21-12-2012, 04:53 PM
A COMPARISON OF PREDICTION SCHEMES PROPOSED FOR A NEW LOSSLESS IMAGE COMPRESSION STANDARD
A COMPARISON OF PREDICTION SCHEMES PROPOSED.pdf (Size: 433.49 KB / Downloads: 21)
INTRODUCTION
Due to the perceived inadequacy of current standards for
lossless image compression, the JBIG/JPEG committee approved
a new work item proposal in early 1994, titled Next
Generation Lossless Compression of Continuous-tone Stall
Pictures. A call soliciting proposals was issued in March
1994. The proposed standard would provide lossless and
near-lossless compression (fkoff). It would target 2 to 16
bit still images and would be suitable for wide variety of
applications including medical, satellite and pre-press images.
It is anticipated that the proposed standard would be
significantly better than current lossless standards.
A series of requirements were imposed on submissions.
For details the reader is refered to [I]. For example. exploitation
of inter-band correlations (in color and satellite
images for example) was prohibited.
A total of nine proposals were submitted in response to
this call. In July 1995, the nine submitted proposals were
evaluated by IS0 on a very large set of test-images using an
objective performance measure that had been announced
prior to the competition. The results of the first round
evaluation are given in [7]. After this competitive round
of evaluations, the standardization effort has now entered a
collaborative phase, where the ideas proposed by different
groups will slowly be converged to yield the best possible
standard. A detailed evaluation of the different schemes
that were proposed in the first round, is currently underway.
ENTROPY CODING THE PREDICTION
In order to get a better comparison of the prediction
schemes that are being evaluated, we further modeled the
prediction errors and entropy coded them to get actual bitrate
figures. The error modeling and entropy coding scheme
we used were taken from the CALIC proposal and the same
entropy coder and error modeling scheme was used in conjunction
with all the predictors being evaluated.
The prediction errors were encoded using the eight primary
contexts A described in the previous section. Conditioning
the error distribution on A, separates the prediction
errors into classes of different variances. Thus entropy coding
of errors using estimated conditional probability p(elA)
improves coding efficiency over using p(e). We used an
adaptive m-ary arithmetic coder, CACM++ package that
was developed and made publicly available by CarpinelLi
and Salamonsen. The software is based on the work by
Moffat, Neal, Witten [4]. However, in order to improve
coding and space efficiency some reversible pre-processing
of errors was done prior to entropy coding. Details are given
in [9].
Performance Comparison
Using the context-based error modeling and feedback
mechanism described above led to the results shown in table
2. We observe the following:
1. Error feedback can some times worsen the performance
2. On an average, error feedback helps reduce the entropy
of a prediction scheme.
of prediction errors.
3. GAP gains most with error feedback.
4. ALCM, LOCO and CALC still give the best results.
5. Error feedback helps ALCM the least. In fact, its performance
deteriorates more often than not.
6. Error feedback caused ALCM to break down for the
"cafe.tif.black" image. This indicates that when using
error feedback with ALCM, the adaptation algorithm
must clip the pixel weights beyond a certain range.
CONCLUDING OBSERVATIONS
Before we make any conclusions we would like to point out
that the current study only evaluated the prediction components
of different proposals. The prediction scheme was
just one component of these proposals which included different
error modeling and encoding schemes along with many
'tricks' to improve computational and compression performance.
So, the results given in these paper in no way reflect
the compression performance of the entire codecs that were
proposed. An evaluation of the compression performance
of the different proposals is given in [7].