changes_from_vm4_2.txt

JPEG2000实现的源码

TESTING NOTES:
We suggest that error resilience be tested with `-Cer_term' at the
encoder and `-Cer 2' at the decoder.  This combination may be used
successfully with any other combination of options (e.g. `-Clazy',
`-Cparallel', `-Clayers', etc.) and should work well regardless of
the number of quality layers in the bit-stream (one is sufficient).
For added error resilience, add `-Csegmark' to `-Cer_term'.  As
previously, please report bugs in the error resilience code to David
Taubman.  It is best to run the decoder in the debug mode so that
assert() statements get caught.  This will allow us to iron out
oversights more quickly.

______________________________________________________________________________
                            Changes in Transform
______________________________________________________________________________
* The boundary termination policy for lifting implementations has been
  corrected to ensure compatibility with the symmetric extension used
  for convolution implementations.  This work was done by Christos
  Chrysafis of HP Labs.

* The PSE implementation no longer appears explicitly in the VM, as
  agreed in Vancouver.

______________________________________________________________________________
                          Changes in Visual Options
______________________________________________________________________________
* The `-Cvis' encoder-only option has been implemented.  This option
  adjusts the distortion metric used during R-D optimization to take
  account of the visual masking phenomenon.  This phenomenon is quite
  distinct from the better known Contrast Sensitivity Function (CSF),
  which may be compensated by applying appropriate weighting tables to
  `-Fweights'.  Whereas the CSF is a global characteristic, mostly
  attributable to MTF of the human optical system (optics + sensor
  integration PSF's) and any assumptions on the MTF of the display
  device (monitor, printer, etc), the masking phenomenon is an
  image-dependent, spatially varying phenomenon.  The CSF depends
  intimately on assumed viewing distance (and possibly display
  properties).  The compensation for visual masking built into
  `-Cvis' is designed to be complementary to any CSF compensation.
  Thus, the visual masking formulation itself is independent of
  assumptions on viewing distance, display device, etc.  This means
  that when nothing is known about such parameters (e.g. the user
  is allowed to zoom in and out of an image arbitrarily on a monitor),
  the `-Cvis' compensation usually gives substantial overall
  improvement in image quality.  When the viewing parameters can
  be guessed, CSF compensation through appropriate weighting tables
  supplied to `-Fweights' usually gives further improvements when
  combined with `-Cvis'.  Note that the behaviour of `-Cvis' is
  informative only from the point of view of the JPEG2000 standard,
  since the EBCOT paradigm embodied in the syntax and decoder support
  arbitrary policies for determining the relative truncation points
  for each code-block.  Note also, that `-Cvis' generally works best
  with smaller code-blocks (typically 32x32); a warning message will
  be generated if it is used with the default 64x64 code-block size.

______________________________________________________________________________
                  Changes to the Global Header of the Bit-Stream
______________________________________________________________________________

* Since the PSE no longer exists explicitly, there is no global
  header entry for it.
* The nominal EBCOT block size parameters are now restricted to powers of 2 in
  the range 4 to 256 and are signaled by 3 bit quantities representing
  log base 2 of the size, minus 2.
* Since we have fixed all entropy coder quantities relating to sub-blocks
  (i.e. the scan), there is no longer any sub-block size identification in
  the header.
* There is a flag indicating whether or not context formation during
  entropy coding should be restricted by a vertically stripe causal
  constraint.
* There is a flag identifying whether or not the probability models should
  be reset at the end of each coding pass.
* There is a flag identifying whether or not lazy coding mode is in use.
* There are two flags relevant to error resilience: a) the segmentation
  marker flag which was there before; and b) a flag indicating whether or
  not the decoder can rely upon the encoder having used a particular
  standard algorithm for terminating codeword segments in the bit-stream
  (if present, the properties of this algorithm can be exploited for
   error detection; otherwise, the encoder can use any method it likes, but
   the decoder cannot exploit spare code space left after termination for
   error resilience).
* ROI temporary CME tag is now a permanent syntax tag.

______________________________________________________________________________
                                  Bug Fixes
______________________________________________________________________________

* A bug in "ebcot_send_bits.c" pointed out by Diego Santa Cruz has been
  fixed.  The bug caused occasional accesses outside allocated memory,
  which became observable in Java.

* A bug in "std_kernel_info.c" was found by Christos and fixed.  This
  "bug" allowed different signs conventions to be used by convolution and
  lifting implementations of a filter, depending upon how the relevant
  parameters appeared in user kernel files and the "std_coeffs.h" file.
  The modification changes the sign of the high pass coefficients for
  the 5X3 filter, relative to what was being used in previous implementations.

* A general syntax bug was reported by Frederick Wheeler.  This bug resulted 
  in array bounds problems, but has now been fixed.

* A problem with retrieving visual weights in the decoder was reported by 
  Wenjun Zeng.  This problem resulted in poor results when masking AC subbands, 
  but has been fixed.

* Paul Skeberdis reported comment in comment compilation warnings generated on 
  his SGI Octane running IRIX 6.4 with SGI's version 7.2.1 compiler.  These 
  warnings have now been fixed.

* David Taubman and others reported that the header size was much larger than 
  necessary.  As a result of fixes put in for this, the VM5 header size should 
  be smaller and reported correctly.

* A component mixing problem reported independently by David Sheppard and 
  Takahiro Fukuhara was fixed.  The result of this bug was that only three 
  component could be processed.  Now the VM should be able to process 256 
  components.

* Several bug reports/complaints concerning tile size implementation were 
  addressed.  

* A -Fsteps problem reported by Rajan Joshi was fixed.  This problem was only
  apparent for multi-component imagery with different step sizes for each
  component.

* A discrepancy between the VM and syntax documentation was reported by Scott 
  Houchin.  This concerned the length of tile packets not being equal to a 
  multiple of 16.  This problem is fixed in the current VM.

______________________________________________________________________________
                            COMMAND LINE OPTIONS
______________________________________________________________________________

Be careful to read the usage statements generated by the compressor
and decompressor, paying attention to the descriptions of the following:

Compressor:
* -Cparallel
* -Clazy
* -Ccausal
* -Cnon_causal
* -Cer_term
Decompressor:
* -Cer
* -Cno_speedup