changes_from_vm3_2a.txt

JPEG2000实现的源码

                         CHANGES IN VM4.0 FROM VM3.2A
	                 ============================
______________________________________________________________________________
                                  Bug Fixes
------------------------------------------------------------------------------

* A bug in the tiling implementation which was found by Motorola.
  Specifically, the tile dimensions supplied on the command line arguments
  identify the maximum allowable dimensions of any tile.  The system computes
  actual tile dimensions subject to these limits based upon desirable
  relationships between different components (component dimensions are
  arbitrary) and the requirement that the nominal tile dimensions are powers
  of 2 so that only tiles on the right and lower boundaries of the image
  may have non-power-of-2 tile dimensions.  This usually means that actual
  tile sizes will be the largest power of 2, not exceeding the supplied
  dimensions, but in the special case where one or more of the supplied
  dimensions is equal to or larger than the largest corresponding dimension
  amongst all the image components, the nominal tile size may be set to
  a power of 2 which is larger than the respective image component dimension
  without violating the constraints.  The bug lay in the fact that this
  special condition was not taken into account in the code which can lead
  to unexpected results when using the software.  The bug fix involves minor
  changes to the code in "std_forward_info.c", within the function,
  "__initialize", immediately following the section comment,
  "Construct tiles and find all dimensions".

* A bug in the implementation of multiple ROI's with arbitrary geometry
  was detected and fixed by David Taubman.  The fix involes a minor change
  to the implementation of file "ref_roi_common.c" in function
  "roi_buffer__access_rows".

______________________________________________________________________________
                                 ROI Changes
------------------------------------------------------------------------------
* Inclusion of the rectangular mask generation code from VM3.0A, which
  had been dropped in VM3.2A.  Included code differs somewhat in
  implementation to minimize redundancy between generic mask generation
  implementation and rectangular mask generation implementation.
  Changes to the code associated with this capability are identified by
  the key phrase "ROIrect", which appears in comments against individual
  lines and in bracketing comments.  The files which include changes of
  this type are:
        -- "ref_forward_roi.c"
        -- "ref_reverse_roi.c"
        -- "ref_roi_common.c"
        -- "ref_roi_common.h"
  The modification takes tiling and frames into account to produce exactly
  the same ROI mask as that which would be generated using the arbitrary
  shape implementation, provided the regions are rectangular in shape.
  A new command-line flag is introduced, `-Rno_rect', which can be
  applied during compression and/or decompression to suppress the use
  of the efficient ROI mask generation implementation when the regions
  all have rectangular geometry.  This allows the user to compare the
  results of the ROI mask generation used for rectangular regions with
  that used for arbitrary shape regions to confirm that they agree.  If
  the option used during compression differs from that used during
  decompression, any differences in the implementations should lead to
  serious artifacts in the reconstructed image and so will easily be
  detected.

* Inclusion of an explicit command-line argument to force the use of the
  `max-shift' method -- in VM3.2A the `max-shift' method was used
  automatically if the scaling factor for the ROI was sufficiently large,
  but this obscured the existence of the mode.  The new flag is
  `-Rmax_shift'.  The modifications to the code are minor and are
  restricted to the files:
        -- "ref_forward_roi.c"
        -- "ref_reverse_roi.c"

_______________________________________________________________________________
                               Error Resilience
-------------------------------------------------------------------------------
* Included error detection marker between bit-planes in the embedded block
  bit-streams.  Code supplied by Sarnoff corporation and modified slightly
  by David Taubman.  There are very few changes.  The changes are identified
  by comment banner lines containing the keyword "Sarnoff" and are
  restricted to the following files:
       -- "ebcot_encoder.c"
       -- "ebcot_decoder.c"
       -- "ebcot_encoder.h"
       -- "ebcot_decoder.h"

______________________________________________________________________________
        Major Changes to the Global Header of the Bit-Stream
------------------------------------------------------------------------------
                                Gene Wu, RSV
* The actual bytes written to the global header have been changed to
  a marker/tag based structure to be compatible with JPEG parts I & III.
  This has the added benefit that the global header is no longer dependent
  on the particular of calls to push_global_header(), likewise the decoder
  no longer needs to ask for items with pull_global_header() in the exact
  order the encoder wrote them.

* This is currently accomplished by giving push_ & pull_ global_header
  algorithmic knowledge about the values being pushed. In the future
  it will be good if various objects request items close to the way
  they are stored in the global header.

______________________________________________________________________________
                                  Tiling
------------------------------------------------------------------------------
                              Michael Gormish, RSV

The arrangement of the coded data written to the bitstream has been
changed.  Formerly, data could be written in "SNR progressive" or
"resolution progressive" order. Now the data is always in "Tile
progressive" order, but within a tile the data can be either "SNR
progressive" or "resolution progressive" order. The coded data is
currently identical to VM3_2A in the case of one tile. There is a
brief tile header (currently 12 bytes) before each tile, containing
Start of Tile (SOT), Tile Length Tile (TLT), and Start of Scan (SOD)
markers. The Tile Length allows spatial access to regions of the image
without decoding even the packet header syntax. The length and
information included in the tile header will change in future
versions.

The basic structure of objects has not been changed. In particular
there is still a linked list of tiles within each subband of each
transform level for all layers and components.

form_packet() in ebcot_send_bits.c is now form_tile_packet(), because
it only forms a packet with header and coded data for one tile. An
equivalent change occures with receive_packet() in
ebcot_receive_bits.c. ebcot_send_bit_stream() has been changed to
loop over tiles first, then layers and resolutions (in either order).

The function write_tile_header() has been added to the bitstream object.
Thus tile headers are independent of packets, and in particular are not
stuffed (because tile headers contain markers).