A Parallel Adaptive Wavelet Processor for Realtime Pattern Recognition Applications

Wavelet transforms have already proved to be very useful tools in many signal processing applications including digital audio and video compression. Although the wavelet-based representations offer more versatility in time-frequency localization than other Fourier methods, they tend to have poor localization especially near high frequencies. To eliminate this difficulty, we propose a technique for analyzing signals of time-varying characteristics using adaptive wavelet packets, which facilitates identification of arbitrarily localized spectral characteristics. For nonstationary signals that contain varying distribution of spectral energy, the analysis must adaptively select filters of different localization properties. Wavelet-packets designed through digital filter banks offer arbitrary tiling of the time-frequency domain. Herley et al. [1993] report a double-tree algorithm that combines selection for the best wavelet-packet expansion with an adaptive time-segmentation of the signal simultaneously. The system we propose here is an extension of the above algorithm together with a choice of a basis function from a library of wavelet packets. We also propose to develop the pattern recognition capabilities based on feature vectors computed from the compressed signals. A real-time prototype wavelet transform processor will be available in Phase II on an inexpensive parallel hardware called PC - CNAPS that operates as an accelerator board to a PC.