideband speech (50 Hz to 7 kHz).