Block multitransform coding of non-stationary signals

We develop a new class of multitransform (which we call “MBR”) algorithms for representing time-varying signals based on a parallel structure and compare their performance and implementation complexity to existing cascade structure algorithms. The algorithms use either a gradient or a non-gradient search over partial sets of several different basis functions to effectively capture different signal characteristics. We see that our proposed parallel structure yields algorithms that require fewer than one-half the number of computations of the previously derived cascade structure to represent the signal with equivalent error. Also, we use a non-gradient based recursive technique, the recursive residual projection (RRP) algorithm, to find the dominant components (DC) of the signals. We find that the MBR algorithms using the RRP search have superior performance in many instances