Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
Optimized Signal Quantification
Related Application
The subject matter of this application is related to the subject matter of the
following commonly owned applications: Serial Number 09/112,668, attorney
docket
number 3486, titled "Apparatus And Method For Entropy Coding', filed on July
9,
1998, also by Kenbe Goertzen; Serial Number , attorney docket number 4754,
titled "Scaleable Resolution Motion Image Recording And Storage System', filed
concurrently, also by Kenbe Goertzen; Serial Number , attorney docket
number 4753, titled "A System And Method For Improving Compressed Image
Appearance Using Stochastic Resonance And Energy Replacement", filed
concurrently, also by Kenbe Goertzen; Serial Number , attorney docket
number 4754, titled "Scaleable Resolution Motion Image Recording And Storage
System', filed concurrently, also by Kenbe Goertzen; Serial Number
attorney docket number 4756, titled "Quality Priority Image Storage and
Commurucatiori', filed concurrently, also by Kenbe Goertzen; the contents of
which
are incorporated by reference as if fully disclosed herein.
Technical Field
This invention pertains to the field of digital signal compression and
quantification. More specifically, the present invention related to a method
for
optimizing signal quantification, particularly the quantification of signals
transmitting still and motion image components.
Shortcomings of Prior Art
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
Multiband compression methods have generally divided a signal into
frequency components and then used some method to quantify the values in each
of
the frequency bands in order to represent the desired signal quality. Some of
the
problems with this approach include:
1. Only a small number of frequency bands are used so the regional
quantification
method is a Less than optimal coarse approximation of the desired function.
2. Unequal quantification of neighboring frequency bands generally increases
the
amount of aliasing in the reconstruction mechanism.
3. Many quantification methods can generate undesirable artifacts in various
degenerate cases.
4. Quantification as a separate process adds time or hardware to the
implementation.
5. Quantification as a separate process can add additional noise.
What is needed is a system and method for quantifying one or more signals in
an image stream such that the method can be easily implemented while offering
better coding efficiency and more degrees of freedom in designing subband
filters.
Summary of Invention
Wavelet compression of images generally consists of subband transforms of an
image into frequency regions. These regions are then quantified to relative
resolutions and entropy coded. The present invention uses a continuous and
frequency specific function to quantify the regions rather than a regional
approximation.
More specifically, the method uses reversible filters before and after signal
quantification as a continuous function of the frequency domain. This allows
the
2
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
scaling function to be either the exact desired function, or more closely
approximate
the desired continuous quantification function. It also allows the
characteristics of any
quantification and aliasing artifacts to be tailored to the particular
application. As a
result, the present invention provides better interpolation of quantification
errors
resulting in less noticeable artifacts in the quantified stream. The present
invention
also provides lower aliasing energy between subbands.
Brief Description of the Drawings
These and other more detailed and specific objects and features of the present
invention are more fully disclosed in the following specification, reference
being had
to the accompanying drawings in which:
Fig.1 is a flowchart illustrating the preferred method of the present
invention;
Fig. 2 is an example of a typical 2D quantification map for a 2 band pyramid
transform;
Fig. 3 is an example of an optimum quantification map for a separable
complete 2 band transform;
Fig. 4 is an example of an optimum quantification map for a non-separable
complete 2 band transform;
Fig. 5 is an example slice of the quantification surfaces demonstrating the
advantages provided by the present invention; and
Fig. 6A and 6B are examples of a bandsplit filter that was made more accurate
by applying the method of optimum quantification.
Detailed Description
Wavelet compression of images generally consists of subband transforms of an
image into frequency regions. These regions are then quantified to relative
3
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
resolutions and entropy coded. Although the method will be described in terms
of
steps, the steps involving application of the quantification and inverse
quantification
functions to the signal can also be incorporated directly into the subband
analysis and
synthesis filters thus avoiding a separate quantification and dequantification
step.
Referring now to figure 1, the method of the present invention is shown. In a
first embodiment, the method begins by applying 110 reversible filters to the
signal in
order to pre-quantify the signal as a continuous function of the frequency
domain.
'This allows the scaling function to be either the exact desired function, or
more
closely approximate the desired continuous quantification function. It also
allows the
characteristics of any quantification and aliasing artifacts to be tailored to
the specific
application or signal.
After filtering the signal, the signal is pre-processed 120. This might
include
performing any number of different processes on the signal, such as converting
the
colorspace. Other forms of preprocessing may be applied depending on the type
of
signal and the desired output. In step 130, subband transforms of the signal
split the
signal into frequency regions. This enables quantification of the signal by
region.
While the region quantification step is described as a separate step from
subband transforms, these two steps may be combined into a single step if
desired. In
this step 140, each region is separately quantified. In the preferred
embodiment of the
present invention, a continuous function is used to quantify the frequency
regions.
This can provide a significant improvement in efficiency over a stepwise
approximation, as well as a reduction in aliasing. Finally, the quantified
signal is
subjected to entropy coding resulting in maximum compression. In the preferred
embodiment, the step of entropy coding is performed in accordance with the
entropy
4
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
coding method described in the related application entitled, "Apparatus and
Method
for Entropy Coding."
An important application of this method is still and motion image encoding.
Images are often sampled as discrete x,y or x,y,z grids of digital values. If
these
values are Iinearized with uncorrelated noise, then sampling theory indicates
that as
the frequency of interest is reduced from the Nyquest frequency, the signal
resolution
increases. It is desirable to quantify the transformed image to a curve
matching the
resolution function established by sampling theory. This function in one
dimension
is:
resolution = (1/frequency)~0.5 * Nyquest resolution
where frequency is in the range of {0..1} where 1 represents the Nyquest
frequency
The separable n dimensional case is:
resolution = (product from 1 to n of (1/Fn)~0.5 ) * Nyquest resolution
Subband image coders typically quantify the subbands using a stepwise
approximation of this function to obtain acceptable images after compression.
If a
continuous function is used for the quantification, it can provide a
significant
improvement in efficiency over a stepwise approximation, as well as a
reduction in
aliasing. Reversible quantification filter functions can also be developed to
provide
other types of continuous quantification functions other than those provided
by
sampling theory. An example would be a continuous quantification function
matching human perception resolution.
Even in cases where an exact continuous filter and its inverse cannot be
designed for the desired quantification function, benefit from the method can
be
obtained by treating the desired function as a lower bound and designing a
filter and
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
its inverse which are bounded by the desired continuous function but better
than a
stepwise approximation by region.
The greater the dynamic range required of the quantification filter and its
inverse, the more difficult they are to design. This problem can be solved by
dividing
the range into regions combined with the subband transform process. Once the
range
has been divided, a family of filters may be used to approximate the desired
continuous function. Alternatively, a single filter can be used recursively.
If the subband mechanism has adequate resolution, the subband filters can be
convolved with the quantification filters to combine the transform steps. In a
preferred embodiment, the quantification filter design is combined with the
subband
filter design to increase the degrees of freedom for design of the subband
filters. This
would enable one to design a subband filter in a similar fashion to
biorthagonal filter
design.
Subband transforms may also be applied separately by dimension. If subband
1 S transforms are to be applied separately by dimension, then the type of
transform is
determined by the desire to generate a smooth and continuous quantification
function matching the resolution specified by sampling theory. This requires
that
only the low frequency halfband from each bandsplit be further subdivided.
In the case of two or more dimensions, this requires a more complete
transform than the typical pyramid subband transform. A two-dimensional two-
band pyramid transform generates 7 regions, a complete transform of low
frequencies
by dimension generates 9 regions, and a full transform would generate 16. Non-
separable multidimensional filters and subband transforms can then be designed
which allow the generation of the smooth, continuous quantification function
for a
6
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
pyramid transform. While the above method was described as the preferred
embodiment, the following embodiments represent alternative compression
processing methods.
The following is an alternate set of steps for a Precompensation -based
subband image compression process.
Step 1: preprocessing such as color space conversion
Step 2: precompensation filter matching sampling theory resolution (may be
combined with transform)
Step 3: subband transform
Step 4: entropy coding
The following is an alternate combined subband image compression process
Step 1: preprocessing such as color space conversion
Step 2: subband transform matching sampling theory resolution
Step 3: entropy coding
Examples and Related Calculations
The following three plots outline the difference for a quantification surface
which will assure 0 dB of loss at the Nyquest frequency. Referring to Figure
2, an
example of a typical 2D quantification map for a 2 band pyramid transform is
shown.
The illustrated transform was generated by using the following plotting
equation:
Plot3D[dB[rs2b[x,y]],{x,O,Pi},{y,O,Pi},PlotPoints->30,PlotRange->{0,18}];
Referring to Figure 3, an example of an optimum quantification map for a
separable complete 2 band transform is shown. The illustrated transform was
generated by using the following plotting equation:
Plot3D[dB[rc2b[x,y]],{x,O,Pi},{y,O,Pi},PlotPoints->30,P1otRange->{0,18}];
7
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
Referring now to figure 4, an example of an optimum quantification map for a
non-separable complete 2 band transform is provided. The illustrated transform
was
generated by using the following plotting equation:
Plot3D[dB[rr2b[x,y]]+3,{x,O,Pi},{y,O,Pi},PlotPoints->30,P1otRange->{0,18}]
Referring now to figure 5, an example slice of the quantification surfaces
demonstrate the advantages provided by the present invention. The plot was
generated using the following plotting equation:
Plot[{dB [rs2b [x,x]],dB [rc2b [x,x]],dB [rr2b[x,x]]+3}, {x,O,Pi},Plo tRange->
{0,20}
The following equations provide a method for calculating the non-separable
and separable advantage over a stepwise approximation for a given channel
resolution.
N[NIntegrate[dB[rs2b[x,y]]-(dB[rr2b[x,y]]+3),{x,O,Pi},{y,O,Pi}]/Pi~2] 3.77673
N[NIntegrate[dB[rs2b[x,y]]-dB[rc2b[x,y]],{x,O,Pi},{y,O,Pi}]/Pi~2] 3.25787
Referring now to figure 6A and 6B, an example of a bandsplit filter which was
made more accurate by applying the method of optimum quantification is shown.
The variables used in the plotting equation are provided below.
tal={1168, 590, -106, -78, 34};
tsl={1168, 590, -106, -78, 34};
tah={990, -454+1, -166, 83-1, 42};
tsh={1390+6, -769, -18-3, 62, 30};
ralt[w_]:=Sum[Cos[n*w]*tal[[Abs[n]+1]],{n,-4,4}]/ (2048)
raht[w_]:=Sum[Cos[n*w]*tah[[Abs[n]+1]],{n,-4,4}]/ (2048)
rslt[w_]:=Sum[Cos[n*w]*tsl[[Abs[n]+1]],{n,-4,4}]/ (2048)
rsht[w_]:=Sum[Cos[n*w]*tsh[[Abs[n]+1]],{n,-4,4}]/(2048)
CA 02361410 2001-08-02
WO 00/46738 PCT/US00/03051
Figure 6A was plotted using the following equation:
Plot[{(raft[x]*rslt[x]+raht[x]*rsht[x]-1)*1000},{x,O,Pi}].
Figure 6B was plotted using the following equation:
Plot[{ralt[x],raht[x],rslt[x],rsht[x] }, {x,O,Pi}].
Although the description above contains many detailed descriptions, these
descriptions should not be construed as limiting the scope of the invention
but merely
as providing illustrations of some of the presently preferred implementations
of this
invention. For example, although this method was described with reference to
standard motion and still images, this method can be used to optimize
quantification
of any signal stream. Thus the scope of the invention should be determined by
the
appended claims and their legal equivalents, rather than by examples given.
9
