Image Compression Using
- Wavelet Packets
BY-Deepti Gupta., Shital Mutha Students-Cummins College of Engineering,Karvenagar,Pune-52
Phone-(020)54672 11/12, sd-bhide@[Link]
Abstract: . '-
Algorithms based on wavelets have heen shown to
work well in image compression. Advances in A multilevel wavelet filter bank involves iterating the low pass
wavelet transform a n d quantisation methods have and high pass filtering and down sampling procedure on the
produced the algorithm capable of surpassing the output of low pass branch and high pass branch of previous
existing image compression standards like the Joint stage to form a full tree decomposition. The new basis function
Photographic Experts Group (JPEG Algorithm) ...' called wavelet packets are defined as follows:
This paper presents new wavelet packets algorithms Let O(t) and v(t) be the scaling and wavelet functions
for image compressions. Extensive experiment respectively, which obey the two scale equations:
results prove that our techniques exhibit m
performance equal to, or in several cases superior to, $(t)=d2 L hke(2t-k) (1)
the current wavelet filters. k= -m
Wavelet Transform: m
Wavelet transform is modeling the signals by V(t)=d2 C grO(2t-k) (2)
combining translations and dilations of a simp!e k= -m
oscillatory function of finite duration called a wavelet.
Wavelet transform is used for analysis of the image at Note that the sequence (hk] and {gk]are the scaling and wavelet
different decomposition levels. These decomposition filter coefficients.
levels contain a number of subbands, which consists of Now let U,, (t)+(t) and ut (t)-v(t) and define
coefficients that describe the horizontal and vertical
spatial frequency characteristics of the input image.
Wavelet transform has assumed considerable practical
significance due to its linking with digital filtering. Its
filter bank structure is
m
Uh+t(t)=dZ z gk ud2t-k) (4)
k= .m
Taking the dyadic rescalings and translations of these functions
yields a library of functions (2-fl U. (2."-k)).
This library is over complete but a proper complete basis can be
found out selecting a subset of library with right set of
fig 1:Digital filter bank structure parameters (n. j, k ) .
Best Tree:
Wavelet Packets: One practical way of constructing wavelet bases is to iterate a
Despite of its general success, the wavelet transform two channel perfect reconstruction filter bank over the low pass
often fails to accurately capture high frequency branch. This results in a dyadic division of the frequency axis
information, especially at lower bit rates where such that works well for most signals we encounter in practice.
information is lost in quantisation noise. A technique However, for signals with strong stationary high pass
have been developed called wavelet packets that is components, it is often argued that wavelet basis will not
better able to represent high frequency information. perform well and improvements can be generally found it will
Wavelet packets are the conventional wavelet search over all possible binary trees for a particular filter set,
transforms in which the details are iteratively filtered. instead of using the fixed wavelet [Link] produced by
these arbitrary subband trees are called wavelet packet
expansions, and mathematical groundwork of wavelet packets
Wavelet packet transforms are obviously more general than
wavelet transform we, they also offer a rich library bases from
which the best one can be chosen for a fixed signal with a fixed
cost function. The advantage of the wavelet packet framework is
thus its universality in adapting the transform to a signal without
training or assuming any statistical property of the signal.
0-7803-7651-WO3617.W 0 2W3 IEEE
uthorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 17,2026 at 07:42:12 UTC from IEEE Xplore. Restrictions apply
., and Video Coding /923
lmaae
The extra adaprivity of the wavelet packet framework is
obtained at a price of added computation in searching Best Tree (Basis Bit Rate)
for the best wavelet packet basis, so an dfficient fast 1. Start.
search algorithm is the key in application. involving 2. Decompose the original image into four subhands.
wavelet packet. 3. Decide the threshold psnr.
4. Now decompose one suhband at a time.
Basis selection process: 5 . Reconstruct the image and calculate its psnr.
One reasonable criterion for selecting decomposition 6. Adjust the bits of the decomposed bands so that the
for the compression of the image is the additive cost threshold psnr is achieved
function. Since most transform-based compression a) Increment if the psnr is less than the threshold
schemes work by truncating or thresholding the small b) Decrement if the psnr is greater than the
coefficients to zero, a cost function that maximizes the threshold
number of near-zero values is a reasonable criteria for 7. Are all the four subbands decomposed one by one?
selecting a “hest”, from a compression viewpoint, No: go to step 4
decomposition. Yes: go to step 8
The optimization algorithm must use the function to 8. Decompose further the subband which requires
minimize the “cost” of decomposition tree’s leaf nodes. minimum average hits if decomposed
Minimal cost leaf nodes should be favored because they 9. Check if all the levels are reached
have more near zero values, which leads to greater No: go to step 4.
compression. Yes: go to step 10.
Selection of best basis for any particular image may be 10. stop.
performed in number of ways. Here we suggest the use
of an additive cost function like rate distortion and Quantizer:
number of bits required to represent all the coefficients The largest factor effecting wavelet coding compression and
in the node. reconstruction error is coefficient quantization. Most widely
used quantizers are uniform quantizers.
Best Tree (Basis Rate Distortion): Quantization operation can be described as
PSNR is the measure of rate distortion and is defined as y = Q (f). (6)
(peak signal to noise ratio) wherey= [yl,y2,......yL].
PSNR (dB) For the uniform quantizer
peak absolute squared value of signal Yk+I-Yk=A ,
=IOlog i.e. for uniform quantizatian step size is constant
mean squared error The difference between f(x,y) and y is the quantization error.
(5) Performance of the quantizer can be compared, using
Here we propose a method that attempts to select the quantization error, which is given as.
set of terminal nodes that are optimum in rate distortion L tk+l
sense. The approach involves: d= C (f-y)z Pdf) df (7)
K=l tk
1. start If the quantization error is greater than step size it is called as
2. Decompose the original image into four overload error. If the quantization error is less than the step size
subhand. it is granular error. Total error is the sum of granular error as
3. Now decompose one subband at a time. well as overload error, which depend on step size. For the given
4. Reconstruct the image and calculate its psnr. probability function step size can be selected such that mean
5 . Are all the four subbands decomposed one by square error gets minimized. Such quantizer is optimal
one. quantizerit is also called as Lloyd -Max quantizer.
No: go to step 3 Centroid quantization is based on the statistical properties of the
Yes: go to step 6 image. Results show that centroid method gives less
6 . Decompose the subbands, which gives -
quantization error as compared to Lloyd [Link] centroid
maximum psnr further. method is more accurate as compared with Lloyd -Max.
I . Check if all the levels are reached Although the above mentioned quantizers are not adaptive,
No: go to step 3. much can be gained from adjusting the quantization levels based
Yes: go to step 8. on the local behavior of an image. In theory, slowly changing
8. stop region can be finely quantized, while the rapidly changing areas
are quantized more coarsely. This approach simultaneously
reduces both granular noise and slope overload, while requiring
only a minimal increase in code rate .The trade off is increased
quantizer complexity.
8 2M)3 IEEE
0-7803-7651-X/03/$17.W
uthorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 17,2026 at 07:42:12 UTC from IEEE Xplore. Restrictions apply
TENCON 2003 /924
Lloyd Max quantizer:
1. start Level 1:diff avg. bits & quantization algos
2. Find max and .min of the matrix to be
quantised. Fig: Eliza
3. Calculate the range =max-min. F
4. Calculate step size=range/(Zb-l) (where b = no Avg Psnr (dB) Psnr (dB)
of bits i.e. b bit quantisation). Bits (Centroid (Lloyd-max method)
5 . Divide each element to be quantised by step method)
size. 28.0251 26.4049
6. Multiply each element by the step size to get 33.4813 33.3908
deauantisation. 38.8442 39.9509
7. stop ig: Cman (cameraman)
Centroid quantizer: Avg Psnr(dB) PSN(dB)
1. Start
Bits (Centroid ' (Lloyd-max method)
2. Find max and min of the matrix to be
method)
quantised.
3 29.0815 27.7584
3. Calculate the range =max-min.
4 34.6338 34.2202
4. Calculate step sizerange/(Zb-l) (where b = no
5 40.5224 40.4595
of bits i.e. b bit quantisation).
5. Divide each element to .be quantised by step 6 46.4840 46.2710
size.
6. Calculate mean of all the values in a particular
level of the original image. Avg Psnr(dB) Psnr (dB)
7. Assign the mean value to the corresponding Bits (Centroid (Lloyd-max metbod)
level in quantised matrix to get dequantisation. method)
8. Stop 3 28.3678 26.6283
4 34.1333 32.4478
Future advancements: 5 42.6663 39.8216
Multi wavelets provide one alternative to the
wavelet transform and wavelet packet transform.
Best tree based on PSNR
Original image is decomposed into four subbands.
Further decomposition of these subbands, one
decomposed at a time, gives following results.
As decomposed HH gives maximum PSNR, only it is
decomposed in level 2.
Four sub bands of HH are decomposed further
one at a time, and following results are
Obtained
2003 IEEE
0-7803-7651-x/03/$17.00~
uthorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 17,2026 at 07:42:12 UTC from IEEE Xplore. Restrictions appl
Image and Video Coding 1925
erli-
Wav (dB)
elets Only LL of HH is decomposed 34.3743
Only HL of HHis decomposed 34.3875
3, ha, 0~~l.H of HHis dccomposed I 31.4312
elel Only HH of HHis decomposed I 31 4176
Transform-Bopardikar
4. Sites: [Link]
5. [Link]
6. Getting started with matlabRudra Pratap.
7. Zixiang Xiong, Kannan Ramachandran. Michaen
Orchard-Wavelet packet image coding using space
frequency quantisation-IEEE Transactions on image
processing-Vol.7. no.6.
Thus the best tree selected is as:
r--
Original image
Level1 avgbit=4 Level2 avgbit=4
Level3 avgbit=4 Best avgbit =4
Fig: Wavelet packet decomposition
(Level 2 decomposition of LL)
References:
1. Gonzales and Woods-Digital image
processing-PHI publication.
2003 IEEE
0-781~3-7651-x/03/$17.000
uthorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 17,2026 at 07:42:12 UTC from IEEE Xplore. Restrictions apply
TENCON 2003 / 926
8. ,Michael Martin and Amy Bell-New image
compression techniques using multiwavelets &
multiwavelet- packets-IEEE Transactions on image
processing-Vol. 10,110.4.
0 2W3 IEEE
0-7803-765l-x103/~17.00
uthorized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on February 17,2026 at 07:42:12 UTC from IEEE Xplore. Restrictions appl