JPEG2000: The New Still Picture Compression Standard
Christopoulos, Ebrahimi, Skodras

Jonathan Ledlie (jonathan@eecs)

Instead of using a discrete coefficient transform (DCT), as in the
original JPEG, JPEG2000 uses a discrete wavelet transform (DWT), which
separates the input into downsampled low-resolution samples and
high-resolution ones.  Thus, at the transformation step, JPEG2000 is
either lossless or lossy; this is unlike JPEG where the initial DCT
transformation was ideally always reversible -- its output then was
perhaps fed into a lossy quantizer.  The default irreversible
transformation is implemented with the Daubechies 9-tap/7-tap filter
and the default reversible one is given with the 5-tap/3-tap filter.
Like JPEG, this step is then followed by a quantization, which is
lossy unless the quantization step is 1.  The final compression step
is again a form of entropy coding; JPEG2000 uses Elias coding as the
basis for a binary arithmetic coding process.  The specification does
not mention Huffman coding as an alternative, like in JPEG.

The most interesting benefits added to the specification are the
ability to specify a region-of-interest (ROI) and the huge improvement
in compression at the same rate of loss (double in their one example).