01/01/1994 09:00 AM
Computer Science
Subband coding compresses grayscale still images of natural scenes by quantization and entropy coding of critically subsampled spatial frequency bands. A specially-designed multi-channel spectral analysis filterbank splits the bandwidth in such a way that a matched synthesis filterbank can exactly reconstruct the original image from the unquantized subbands and, from the quantized subbands, recover the image with distortion little or no worse than quantizing the image directly. With no change in the total sample count, compression to less than 1 bit/pixel comes from entropy coding of high- frequency channels, which quantize to almost all zeroes, due to the subband transformation packing most of the energy into the low-frequency channel and allowing quantization to be perceptually weighted by spatial frequency. FIR filterbanks are obvious candidates for parallel, pipelined VLSI implementation. Acceptance has been held back primarily by the computational expense of high-quality FIR filters and the lack of standardization. This Dissertation Proposal addresses both issues with a way of reducing the number of additions and multiplications for certain linear-phase exact-reconstruction two- channel filterbanks, constructed for this formulation gave better rate- distortion than any fo the Daubechies maximally-flat wavelet filters, yet the simplest requires only 6 additions and 2 shifts per pair of samples; including its modification at the image boundaries, its implementation in C is only 8 lines. Notes: Ph.D. Dissertation Proposal