Int. J. Electron. Commun.

(AEÜ) 73 (2017) 183–192

Contents lists available at ScienceDirect

International Journal of Electronics and

Communications (AEÜ)
journal homepage: www.elsevier.com/locate/aeue

Regular paper

An improved wavelet-based image coder for embedded greyscale and

colour image compression
Tahar Brahimi a, Farid Laouir a, Larbi Boubchir b,⇑, Arab Ali-Chérif b
a
L2EI Research Lab., Department of Electronics, Jijel University, BP 98 Ouled Aissa, Jijel 18000, Algeria
b
LIASD Research Lab., University of Paris 8, 2 rue de la Liberté, 93526 Saint-Denis, France

a r t i c l e i n f o a b s t r a c t

Article history: The embedded zero-tree wavelet (EZW) coding algorithm is a very effective technique for low bitrate still
Received 8 March 2016 image compression. In this paper, an improved EZW algorithm is proposed to achieve a high compression
Accepted 11 January 2017 performance in terms of PSNR and bitrate for lossy and lossless image compression, respectively. To
reduce the number of zerotrees, the scanning and symbol redundancy of the existing EZW; the proposed
method is based on the use of a new significant symbol map which is represented in a more efficient way.
Keywords: Furthermore, we develop a new EZW-based schemes for achieving a scalable colour image coding by
Embedded greyscale and colour image
exploiting efficiently the interdependency of colour planes. Numerical results demonstrate a significant
coding
Lossy/lossless image coding
superiority of our scheme over the conventional EZW and other improved EZW schemes with respect to
EZW algorithm both objective and subjective criteria for lossy and lossless compression applications of greyscale and col-
Zerotree coding our images.
Wavelet transform Ó 2017 Elsevier GmbH. All rights reserved.

1. Introduction most complex techniques but also extremely fast in execution. In

Digital images are widely used in computer applications.
Uncompressed digital images require considerable storage capac-
ity and transmission bandwidth. Efficient image compression solu-
tions are becoming more critical with the recent growth of data
intensive, and multimedia-based web applications. The data com-
pression have a goal to reduce the volume of necessary data to rep-
resent a certain information quantity. Transform coding is a well-
known and widely-used technique in image compression. The pur-
pose of the transformation serves to produce decorrelated coeffi-
cients and remove redundancy [1,2].
Through the last decade, the discret wavelet transform (DWT)
has matured and has become very effective for the compression
of digital images. The statistical properties of the wavelet trans-
form have been studied extensively and have been exploited by
wavelet-based image coders [3–9]. In particular, the pyramidal,
or dyadic, wavelet decomposition has shown excellent energy
compaction, which is helpful in image compression [10,11,3,13,1
2,2,14–16,1,17,18]. Among the wavelet based coding schemes,
embedded zerotree coding (EZW) by Shapiro [16] was the first to
apply an embedded zerotree using wavelet for image compression
[30]. EZW was not only competitive in its performance with the
⇑ Corresponding author.
E-mail address: larbi.boubchir@ai.univ-paris8.fr (L. Boubchir).
addition, it also has several other merits: it requires no pre-
stored tables, no training, and produces fully embedded codes. This
latter property enables one to send images in a progressive manner
and to encode images at any target bitrate [16]. The good perfor-
mance of EZW is based on four concepts: (1) subband decomposi-
tion of an image by using the discrete wavelet transform, (2)
prediction of abscence of significance information across scales
by exploiting the self similarity in images, (3) successive approxi-
mation quantization, and (4) adaptive arithmetic of symbols.
During the coding process of original EZW, it was observed that
a noticeable number of significant coefficients have insignificant
descendants when compared with significant coefficients whose
descendants are significant. As a consequence, many zerotree sym-
bols are generated by EZW to encode the insignificant descendants
of significant coefficients, which are useless to be encoded again.
So in order to improve the lossy compression performance of the
original EZW, a new enhanced algorithm, namely IMP1EZW, is pro-
posed. The proposed algorithm is a modification of basic EZW algo-
rithm making it efficient for image compression applications. The
key idea in IMP1EZW is the introduction of a new test for signifi-
cant wavelet coefficients (Swc), which allows to determine if Swc
has or no significant children. Therefore, a new data structure of
coding significant coefficient is introduced. This data consists of
six symbols instead of the four symbols existing in the original
EZW. Four symbol deal with significant coefficients in order to

http://dx.doi.org/10.1016/j.aeue.2017.01.008
1434-8411/Ó 2017 Elsevier GmbH. All rights reserved.
184 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192

avoid the encoding of the zerotree of Swc and generate fewer bits ordinate list, and their positions in the image are filled with zeroes.
for a given data, saving many bits, by reducing the number of over- This is to ensure that they will not be coded more than once. The
all zerotrees, and making IMP1EZW a simple, fast and efficient second pass, the subordinate pass, is the refinement pass whose
extension of EZW to lossy image coding. Furthermore, an adapta- objective is to refine the representation of the magnitude of those
tion of IMP1EZW is also presented, to improve its lossless coding significant coefficients stored in the subordinate list. After subordi-
performance. nate pass is finished, the current threshold is halved. A new round
EZW and SPIHT (Set Partionning In Hierarchical Trees) [15] of dominant/subordinate coding is operated. This iterative proce-
were primarily developed for greyscale images. A simple applica- dure continues until a desired bitrate is reached up or a predefined
tion to colour images consists of coding each wavelet transformed distortion objective is met. EZW technique can be enhanced using
colour plane separately. However, this approach requires alloca- entropy coding before transmission, to achieve further
tion of bits. It does not also keep the full embeddedness and the compression.
precise rate control of the coder, since the decoder needs to wait
until the full bitstream arrives to reconstruct the colour images.
3. Proposed improvement to EZW for greyscale images
In [20,21] a colour embedded zerotree wavelet, called CEZW [22],
has been proposed using the YUV colour space. It uses new spatial
EZW algorithm is the first successful coding scheme developed
orientation tree (SOT) for luminance coefficients (Y) to take advan-
for the wavelet transform and has attracted great attention and is
tage of the interdependence of chrominance planes (U and V). In
widely used in a number of applications [27,24,28]. Many wavelet
this paper, we propose also two fully embedded colour EZW. The
transform coding schemes developed afterward [15,23,3,13] were
first one, which is the extension of IMP1EZW to colour images, uses
more or less influenced by the idea of zerotree data structure. To
the same SOT of conventional EZW. It treats all colour planes,
reduce complexity and increase the compression performance,
already separated and decomposed in the wavelet domain, as
we propose an improved EZW image coding algorithm (IMP1EZW)
one unit during the coding stage to generate one mixed bit-
which can be viewed as an extension of the EZW coding technique.
stream. The second one combines the SOT proposed in [20] with
Experimentally, it has been noticed during the EZW coding of
IMP1EZW to efficiently exploit the interdependencies in chromi-
different test images that there is a considerable number of signif-
nance planes.Experimental results show that our algorithms
icant wavelet coefficients (Swc) whose descendants are insignifi-
achieve better coding performance than the conventional EZW,
cant noted (Zswc) compared to those coefficients whose
and other improved EZW coder [23,20,21,24] for both greyscale
descendants are found to be significant (Iswc) (see Table 1). Conse-
and colour images.
quently, EZW consumes an important amount of bits to encode the
The paper is organized as follows. The EZW algorithm is briefly
descendants of Zswc, as zerotree symbols, although they are not
explained in Section 2. Then, a description of the proposed modifi-
significant [3]. Their coding as zeortree is mainly due to the fact
cation of EZW scheme for greyscale images is given in Section 3.
that they have significant parents (positive or negative).
The lossless scheme is presented in Section 4. Section 5 provides
Based on this observation, and in order to avoid the coding of
a brief description of embedded colour image compression process,
the zerotree found from Swc, we suggest adding a new test on
whereas Section 6 describes our proposed embedded colour EZW
Swc. Therefore, a new data structure is defined to improve the
schemes. Numerical results are presented in Section 7. The last sec-
compression of significance maps of wavelet coefficients by reduc-
tion concludes the paper.
ing the number of zerotree [3], and hence increases the number of
coefficients not to be encoded, as follows: Sa: if wavelet coefficient
is positive significant but all its descendants are insignificant. Sb: if
2. Review of original EZW algorithm
wavelet coefficient is negative significant but all its descendants
are insignificant. Sc: if wavelet coefficient is positive significant
After performing the discret wavelet transform, image coding is
and has at least one descendant that is significant. Sd: if wavelet
carried out with the embedded zerotree wavelet coder. Since its
coefficient is negative significant and has at least one descendant
publication, EZW has attracted a lot of attention and has been
that is significant. The proposed method is similar to Shapiro’s
widely used in many applications [25,26]. The design idea evolve
EZW coding in the dominant pass but it has six symbols mostly
from the empirical true hypothesis that if a wavelet coefficient at
dealing with significant coefficients.
a coarse scale is insignificant with respect to a threshold T, then
We develop a modified version called IMP1EZW which is per-
all wavelet coefficients of the same orientation in the same spatial
formed in two passes in a recursive manner, as in EZW coding. In
location at finer scale are likely to be insignificant with respect to T.
the dominant pass, when a coefficient is found to be significant,
The first step in the EZW coding algorithm is to determine the ini-
all its descendants must also be tested [3]. Then four symbols:
tial threshold T 0 which is the largest as power of two less than or
Sa; Sb; Sc or Sd, are introduced to encode this coefficient as indi-
equal to the largest coefficient.
cated previously. Note that the symbols Sa; Sb, use two bits for
During the encoding procedure of EZW, two distinct operations
encoding the coefficients in the region representing the detail at
are performed alternatively for each threshold. They are called the
the highest frequency subbands (HL, LH and HH), since there is
Dominant Pass and the Subordinate Pass. In the first pass, the dom-
no child. Fig. 1 depicts the block diagram of the dominant pass of
inant pass, the wavelet coefficient matrix is scanned and a symbol
IMP1EZW.
is output for every coefficient. EZW’s significance map coding
In fact, this modification simplifies the coding technique as well
requires absolutely no training, no pre-stored table, and no prior
as reduces the number of zerotrees and increases the number of
knowledge of the image source. Only the following four symbols
coefficients not to be coded, speeds up the process and improves
are used to describe the significance map coding [16]: ZTR: the
the coding efficiency.
coefficient is classified as a zerotree Root (ZTR), which itself and
all its descendants are insignificant with respect to the current
threshold. IZ: if the wavelet coefficient is insignificant and has at 4. Adaptation of IMP1EZW for lossless image compression:
least one descendent that is significant. Pos: if coefficient is a sig- IMP2EZW
nificant positive. Neg: if coefficient is a significant negative. Finally,
all the coefficients that are in absolute value larger than the current Experimentally, it has been noticed during IMP1EZW coding of
threshold are extracted and placed without their sign on the sub- different test images that the number of the coefficients Iswc
 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192 185

Table 1
Rate (%) of significant coefficients whose descendants are insignificant (Zswc) and significant coefficients whose descendants are significant (Iswc) at different compression ratios
for several greyscale test images using a 9/7 wavelet transform.

Compression ratio 4 8 16 32 64 128

Airplane Zswc 70.2105 76.0133 77.7247 75.4973 75.1279 75.5857
Iswc 29.7895 23.9867 22.2753 24.5027 24.8721 24.4143
Boat Zswc 61.9562 68.2379 72.2192 72.5440 74.2688 80
Iswc 38.0438 31.7621 27.7808 27.4560 25.7312 20
Lena Zswc 60.5401 74.6552 79.4872 80.0446 81.3992 79.1434
Iswc 39.4599 25.3448 20.5128 19.9554 18.6008 20.8566
Peppers Zswc 52.1566 70.4730 78.5957 79.1900 79.2046 76.0355
Iswc 47.8434 29.5270 21.4043 20.8100 20.7954 23.9645
Barbara Zswc 65.3610 72.4036 76.7840 81.1349 76.6367 81.2756
Iswc 34.6390 27.5964 23.2160 18.8651 23.3633 18.7244
Goldhill Zswc 56.4507 65.1203 69.8074 73.8161 78.0986 74.1489
Iswc 43.5493 34.8797 30.1926 26.1839 21.9014 25.8511
Sailboat Zswc 61.0062 69.2376 72.8032 71.4531 72.1227 78.3224
Iswc 38.9938 30.7624 27.1968 28.5469 27.8773 21.6776

Bold values means that the highest rates of Iswc are at the compression ratio.

Sc IMP2; Sd IMP2: the number of positive significant and nega-

tive significant coefficients found to be in zone 1 which have at
least one descendant that is significant of IMP2EZW.
S IMP2: is the sum of symbols Sa1 IMP2; Sb1 IMP2; Sc IMP2
and Sd IMP2 found to be in zone 1 of IMP2EZW.
Since the number of zerotree in IMP2EZW is reduced, the gain
of zerotree (Gain2 ztr) is given by:

Gain2 Ztr ¼ ðZt1 EZW  Zt1 IMP2Þ  2 þ ðZt2 EZW

 Zt2 IMP2Þ ð1Þ
In view of the fact that the sum of significant coefficient in zone
1 of EZW is equal to the sum of significant coefficients in zone 1 of
IMP2EZW, we have the following equation:

ðP1 EZW þ N1 EZWÞ ¼ ðSa1 IMP2 þ Sb1 IMP2 þ Sc IMP2

Fig. 1. Flow chart for encoding a coefficient in IMP1EZW scheme.
þ Sd IMP2Þ þ ðP1 IMP2 þ N1 IMP2Þ ð2Þ

(Sc þ Sd) increases compared to the number of coefficients Zswc Note that since the symbols
(Sa þ Sb) as the threshold decreases. On the other hand, the sum P1 EZW; N1 EZW; P1 IMP2 þ N1 IMP2 are coded by 2 bits and
of Sa; Sb; Sc and Sd symbols surpasses the gain of zerotree at a cer- the symbols Sa1 IMP2; Sb1 IMP2; Sc IMP2; Sd IMP2 are coded by
tain value of the threshold. Consequently, IMP1EZW consumes an 3 bits, IMP2EZW loses
important amount of bits to encode the symbols Sa; Sb; Sc and ðSa1 IMP2 þ Sb1 IMP2 þ Sc IMP2 þ Sd IMP2Þ bits on the encoding
Sd. One way to improve the performance of IMP1EZW is to switch of these coefficients.
to the original EZW algorithm at this threshold value. The com- The gain G of IMP2EZW over EZW for lossless image coding is as
bined algorithm which uses IMP1EZW and then switches to EZW follows:
is referred to as IMP2EZW. G ¼ Gain2 Ztr  S IMP2 ð3Þ
The condition of switching IMP1EZW to EZW, is given by calcu-
lating the following value:
4.1. Gain of IMP2EZW over EZW for lossless image coding
a ¼ ðZt1 EZWit  Zt1 IMP1itÞ  2 þ ðZt2 EZWit
We denote by ‘zone 1’ the region that contains approximation  Zt2 IMP1itÞ  Sit ð4Þ
and detail subbands which have a decomposition level higher than
1. On the other hand, the remaining part which includes the detail with: (Zt1 EZWit; Zt2 EZWit): the numbers of EZW’s zerotree
subbands (HL, LH and HH) of the first level of decomposition is located in zones 1 and 2 at a threshold value defined as:
referred to as ‘zone 2’. T i ¼ 2ki ði ¼ 0; . . . ; kÞ, with k ¼ floorðlog2 ðmaxðabsðC ij ÞÞÞÞ. Where C ij
Let us denote by: is the wavelet coefficient, and floorðxÞ rounds the elements of x to
(Zt1 EZW; Zt2 EZW), (Zt1 IMP2; Zt2 IMP2): are the numbers of the nearest integers towards minus infinity.
zerotree of EZW and IMP2EZW located in zones 1 and 2, Zt1 IMP1it; Zt2 IMP1it: the numbers of IMP1EZW’s zerotree
respectively. located in zones 1 and 2 at a threshold value T i .
P1 IMP2; N1 IMP2: the number of positive significant and neg- Sit: is the sum of Sa1 IMP1; Sb1 IMP1; Sc IMP1 and Sd IMP1 at
ative significant coefficients found to be in zone 1 of IMP2EZW. threshold value T i .
Sa1 IMP2; Sb1 IMP2: the number of positive significant and Sa1 IMP1; Sb1 IMP1: the number of positive significant and neg-
negative significant coefficients found to be in zone 1 where all ative significant coefficients found to be in zone 1 where all their
their descendants are insignificant of IMP2EZW. descendants are insignificant of IMP1EZW.
186 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192
Sc IMP1; Sd IMP1: the number of positive significant and nega-
tive significant coefficients found to be in zone 1 which have at
least one descendant that is significant of IMP1EZW.
The Switching process from IMP1EZW to EZW is performed if
the parameter a is negative (usually at threshold T = 4).
5. Embedded colour image compression
EZW was previously developed for greyscale images. A simple
approach to code colour images is to treat the three colour planes
as three greyscale images and the same coding technique is then
used for each plane. Although RGB to luminance and chrominance
(LC) colour transformations are employed to reduce the correlation
found to be in colour images, there are some redundancies that can
further be investigated in LC space [19,29]. These redundancies
occur for those coefficients within chrominance planes at the same
spatial location which provide the same significance test result
when evaluated against a given threshold. That is, if a wavelet
luminance and its descendants in luminance plane are found to
be insignificant according to a given threshold, chrominance coef-
ficients at the same location are also likely to be insignificant. In
[20,21] a spatial orientation tree have been proposed to exploit
the interdependence between the colour planes, known as the col-
our embedded zerotree wavelet (CEZW) [22] as illustrated by
Fig. 2. In Table 2, we provide the percentage of insignificant lumi-
nance wavelet coefficients which do not have any significant
descendants, including those in luminance and chrominance
planes, denoted as P Ztr in Table 2. Note that the descendants of
a luminance coefficient in chrominance planes are at the same
location. Let denote by P Iz as the percentage of luminance coeffi-
cients which have at least one significant descendant (in lumi-
nance or chrominance planes). It can be seen easily form Table 2
that the chrominance coefficients show high similarities with
luminance coefficients at the same location since the percentage
of zerotree is higher than that of the percentage of Iz symbols.
6. Proposed embedded colour EZW coders: CCIMPEZW and
CIMPEZW
We refer to CCIMPEZW the proposed EZW coder, described in
Section 3, applied to colour images using the same SOT described
in conventional EZW. The chrominance planes are scanned after
Fig. 2. Diagram of parent descendent relations in [20].
the luminance plane is scanned during dominant pass. Therefore,
a one mixed bit-stream is generated. The basic idea of the second
proposition of EZW coder for colour images, namely CIMPEZW is
to apply the SOT described previously [20,21] to luminance coeffi-
cient found to be insignificant during the dominant pass coding
stage. CIMPEZW completes two passes (dominant pass and subor-
dinate pass) to encode a given colour image. In the dominant pass,
the luminance is first scanned. Six symbols (Sa; Sb; Sc; Sd; Ztr and
IZ) are used to encode a given coefficient. For each luminance coef-
ficient a significant test is performed against the current threshold.
If the luminance coefficient is significant, only its descendants in
luminance plane are tested. If its descendants are insignificant
the symbol Sa (or Sb) is assigned if the coefficient is positive (or
negative) respectively. Otherwise, the symbol Sc is attributed for
positive coefficient and Sd for negative coefficient. On the other
hand, if the luminance coefficient is found to be insignificant, all
its descendants, in luminance and chrominance planes, should be
tested. If they are all insignificant, the zerotree symbol is assigned.
Otherwise, the symbol Iz is transmitted. Once, the luminance plane
is scanned, the two chrominance planes are alternately scanned.
Note that coefficients of chrominance plane previously encoded
as zerotree symbols while scanning the luminance are not tested.
The subordinate pass is identical as that in conventional EZW.
Practically, the magnitude of luminance coefficients is lager
than those in chrominance planes. Therefore, the chrominance
components are examined only if their corresponding threshold
is greater than, or equal to, the current threshold.
7. Simulation results
7.1. Progressive lossy compression of greyscale image
To assess the performance of the proposed technique, numeri-
cal experiments have been carried out on a number of natural
images of size 512  512, each sample is quantized at 8 bits per
pixel (bpp). A 6-level dyadic wavelet decomposition of the images
is performed using the biorthogonal 9/7 wavelet filters. For a fair
comparison, the results are of binary coded versions. The proposed
EZW-based coder has been compared against the conventional
EZW and its improved version reported in [23] while using the
same wavelet transform. The results obtained are reported in
Table 3. The gains of IMP1EZW over EZW and IEZW are denoted
by G1 and G2, respectively.
As can be shown from Table 3, our proposed method outper-
forms baseline EZW for all bit-rates on the test images. At the same
compression ratio, the PSNR values of the reconstructed images
with IMP1EZW technique are higher than those for EZW. Clearly,
the improvement over the standard EZW is significant, for exam-
ple, the gain reaches 2.37 dB for ‘Airplane’, 2.25 dB for ‘Barbara’,
1.82 dB for ‘Peppers’ and 1.62 dB for ‘Lena’. Moreover, the experi-
mental results show that the performance of IMP1EZW coder sur-
passes that of IEZW coder [23] over a wide range of bit rates for
different images. The improvement is notably considerable for sev-
eral bitrates, as an example at the bitrates 1, 1, 0.25 and 1 bits per
pixel (bpp), using IMP1EZW will elevate the PSNR by 2.23 dB for
‘Barbara’, 2.21 dB for ‘Airplane’, 1.51 dB for ‘Peppers’ and 1.47 dB
for ‘Lena’ respectively. Overall, the improvements obtained over
EZW for greylevel images vary between 0.42 dB and 2.37 dB.
Fig. 3 shows the decoded ‘Cameraman’ and ‘Boat’ images of size
256  256 for EZW and IMP1EZW algorithms at 0.25 bpp. In this
subjective test, we compare the perceptual quality between the
two coders. It can be observed that the visual quality of the images
in Figs. 3-(b) and -(d) is better than that in Figs. 3-(a) and-(c).
It should be noted that IMP1EZW reduces the number of zero-
trees and consequently increases the number of coefficients not
 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192 187

Table 2
Percentage(%) of Ztr and Iz symbols for several threshold values during the coding of insignificant luminance coefficients.

Lena Peppers Airplane

Threshold P Ztr P Iz P Ztr P Iz P Ztr P Iz
8192 100 0 100 0 100 0
4096 100 0 98.9362 1.0638 100 0
2048 100 0 100 0 99.4819 0.5181
1024 98.5075 1.4925 93.6170 6.3830 96.6981 3.3019
512 94.3020 5.6980 91.5709 8.4291 89.2157 10.7843
256 93.2377 6.7623 90.9242 9.0758 89.0158 10.9842
128 89.5611 10.4389 90.6761 9.3239 89.0385 10.9615
64 91.1486 8.8514 90.9061 9.0939 89.0501 10.9499
32 90.4791 9.5209 90.7003 9.2997 90.1426 9.8574
16 86.6314 13.3686 84.9721 15.0279 86.9432 13.0568

Bold values mean that the percentage of Ztr symbols is higher than that of the percentage of Iz symbols at all estimated threshold values.

Table 3
Lossy performance comparison of IMP1EZW, IEZW [23] and EZW image codecs for several greyscale test images. G1 and G2 are, respectively, the gains of the proposed coder
IMP1EZW over EZW and IEZW coders.

Image Bitrate (bbp) EZW IEZW [23] IMP1EZW G1 G2

Airplane 0.03125 23.0369 23.8403 23.8106 0.7737 0.0297
0.0625 25.2801 25.44 25.7746 0.4945 0.3346
0.125 27.0272 27.5713 28.1217 1.0945 0.5504
0.25 29.8712 30.3278 31.2195 1.3483 0.8917
0.5 33.5873 33.7779 34.9951 1.4078 1.2172
1 37.9931 38.1555 40.3689 2.3758 2.2134
Boat 0.03125 22.0179 22.3037 22.5333 0.5154 0.2296
0.0625 23.6797 24.2149 24.6006 0.9209 0.3857
0.125 25.3461 26.0021 26.3312 0.9851 0.3291
0.25 27.7107 28.0317 28.7099 0.9992 0.6782
0.5 31.0003 31.1332 31.9414 0.9411 0.8082
1 35.168 35.2571 36.0863 0.9183 0.8292
Lena 0.03125 23.1295 23.6435 23.8158 0.6863 0.1723
0.0625 25.1688 25.8174 26.2563 1.0875 0.4389
0.125 27.4979 28.1392 28.9792 1.4813 0.8400
0.25 30.1838 30.69 31.6442 1.4604 0.9542
0.5 33.4046 33.6187 34.7579 1.3533 1.1392
1 37.0065 37.1582 38.6311 1.6246 1.4729
Peppers 0.03125 23.5973 24.3384 24.5045 0.9072 0.1661
0.0625 26.0735 26.5059 27.0365 0.9630 0.5306
0.125 28.3923 28.8523 29.8714 1.4791 1.0191
0.25 31.1927 31.4977 33.0125 1.8198 1.5148
0.5 34.4034 34.5392 35.7291 1.3257 1.18998
1 37.0137 37.1188 38.3256 1.3119 1.2068
Barbara 0.03125 20.7364 20.8728 21.1572 0.4208 0.2844
0.0625 21.8887 22.1424 22.4375 0.5488 0.2951
0.125 22.9503 23.1445 23.4877 0.5374 0.3432
0.25 24.2573 24.3393 25.925 1.8400 1.7809
0.5 27.1892 27.2483 29.0292 1.3257 1.1899
1 31.5741 31.5936 33.8263 2.2522 2.2327
Goldhill 0.03125 22.656 23.5243 23.5271 0.8711 0.0028
0.0625 24.4855 24.8229 25.1736 0.6881 0.3507
0.125 26.0731 26.4052 26.9261 0.8530 0.5209
0.25 28.0296 28.2716 28.7573 0.7277 0.4857
0.5 30.3132 30.5538 31.1567 0.8435 0.6029
1 33.6686 33.7536 34.4445 0.7759 0.6909
Sailboat 0.03125 21.5291 21.8738 21.9762 0.4471 0.1024
0.0625 23.031 23.602 23.8868 0.8558 0.2848
0.125 24.6114 25.2835 25.6648 1.0534 0.3813
0.25 27.12 27.3304 28.0411 0.9211 0.7107
0.5 29.896 30.2623 30.8656 0.9696 0.6033
1 33.3393 33.6265 34.1861 0.8468 0.5596

Bold values indicate the best performances obtained in terms of PSNR (dB).

to be encoded. Furthermore, it increases the number of significant is, the sets of distortion measures are acquired from the same file.
coefficients which provides a better reconstruction. It is worth In other words, the decoder reads the first bytes of the file up to the
pointing out that the coding results obtained here for the three desired bit rate, performing the decoding and inverse transforma-
coders are based on the progressive transmission property, that tion, and then compares the recovered image with the original.
188 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192

Fig. 3. Subjective evaluation of IMP1EZW and EZW algorithms for ‘Cameraman’ and ‘Boat’ images. (a) Reconstructed image by Original EZW with PSNR of 23.4095 dB at
0.25 bpp. (b) Reconstructed image by IMP1EZW with PSNR of 24.3672 dB at 0.25 bpp. (c) Reconstructed image by original EZW with PSNR of 24.0212 dB at 0.25 bpp. (d)
Reconstructed image by IMP1EZW with PSNR of 25.0663 dB at 0.25 bpp. (e) A enlarged part of reconstructed ‘Cameraman’ image by EZW. (f) A enlarged part of reconstructed
‘Cameraman image’ by IMP1EZW.

0 1
7.2. Progressive lossy compression of colour image 2552
PSNR ¼ 10 log10 @
A ð4Þ
MSEðYÞþMSEðIÞþMSEðQÞ
In the following, we present a set of coding results operating on 3
five colour images of size 512  512 with 24 bpp: ‘Airplane’,
‘Baboon’, ‘Lena’, ‘Peppers’ and ‘Sailboat’. Six decomposition level where MSE is the mean square error.
were performed using 9/7 biorthogonal filters. For a fair evaluation The results are reported in Table 4 and Fig. 4. Obviously, Table 4
of the performance, binary coded version of conventional colour shows that CCIMPEZW outperforms significantly conventional col-
EZW (CCEZW), CEZW [20,21], and proposed techniques CCIMPEZW our EZW in all test cases. The coding gain exceeds 0.7 dB in the
and CIMPEZW are used, in YIQ colour space, for the purposes of range of bitrates [0.125–1] bpp of the vast majority of the test
comparison. The overall PSNR is obtained by: images. Moreover, CCIMPEZW achieves a gain more than 1 db for
 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192 189

Table 4
Lossy coding performance comparison of different embedded colour images coders. G3 and G4 represent, respectively, the gain of the proposed CIMPEZW colour coder over
CCEZW and CEZW colour coders.

Image Bitrate (bbp) CCEZW CEZW Proposed CCIMPEZW Proposed CIMPEZW G3 G4

Airplane 0.015625 19.0895 19.1425 20.2548 20.2805 1.19 1.14
0.03125 21.0216 21.128 21.8509 21.8987 0.88 0.77
0.0625 22.5446 22.6292 23.2606 23.2966 0.75 0.67
0.125 24.0172 24.1639 24.8814 24.9282 0.91 0.76
0.25 26.0847 26.1468 27.6148 27.7486 1.66 1.60
0.5 28.9971 29.056 30.7694 30.9455 1.95 1.89
1 32.5429 32.7409 34.2350 34.4004 1.86 1.66
Lena 0.015625 18.3272 19.999 20.3076 20.7426 2.42 0.74
0.03125 21.2298 21.7729 22.3324 22.6596 1.43 0.89
0.0625 23.2072 23.6418 24.2711 24.42 1.21 0.78
0.125 25.3461 25.683 26.4282 26.5266 1.18 0.84
0.25 27.4117 27.9234 28.6571 28.7462 1.33 0.82
0.5 29.9251 30.4754 31.0746 31.1601 1.23 0.68
1 32.4212 33.0934 33.4965 33.5582 1.14 0.46
Peppers 0.015625 17.1220 17.8597 18.3211 18.4234 1.30 0.56
0.03125 19.0736 19.7944 20.1309 20.3641 1.29 0.57
0.0625 20.8391 21.7564 22.1675 22.2822 1.44 0.62
0.125 22.7927 23.2696 24.3562 24.5546 1.76 0.53
0.25 25.1489 25.3662 26.7241 27.0071 1.86 1.64
0.5 27.7671 28.0852 29.1697 29.7432 1.98 1.66
1 30.4787 30.977 31.4955 31.9577 1.48 0.98
Baboon 0.015625 16.702 16.913 17.267 17.653 0.95 0.74
0.03125 17.9749 18.1208 18.38 18.5273 0.55 0.41
0.0625 18.7632 18.8897 18.9499 18.9912 0.23 0.10
0.125 19.3299 19.418 19.8037 19.8494 0.52 0.43
0.25 20.3858 20.4815 20.6648 20.6879 0.30 0.21
0.5 21.7085 21.8575 22.1712 22.2057 0.50 0.35
1 23.7006 23.8048 24.2855 24.3394 0.64 0.53
Sailboat 0.015625 16.7956 17.0018 17.6737 20.2805 3.48 3.28
0.03125 18.1385 18.4472 19.014 21.8987 3.76 3.45
0.0625 19.6013 19.8800 20.4825 20.5588 0.96 0.68
0.125 21.4530 21.4988 22.1691 22.2692 0.82 0.77
0.25 22.9719 23.1678 23.7578 23.8162 0.84 0.65
0.5 24.8138 25.2196 25.6861 25.7879 0.97 0.57
1 26.9894 27.5327 27.8622 27.9828 0.99 0.45

Bold values indicate the best performances obtained in terms of PSNR (dB).

Fig. 4. Coding performance assessment of proposed embedded colour image coders (CIMPEZW, CCIMPEZW), CEZW and CCEZW algorithms for ‘Airplane’, ‘Lena’, ‘Peppers’,
‘Baboon’ and ‘Sailboat’ colour images: (a) Average value of PSNR. (b) Average value of gains.

‘Airplane’, ‘Peppers’ and ‘Lena’ in the range [0.25–1] bpp. Particu-
larly, the gains are up by 1.77 dB and 1.98 dB at 0.5 bpp and the
very low bitrate for ‘Airplane’ and ‘Lena’ test images respectively.
On average, CIMPEZW makes a gain of [0.84–1.38] dB. The
improvements obtained in overall average PSNR over CEZW are
also significant and vary between 0.47 dB and 0.87 dB. The gain
in efficiency depends on the test images and the bitrate. In partic-
ular, CCIMPEZW outperforms CEZW up to 1.46, 1.71 and 1.49 dB
for ‘Airplane’ at the bitrates 0.25, 0.5 and 1 bpp, up to (1.09,1.36
1.08) dB, (0.74, 0.73, 0.6) dB at the range of bitrates (0.125,0.25,
0.5) bpp for ‘Peppers’ and ‘Lena’ respectively. The second proposed
technique CIMPEZW provides also significant improvements
where a gain outperforms 1.66, 1.95 and 1.85 dB for ‘Airplane’ at
0.25, 0.5 and 1 bpp when compared to CCEZW. The gain alternate
between 1.42 and 1.98 db in the range [0.06125–1] bpp for ‘Pep-
pers’. It up to 2.33 dB and 3.48 dB at low birtaes for ‘Lena’ and ‘Sail-
190 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192

boat’. CCIMPEZW stands to gain at least 1.18 dB for ‘Peppers’ and

Fig. 5. Performance comparison of proposed coders CIMPEZW and CCIMPEZW
against the method in [24] (in PSNR) for Lena colour image at the same range of
bitrates [0.06125,0.5](bpp).
‘Lena’ at all interest bitrates. If one refers to CEZW, the use of the
proposed technique CIMPEZW aims at improving significantly
the performance. In particular, CIMPEZW shows an improvement
of (1.60, 1.89 1.66) dB and (1.64 1.66 0.98) dB at the bitrates
(0.125, 0.25 0.25) bpp for ‘Airplane’ and ‘Peepers’, respectively.
The improvement is up to 3 dB for ‘Sailboat’ image at very low
bitrates. On average, CIMPEZW demonstrates an improvement in
terms of PSNR varying between 0.92 and 2.11 dB over CCEZW,
and from 0.55 to 1.6 dB over CEZW as shown in Table 4 (see gains
G3 and G4). The improvement varies according to the image con-
tent and the interest bitrate. It should be noted that CEZW shows
a medium average gain of [0.22–0.89] dB in comparison with
CCEZW. For ‘Airplane’ image, small gain was achieved less than
0.2 dB. Fig. 6 illustrates the decoded ‘Peppers’ image of size
256  256 compressed using CCEZW and CIMPEZW at 0.25 bpp
and 0.5 bpp. According to Fig. 5 we can notice some improvement
on the images quality showed in Figs.6-(b) and -(d). Overall, CCIM-
PEZW and CIMPEZW versions are a clear improvement on their
predecessors (CCEZW, CEZW). The results demonstrate clearly that
the interdependency between colour planes is better exploited by
CIMPEZW than that of CEZW. The proposed colour EZW coders

Fig. 6. Subjective assessment of CIMPEZW and CCEZW algorithms for ‘Peppers’ colour image: (a) Reconstructed image by CCEZW with PSNR of 22.2249 dB at 0.25 bpp. (b)
Reconstructed image by CIMPEZW with PSNR of 23.3912 dB at 0.25 bpp. (c) Reconstructed image by CCEZW with PSNR of 24.1628 dB at 0.5 bpp. (d) Reconstructed image by
CIMPEZW with PSNR of 26.2026 dB at 0.5 bpp.
 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192 191

Table 5
Coding and decoding time (in seconds) of different embedded coding techniques for common greyscale test images.

Image Bitrate (bbp) EZW IEZW [23] IMP1EZW

Airplane 0.125 3.3000 2.0400 3.1100
0.25 4.4500 3.2000 4.3200
0.5 7.7800 6.7600 7.7100
0.75 14.3900 12.9900 14.0100
Boat 0.125 3.2300 1.6500 2.9300
0.25 4.2300 2.7200 4.3100
0.5 7.7700 6.9700 7.8800
0.75 13.4200 11.9400 14.0700
Lena 0.125 3.2400 1.4900 2.9500
0.25 4.2300 2.4900 4.3100
0.5 7.8900 6.0500 7.6200
0.75 13.8900 12.9800 13.0900
Peppers 0.125 3.3000 1.6500 3.0400
0.25 4.2000 2.6700 4.1700
0.5 7.7300 6.0300 8.2600
0.75 13.5300 12.0900 13.1600
Barbara 0.125 3.2000 1.6500 2.8200
0.25 3.9200 2.4700 3.9600
0.5 7.3400 5.7300 7.5000
0.75 15.6300 12.9500 14.0300
Goldhill 0.125 3.2100 1.4800 2.8200
0.25 4.2300 2.5800 3.9100
0.5 8.0100 6.1100 7.2800
0.75 12.6900 10.8200 12.8100
Sailboat 0.125 3.0700 1.4000 2.7700
0.25 3.8400 2.3400 3.9200
0.5 7.0000 5.9300 7.3700
0.75 13.3500 11.9300 12.6600
Average 0.125 3.2214 1.6229 2.9200
0.25 4.1571 2.6386 4.0800
0.5 7.6457 6.2257 7.6600
0.75 13.8429 12.2429 13.4043

Table 6
Coding and decoding time (in seconds) of different embedded colour coding techniques for colour test images.

Image Bitrate (bbp) CCEZW CEZW CCIMPEZW CIMPEZW

Airplane 0.125 8.7700 9.1500 7.6900 11.2100
0.25 16.7600 25.4700 15.3200 18.9000
0.5 69.8900 84.2400 90.9800 93.3300
Lena 0.125 7.6600 6.9600 7.1400 9.4900
0.25 9.4000 10.1800 11.1700 15.5700
0.5 15.6600 23.3300 32.6600 44.6000
Peppers 0.125 10.4900 9.5100 9.4200 11.7100
0.25 21.9400 35.6800 36.0600 35.4000
0.5 131.7500 155.2500 147.2700 144.2500
Baboon 0.125 11.0600 11.3900 10.6600 12.8000
0.25 50.7700 52.8900 49.1900 52.8000
0.5 231.5800 244.8600 214.2300 217.3400
Sailboat 0.125 12.7900 14.9400 16.8900 17.3400
0.25 67.0600 73.2600 64.8200 69.8900
0.5 285.9000 300.4400 261.0200 264.0900
Average 0.125 10.1540 10.3900 10.3600 12.5100
0.25 33.1860 39.4960 35.3120 38.5120
0.5 146.9560 161.6240 149.2320 152.7220

CCIMPEZW and CIMPEZW have been assessed and compared with
a recent EZW colour image compression technique [24] of which
the results obtained on a single colour image ‘Lena’ are presented
in the form of a curve for the following range of bitrates [0.625–
0.5] bpp. Fig. 5 illustrates the results provided by the three codecs,
and assessed at the same bitrates. According to Fig. 5 we can con-
clude that both the proposed coders CCIMPEZW and CIMPEZW sig-
nificantly outperform the method [24] in terms of PSNR values. The
gain achieved is significant and exceeds 2 dB at all interest bitrates.
7.3. Complexity analysis
Like EZW, the proposed coders offer a variety of property which
are desirable in many image coding applications including progres-
sive transmission, embeddness, acceptable values of PSNR, the
ability to perform lossless compression and to code to exact bitrate
in the case of lossy compression. To explore the complexity the
four methods, the speed running is measured in term of computing
the time required for performing coding/decoding of each method.
192 T. Brahimi et al. / Int. J. Electron. Commun. (AEÜ) 73 (2017) 183–192
The programs was established in Matlab and tested in the same
computing environment using the following platform (MacBook,
Processor 2.8 GHz Intel Core i7, Memory 16 GB, 1600 MHz
DDR3). The results obtained are presented in Table 5. As expected,
the results achieved by the proposed coder are better than that of
original EZW for almost all test images. According to average value
in Table 5, IEZW[23] spends short time during the coding/decoding
process. This is due to the fact that this algorithm adopts a subband
classification process allowing to scan only significant subbands.
Table 6 compares the performance of colour images. On average,
the proposed coder CCIMPEZW provides a comparable perfor-
mance with that obtained by CCEZW. Moreover, CCIMPEZW shows
better performances than that of CEZW coder for all reported
bitrates. Note that the decoder is typically faster than the encoder
which always needs to visit all the pixels of a set, unlike the deco-
der which can skip several sets of zero coefficients.
8. Conclusion
In this paper, an efficient image coding method based on the
EZW coder, namely IMP1EZW, is proposed. According to an effi-
cient significant symbol map, the number of zero-trees, the scan-
ning and symbol redundancy of the original EZW have been
reduced as the coding process sped up. Moreover, we have pro-
posed a development of IMP1EZW to embedded colour image cod-
ing by taking advantage of the interdependency of colour
components. The proposed EZW image schemes have shown good
results for both greyscale and colour image coding applications at
any given bit rate for all test images.
References
[1] Usevitch BE. A tutorial on modern lossy wavelet image compression:
foundations of JPEG 2000. IEEE Signal Process Mag 2001;18(5):22–35.
[2] Gargour C, Gabrea M, Ramanchandran V, Lina JM. A short introduction to
wavelets and their applications. IEEE Circuits Syst Mag 2009;9(2):57–68.
[3] Brahimi T, Laouir F, Kechacha N. An efficient wavelet-based image coder. In:
3rd IEEE international conference on information and communication
technologies: from theory to applications (ICTTA); 2008. pp. 1–4.
[4] Brahimi T, Boubchir L, Fournier R, Nait-Ali A. An improved multimodal signal-
image compression scheme with application to natural images and biomedical
data. Multimedia Tools and Applications; 2016. https://link.springer.com/
article/10.1007/s11042-016-3952-7.
[5] Ghanbari M. Standard codecs image compression to advanced video coding.
3rd ed. London, UK: IET; 2011.
[6] Salomon D. Data compression: the complete reference. 4th ed. Springer; 2011.
[7] Pearlman WA, Said A. Digital signal compression: principle and
practice. United Kingdom: Cambridge University Press; 2011.
[8] Xie K, Bai Z, Yu W. Fast seismic data compression based on high-efficiency
SPIHT. Electron. Lett. 2014;50(5):365–7.
[9] Xu G, Han J, Zou Y, Zeng X. A 1.5-D multi-channel EEG compression algorithm
based on NLSPIHT. IEEE Signal Process Lett 2015;22:1118–22.
[10] Antonini M, Barlaud M, Mathieu P, Daubechies I. Image coding using wavelet
transform. IEEE Trans Image Process 1992;1(2):205–20.
[11] Boubchir L, Brahimi T, Fournier R, Nait-Ali A. A novel multimodal compression
scheme based on a spiral insertion function in the wavelet domain. In: 11th
international conference on information sciences, signal processing and their
applications (ISSPA); 2012. pp. 97–102.
[12] Brahimi T, Boubchir L, Fournier R, Nait-Ali A. An improved multimodal signal-
image compression scheme with application to natural images and biomedical
data. Multimedia Tools Appl 2016:1–23. http://dx.doi.org/10.1007/s11042-
016-3952-7.
[13] Brahimi T, Melit A, Khelifi F. An improved SPIHT algorithm for lossless image
coding. Digital Signal Processing 2009;19(2):220–8.
[14] Mallat S. A wavelet tour of signal processing. Third Ed. Academic press; 2008.
[15] Said A, Pearlman WA. A new fast and efficient image codec based on set
partitioning in hierarchical trees. IEEE Trans Circuits Syst Video Technol
1996;6(3):243–50.
[16] Shapiro JM. Embedded image coding using zerotrees of wavelet coefficients.
IEEE Trans Signal Process 1993;41(12):3445–62.
[17] Wu MS. Genetic algorithm based on discrete wavelet transformation for fractal
image compression. J Vis Commun Image Represent 2014;25(8):1835–41.
[18] Kumar R, Kumar A, Singh GK. Electrocardiogram signal compression based on
singular value decomposition (SVD) and adaptive scanning wavelet difference
reduction (ASWDR) technique. AEU-Int J Electron Commun 2015;69
(12):1810–22.
[19] Kassim AA, Lee WS. Embedded color image coding using SPIHT with partial
linked spatial orientation trees. IEEE Trans Circuits Syst Video Technol 2003;13
(2):203–6.
[20] Shen K, Delp EJ. Color image compression using an embedded rate scalable
approach. Proceedings of the international conference on image process
1997;3:34–7.
[21] Shen K, Delp EJ. Wavelet based rate scalable video compression. IEEE Trans
Circuits Syst Video Technol 1999;9(1):109–22.
[22] Saenz M, Salama P, Shen K, Delp EJ. An evaluation of color embedded wavelet
image compression techniques. In: Proceedings of the SPIE/IS&T Conference on
Visual Communications and Image Processing (VCIP), pp. 282–293, 1999.
[23] Kang ES, Tanaka T, Ko SJ. Improved embedded zerotree wavelet coder. Proc IEE
Electron Lett 1999;35(9):705–6.
[24] Pujol FA, Mora H, Jimeno A, Sánchez JL. Colour image compression based on
the embedded zerotree wavelet. In: International work-conference on artificial
neural networks. Berlin Heidelberg: Springer; 2009. p. 612–5.
[25] Tohumoglu G, Sezgin KE. ECG signal compression by multi-iteration EZW
coding for different wavelets and thresholds. Comput Biol Med 2007;37
(2):173–82.
[26] Yang X, Ren H, Li B. Embedded zerotree wavelets coding based on adaptive
fuzzy clustering for image compression. Image Vis Comput 2008;26(2):812–9.
[27] Pujol FA, Mora H, Sánchez JL, Jimeno A. EZW-based image compression with
omission and restoration of wavelet subbands. In: Progress in pattern
recognition, image analysis and applications volume 4756 of the series
lecture notes in computer science; 2007. pp. 134–141.
[28] Christophe E, Duhamel P, Mailhes C. Adaptation of zerotrees using signed
binary digit representations for 3 dimensional image coding. EURASIP J Image
Video Process 2007:1–7.
[29] Kassim AA, Tan EH, Lee WS. 3D color set partitioning in hierarchical trees. Circ
Syst Signal Process 2009;28:41–53.
[30] Kumar M, Vaish A. Prediction error based compression of color images using
WDR coding. AEU-Int J Electron Commun 2016;70(9):1164–71.
Tahar Brahimi received the ‘‘Ingénieur d’état” degree in electrical engineering, the
‘‘Magistère” degree ‘in electronics, and PhD degree in electronics from the
University of Jijel, Algeria, in 2000, 2003 and 2012, respectively. He is now Assistant
Professor within the Department of Electronics at Jijel University, Algeria. His
research interests include image coding, multimodal compression, image denoising,
and digital signal and image processing.
Farid Laouir received the ‘‘Ingénieur d’état” degree in electronics from the
University of Jijel, Algeria in 2007. He was with the electronic department between
2004 and 2007. His research interests include image coding, image processing and
computer networking.
Larbi Boubchir received an ‘‘Ingénieur d’état” degree (B.Eng.) in Computer Science
from the University of Sciences and Technologies of Oran, Algeria, an ‘DEA’ degree
(Master of Advanced Studies) in Computer Science from the Polytechnic of the
University of Tours (Polytech’Tours), France, and an Ph.D. degree in Signal and
Image Processing from the University of Caen, France. He is now Associate Professor
of Computer Science at the University of Paris 8, France; within the department of
Computer Science and LIASD research Lab. Before joining University of Paris 8, he
was a Researcher at the French National Centre for Scientific Research (CNRS),
France, and then he held a Research Fellow position at Northumbria University at
Newcastle, U.K., within the Computer and Electronic Security Systems group of the
department of Computer Science and Digital Technologies. His research interests
include image processing, biomedical signal processing, biomedical engineering,
brain–computer interface and biometrics. He has authored and co-authored more
than 60 publications. Dr. Boubchir is a Senior Member of IEEE and also a member of
IEEE SPS, IEEE EMBS and IARP.
Arab Ali-Chérif received M.Sc. and Ph.D. degrees in Computer Science from the
University of Paris 8, France, in 1988 and 1995, respectively. From 1998 to 2001, he
was an Assistant professor at the University of Paris 8 within the Advanced Com-
puting Laboratory of Saint-Denis (LIASD). In 2001, he received the Habilitation
degree from the same university; where he is currently a Full Professor at the
department of Computer Science. His research interests include soft computing,
data fusion, localization, routing protocols and security with applications in wire-
less sensor networks, robotics and medicine. He is a Leader of LIASD research lab-
oratory and scientific director of several research projects. In 1996, he obtained gold
medal in robotics at Los Alamos, USA. He has published more than 100 papers in
scientific journals, books and conferences. He also supervised more than 20 Ph.D.
theses and 5 Habilitation degrees.