Incremental refinement structures for approximate signal processing
Joseph M. Winograd · 1997
This work investigates approximate signal processing as a design philosophy supporting the realization of efficient, robust, and flexible digital signal processing systems through the use of incremental refinement structures that allow tradeoffs to be easily made between the accuracy or optimality of results and the utilization of computing resources such as time, power, and chip area. The value of this approach is demonstrated through the theoretical development of incremental refinement structures for signal detection using the fast Fourier transform (FFT), image decoding using the two-dimensional inverse discrete cosine transform (2D IDCT), and spectral analysis using the discrete Fourier transform (DFT). Using both deterministic and probabilistic techniques, the theoretical performance of these structures under various resource constraints is quantified in terms of well-defined measures such as probability of detection, SNR, and frequency resolution. These analyses are verified for...