changes_from_vm8.0.txt

JPEG2000实现的源码

                        CHANGES IN VM8.1 FROM VM8.0
                       ============================

Additions:
----------

* The transform object has been modified to allow even length convolutional 
  filter processing with arbitrary canvas dimensions and arbitrary canvas 
  offsets.  Note that this addition currently works only when the `-Fconv' 
  flag is specified for both the encoder and decoder.  Only 2-tap lowpass 
  analysis filters have been tested with these modifications.  Three such 
  tested filters are included as kernels 2,3 and 4 in the 
  kernels/float_lifting directory.

* A (7,5) lifting filter has been included in the kernels/float_lifting 
  directory.  This filter is in file kernel-005 of that directory and may be 
  specified at the encoder with the `-Fkernels 5' command line option.


General Changes:
----------------

* The stream_out object has been modified to force inclusion of the EOC 
  marker in every encoded codestream.  This object was also adjusted to 
  properly account for the two byte length of this marker when determining 
  encoded file sizes.

* The variable DC offset code has been modified to allow floating point 
  offsets.  Offsets provided through the `-Fdc' encoder argument must be in 
  the range [-32768, 32768), but are store in the codestream as fixed point 
  16-bit numbers.  Also, the variable DC offset code was also extended to 
  allow different DC offsets for each tile.

* The variable deadzone capability for quantization was modified to allow 
  floating point numbers in the range (-1.0, 1.0).  A 16-bit fixed point 
  approximation to the number specified via `-Qnz' is store in the 
  codestream for the decoder.

* The component_mix object has been modified to ensure that the number of 
  guard bits is high enough to accommodate `-Mlin' linear transform 
  processing.

* The raw multi-component I/O routines available with the `-imcraw' and
  `-omcraw' encoder and decoder switches have been modified to compile
  and run on Sun, HP/Convex, and Win98/NT platforms.

* The calculation of implicit step sizes was modified to properly account
  for half level processing which can occur with general decompositions.


Bug Fixes:
----------

* Incorrect component sub-sampling for build_partitions() routines in both 
  encoder and decoder.

* Unexpected COC markers inserted into header information.

* NULL strings in component wavelet transform kernel ID's.