changes_from_vm6_0.txt

JPEG2000实现的源码

                        CHANGES IN VM6.0 FROM VM6.1
                       =============================

_____________________________________________________________________________
                 Changes for processing of overlapped tiles
_____________________________________________________________________________

* The `-Fovlp' encoder option has been added to specify which tile
  boundaries contain overlap.  Every tile can include overlap from other 
  tiles and the boundaries which contain overlap can be controlled.  See the 
  usage for `-Fovlp' for more details.

* Extra logic has been added to force SSO-DWT boundary extensions with 
  overlapping tiles.  The encoder flag which triggers this extension mode is 
  `-Xsso_ext' and should be used in combination with the `-Fovlp' option.  
  Refer to the usage for more details, but note that the encoder option 
  "-Fovlp LRTB -Xsso_ext" will cause full SSO-DWT processing across all tile 
  boundaries.

_____________________________________________________________________________
                       Changes for Component Mixing
_____________________________________________________________________________

* The energy weights are now optimally scaled to account for both color 
  transforms used with the `-Mycc' encoder option.  This will improve the 
  order of encoded bit-streams and improve the quality of decompressed color 
  imagery without necessarily specifying a `-Fweights' file.  Energy weights 
  will be automatically scaled (using YCrCb scaling constants) when exactly 
  three components are input to the VM encoder, and the energy weights for 
  the first three components will be appropriately scaled (using either 
  YCrCb constants for lossy compression or RCT constants for lossless 
  compression) whenever the `-Mycc' encoder option is used.  See the 
  discussion with the modify_third_d_filter_gains() routine inside 
  std_forward_info.c for details concerning these scaling factors.  Note 
  that use of the `-Fweights' encoder option will always override the 
  automatic scaling of energy weights.

* A discrete wavelet mixing object has been added to the VM.  When the 
  `-Mtdt' option is specified on the encoder command line, this object 
  currently uses either the 7X9 or 5X3 floating point and integer filters to 
  mix components for lossy and lossless compression.  

* A generic linear component mixing object has also been added to the VM.  
  This option is triggered with the `-Mlin' encoder command line.  As the 
  usage statement indicates, this option expects at least one argument 
  which specifies the location of a floating point forward matrix file and 
  can optionally read an explicit inverse matrix file.  Note that this 
  option currently stores the inverse matrix file inside the encoded 
  bit-stream as floating point numbers.  Therefore, as implemented, this 
  option is currently not platform independent.

_____________________________________________________________________________
                                   Bug Fixes
_____________________________________________________________________________

* Several small bugs when LRA and Xmask are used together were fixed.  These 
  bugs resulted in poorly collected statistics for the LRA technique as a 
  result of the sample-and-hold Xmask power law LUT's.  As a result, 
  erroneous step sizes were computed and a failed assertion occurred.  The 
  correction for this bug now uses linear interpolation with the the alpha 
  and beta power law LUT's.

* A significant problem in the general collection of LRA statistics with 
  tiling was fixed.  This problem could give more weight to subbands from 
  small tiles on the right and bottom of an image.  The LRA statistics 
  should now be uniform throughout.

* A memory allocation problem inside the Xmask implementation was fixed.  
  This bug caused segmentations faults when Xmask was used in combination 
  with tiling and memory for future tiles was deallocated.

* A residual bug from VM5 which caused the lifting extensions to work 
  incorrectly was fixed.  Note that this bug only occurred for very small (2 
  or 3 pixel) tiles.

* A problem with the COD/COC translation of packet-partition markers is now 
  fixed.  In VM6.0, this translation erroneously assumed that a valid Part-I 
  marker should have exactly one packet-partition byte, whereas in reality, 
  it should have L+1 bytes, where L is the number of decomposition levels.

* A bug which restricted packet partition dimensions to be square was 
  fixed.  Now packet partitions can obey rectangular dimensions.

* A bug which incorrectly masked 32-bit signed data was fixed.  This problem 
  dated back to VM3.0A and basically caused many bit-planes of 32-bit 
  imagery to be zeroed out.

* Incorrect layering logic which produced the wrong number of desired layers 
  was corrected.  

* The EOI marker was missing from VM6.0 encoded files.  Files encoded with 
  VM6.1 now have this marker at the end of the file.