The JPEG 2000 still image compression standard

The JPEG 2000 still image compression standard
The JPEG 2000 Still Image
Compression Standard
Athanassios Skodras, Charilaos Christopoulos, and Touradj Ebrahimi
new call for contributions was launched for the
development of a new standard for the compression of
still images, the JPEG 2000 standard [28], [29]. This
project, JTC 1.29.14 (ISO/IEC 15444-1 or ITU-T Rec.
T.800), was intended to create a new image coding system for different types of still images (bilevel, gray level,
color, multicomponent), with different characteristics
(natural images, scientific, medical, remote sensing, text,
rendered graphics, etc.) allowing different imaging models (client/server, real-time transmission, image library
archival, limited buffer and bandwidth resources, etc.)
preferably within a unified system. This coding system
should provide low bit-rate operation with rate distortion and subjective image quality performance superior
to existing standards, without sacrificing performance at
other points in the rate-distortion spectrum, and at the
same time incorporating many interesting features. One
of the aims of the standardization committee has been the development of
Part I, which could be used on a royalty- and fee-free basis. This is important for the standard to become widely
accepted, in the same manner as the
original JPEG with Huffman coding is
now. The standardization process,
which is coordinated by the
JTC1/SC29/WG1 of ISO/IEC (Fig.
1) has already (since December 2000)
produced the International Standard
(IS) for Part I [41].
In this article the structure of Part I
of the JPEG 2000 standard is presented
and performance comparisons with established standards are reported. This
article is intended to serve as a tutorial
for the JPEG 2000 standard. In the next
section the main application areas and
their requirements are given. The architecture of the standard follows afterwards, with the description of the tiling,
multicomponent transformations, wavelet transforms,
quantization and entropy coding. Some of the most significant features of the standard are presented next, such as
region-of-interest coding, scalability, visual weighting, error resilience and file format aspects. Finally, some com-
he development of standards (emerging and established) by the International Organization for
Standardization (ISO), the International Telecommunications Union (ITU), and the International Electrotechnical Commission (IEC) for audio,
image, and video, for both transmission and storage, has
led to worldwide activity in developing hardware and software systems and products applicable to a number of diverse disciplines [7], [22], [23], [55], [56], [73]. Although
the standards implicitly address the basic encoding operations, there is freedom and flexibility in the actual design
and development of devices. This is because only the syntax
and semantics of the bit stream for decoding are specified
by standards, their main objective being the compatibility
and interoperability among the systems (hardware/software) manufactured by different companies. There is, thus,
much room for innovation and ingenuity.
Since the mid 1980s, members from
both the ITU and the ISO have been
working together to establish a joint international standard for the compression
of grayscale and color still images. This
effort has been known as JPEG, the Joint
Photographic Experts Group. (The
“joint” in JPEG refers to the collaboration between ITU and ISO; see Fig. 1).
Officially, JPEG corresponds to the
ISO/IEC international standard
10928-1, digital compression and coding of continuous-tone (multilevel) still
images or to the ITU-T Recommendation T.81. The text in both these ISO
and ITU-T documents is identical. The
process was such that, after evaluating a
number of coding schemes, the JPEG
members selected a discrete cosine transform(DCT)-based method in 1988.
From 1988 to 1990, the JPEG group continued its work by
simulating, testing and documenting the algorithm. JPEG
became Draft International Standard (DIS) in 1991 and
International Standard (IS) in 1992 [55], [73].
With the continual expansion of multimedia and
Internet applications, the needs and requirements of the
technologies used grew and evolved. In March 1997, a
parative results are reported and the future parts of the
standard are discussed.
Why Another
Still Image Compression Standard?
The JPEG standard has been in use for almost a decade
now. It has proved a valuable tool during all these years,
but it cannot fulfill the advanced requirements of today.
Today’s digital imagery is extremely demanding, not only
from the quality point of view, but also from the image
size aspect. Current image size covers orders of magnitude, ranging from web logos of size of less than 100
Kbits to high quality scanned images of approximate size
of 40 Gbits [20], [33], [43], [48]. The JPEG 2000 international standard represents advances in image compres-
sion technology where the image coding system is optimized not only for efficiency, but also for scalability and
interoperability in network and mobile environments.
Digital imaging has become an integral part of the
Internet, and JPEG 2000 is a powerful new tool that provides power capabilities for designers and users of networked image applications [41].
The JPEG 2000 standard provides a set of features that
are of importance to many high-end and emerging applications by taking advantage of new technologies. It addresses areas where current standards fail to produce the
best quality or performance and provides capabilities to
markets that currently do not use compression. The markets and applications better served by the JPEG 2000
standard are Internet, color facsimile, printing, scanning
(consumer and prepress), digital photography, remote
SC 29
WG 1
WG 11
WG 12
▲ 1. Structure of the standardization bodies and related terminology. ISO: International Organization for Standardization, ITU: International Telecommunications Union, IEC: International Electrotechnical Commission, JTC: Joint Technical Committee on Information
Technology, SC: Subcommittee, SG: study group, WG: working group, AHG: ad hoc group, JPEG: Joint Photographic Experts Group,
JBIG: Joint Bi-Level Image Group, MPEG: Moving Picture Experts Group, MHEG: Multimedia Hypermedia Experts Group, AG: advisory
group, AGM: Advisory Group on Management, and RA: registration authority.
feature include medical images, where loss is not always
tolerated; image archival applications, where the highest
quality is vital for preservation but not necessary for display; network applications that supply devices with different capabilities and resources; and prepress imagery. It
is also desired that the standard should have the property
of creating embedded bit stream and allow progressive
lossy to lossless buildup.
▲ Progressive transmission by pixel accuracy and resolution:
Progressive transmission that allows images to be reconstructed with increasing pixel accuracy or spatial resolution is essential for many applications such as web
browsing, image archival and printing.
▲ Region-of-interest (ROI) coding: Often there are parts of
an image that are of greater importance than others. This
feature allows users to define certain ROIs in the image to
be coded and transmitted in a better quality and less distortion than the rest of the image.
▲ Open architecture: It is desirable to allow open architecture to optimize the system for different image types and
applications. With this feature, a decoder is only required
to implement the core tool set and the parser that understands the code stream.
▲ Robustness to bit errors: It is desirable to consider robustness to bit errors while designing the code stream. One
application, where this is important, is transmission over
wireless communication channels. Portions of the code
stream may be more important than others in determining decoded image quality. Proper design of the code
stream can aid subsequent error correction systems in alleviating catastrophic decoding failures.
▲ Protective image security: Protection of a digital image
can be achieved by means of different approaches such
as watermarking, labeling, stamping, or encryption.
JPEG 2000 image files should have provisions for such
JPEG 2000 represents advances
in image compression technology
where the image coding system
is optimized for efficiency,
scalability, and interoperability in
network and mobile
sensing, mobile, medical imagery, digital libraries/archives, and E-commerce. Each application area imposes
some requirements that the standard, up to a certain degree, should fulfill. Some of the most important features
that this standard should possess are the following [33]:
▲ Superior low bit-rate performance: This standard should
offer performance superior to the current standards at
low bit rates (e.g., below 0.25 b/p for highly detailed
gray-scale images). This significantly improved low
bit-rate performance should be achieved without sacrificing performance on the rest of the rate-distortion spectrum. Network image transmission and remote sensing
are some of the applications that need this feature.
▲ Continuous-tone and bilevel compression: It is desired to
have a coding standard that is capable of compressing
both continuous-tone and bilevel images. If feasible, this
standard should strive to achieve this with similar system
resources. The system should compress and decompress
images with various dynamic ranges (e.g., 1 to 16 bits)
for each color component. Examples of applications that
can use this feature include compound documents with
images and text, medical images with annotation overlays, and graphic and computer generated images with binary and near to binary regions, alpha and transparency
planes, and facsimile.
▲ Lossless and lossy compression: It is desired to provide
lossless compression naturally in the course of progressive decoding. Examples of applications that can use this
Image Data
The JPEG 2000 Compression Engine
The JPEG 2000 compression engine (encoder and decoder) is illustrated in block diagram form in Fig. 2. At
Entropy Encoding
Compressed Image Data
Image Data
or Transmit
Entropy Decoding
Compressed Image Data
▲ 2. General block diagram of the JPEG 2000 (a) encoder and (b) decoder.
the encoder, the discrete transform is first applied on the
source image data. The transform coefficients are then
quantized and entropy coded before forming the output
code stream (bit stream). The decoder is the reverse of the
encoder. The code stream is first entropy decoded,
dequantized, and inverse discrete transformed, thus resulting in the reconstructed image data. Although this
general block diagram looks like the one for the conventional JPEG, there are radical differences in all of the processes of each block of the diagram. A quick overview of
the whole system is as follows:
▲ The source image is decomposed into components.
▲ The image components are (optionally) decomposed
into rectangular tiles. The tile-component is the basic unit
of the original or reconstructed image.
▲ A wavelet transform is applied on each tile. The tile is
decomposed into different resolution levels.
▲ The decomposition levels are made up of subbands of
coefficients that describe the frequency characteristics of
local areas of the tile components, rather than across the
entire image component.
▲ The subbands of coefficients are quantized and collected into rectangular arrays of “code blocks.”
▲ The bit planes of the coefficients in a code block (i.e.,
the bits of equal significance across the coefficients in a
code block) are entropy coded.
▲ The encoding can be done in such a way that certain regions of interest can be coded at a higher quality than the
▲ Markers are added to the bit stream to allow for error
▲ The code stream has a main header at the beginning
that describes the original image and the various decomposition and coding styles that are used to locate, extract,
decode and reconstruct the image with the desired resolution, fidelity, region of interest or other characteristics.
For the clarity of presentation we have decomposed
the whole compression engine into three parts: the pre-
processing, the core processing, and the bit-stream formation part, although there exist high inter-relation between them. In the preprocessing part the image tiling,
the dc-level shifting and the component transformations
are included. The core processing part consists of the discrete transform, the quantization and the entropy coding
processes. Finally, the concepts of the precincts, code
blocks, layers, and packets are included in the bit-stream
formation part.
Image Tiling
The term “tiling” refers to the partition of the original
(source) image into rectangular nonoverlapping blocks
(tiles), which are compressed independently, as though
they were entirely distinct images [13], [24], [38], [39],
[48], [63]. All operations, including component mixing,
wavelet transform, quantization and entropy coding are
performed independently on the image tiles (Fig. 3). The
tile component is the basic unit of the original or reconstructed image. Tiling reduces memory requirements, and
since they are also reconstructed independently, they can
be used for decoding specific parts of the image instead of
the whole image. All tiles have exactly the same dimensions, except maybe those at the boundary of the image.
Arbitrary tile sizes are allowed, up to and including the entire image (i.e., the whole image is regarded as one tile).
Components with different subsampling factors are tiled
with respect to a high-resolution grid, which ensures spatial consistency on the resulting tile components. As expected, tiling affects the image quality both subjectively
(Fig. 4) and objectively (Table 1). Smaller tiles create more
tiling artifacts compared to larger tiles (PSNR values are
the average over all components). In other words, larger
tiles perform visually better than smaller tiles. Image degradation is more severe in the case of low bit rate than the
case of high bit rate. It is seen, for example, that at 0.125
b/p there is a quality difference of more than 4.5 dB be-
DWT on Each Tile
DC Level
▲ 3. Tiling, dc-level shifting, color transformation (optional) and DWT of each image component.
ward component transform (Figs.
3 and 5). At the decoder side, inverse dc level shifting is performed
on reconstructed samples by adding to them the bias 2P-1 after the
computation of the inverse component transform.
Component Transformations
JPEG 2000 supports multiple-component images. Different
components need not have the
same bit depths nor need to all be
signed or unsigned [38], [39]. For
reversible (i.e., lossless) systems,
the only requirement is that the bit
depth of each output image component must be identical to the bit
depth of the corresponding input
image component.
Component transformations
improve compression and allow
for visually relevant quantization.
The standard supports two different component transformations,
▲ 4. Image :”ski” of size 720 × 576 (courtesy of Phillips Research, UK): (a) original image,
one irreversible component trans(b)-(d) reconstructed images after JPEG 2000 compression at 0.25 bpp: (b) without tiling,
formation (ICT) that can be used
(c) with 128 × 128 tiling, and (d) with 64 × 64 tiling.
for lossy coding and one reversible
tween no-tiling and tiling at 64 × 64, while at 0.5 b/p this
component transformation (RCT) that may be used for
difference is reduced to approximately 1.5 dB.
lossless or lossy coding, and all this in addition to encoding without color transformation. The block diagram of
the JPEG 2000 multicomponent encoder is depicted in
DC Level Shifting
Fig. 5. (Without restricting the generality, only three
Prior to computation of the forward discrete wavelet
components are shown in the figure. These components
transform (DWT) on each image tile, all samples of the
could correspond to the RGB of a color image.)
image tile component are dc level shifted by subtracting
the same quantity 2P-1, where P is the component’s preciSince the ICT may only be used for lossy coding, it
sion. DC level shifting is performed on samples of commay only be used with the 9/7 irreversible wavelet transponents that are unsigned only. Level shifting does not
form. (See also next section.) The forward and the inverse
affect variances. It actually converts an unsigned repreICT transformations are achieved by means of (1a) and
sentation to a two’s complement representation, or vice
(1b), respectively [7], [38], [56]
versa [55], [56]. If color transformation is used, dc level
  R
 Y   0299
shifting is performed prior to the computation of the for  
  
 ⋅G
 C b  =  −016875
Table 1. The Effect of Tiling on Image Quality.
  
  B
−0.41869 −008131
 C r   05
  
Bit Rate
No Tiling
Tiles of
128 × 128
Tiles of
64 × 64
PSNR (in dB) for the color image “ski” (of size 720 × 576
pixels per component)
1.402   Y 
 R   10
  
  
. −034413
 ⋅ Cb .
 G  =  10
  
  
  
 Cr 
Since the RCT may be used for lossless or lossy coding,
it may only be used with the 5/3 reversible wavelet transform. (See also next section). The RCT is a decorrelating
transformation, which is applied to the three first components of an image. Three goals are achieved by this transformation, namely, color decorrelation for efficient
Component Image
Component 1
DC Level
Component 2
DC Level
Component 3
DC Level
JPEG 2000
JPEG 2000
Compressed Image Data
JPEG 2000
▲ 5. The JPEG 2000 multiple component encoder. Color transformation is optional. If employed, it can be irreversible or reversible.
compression, reasonable color space with respect to the
human visual system for quantization, and ability of having lossless compression, i.e., exact reconstruction with
finite integer precision. For the RGB components, the
RCT can be seen as an approximation of a YUV transformation. All three of the components shall have the same
sampling parameters and the same bit depth. There shall
be at least three components if this transform is used. The
forward and inverse RCT is performed by means of (2a)
and (2b), respectively, where the subscript r stands for reversible
 Yr
 Vr
U r
  R + 2G + B  
 
 
 
R −G
 =
 
 
 U r + Vr
 G   Yr − 
  
  
  
Ur +G
 B 
 
where  a  is the largest integer not exceeding a.
A subjective quality evaluation of the different color
spaces can be found in [37] and [50]. Performance comparisons between lossless compression (i.e., using RCT
and the 5/3 filter) and decompression at a certain bitrate,
and lossy compression (i.e., using ICT and the 9/7 filter)
and decompression at the same bit rate, has shown that
the later produces substantially better results, as shown in
Table 2.
An effective way to reduce the amount of data in JPEG
is to use an RGB to YCrCb decorrelation transform followed by subsampling of the chrominance (C r ,C b ) components. This is not recommended for use in JPEG 2000,
since the multiresolution nature of the wavelet transform
may be used to achieve the same effect. For example, if the
HL, LH, and HH subbands of a component’s wavelet decomposition are discarded and all other subbands retained, a 2:1 subsampling is achieved in the horizontal
and vertical dimensions of the component.
Core Processing
Wavelet Transform
Wavelet transform is used for the analysis of the tile components into different decomposition levels [9], [26],
[70], [71]. These decomposition levels contain a number
of subbands, which consist of coefficients that describe
the horizontal and vertical spatial frequency characteristics of the original tile component. In Part I of the JPEG
2000 standard only power of 2 decompositions are allowed in the form of dyadic decomposition as shown in
Fig. 6 for the image "Lena".
To perform the forward DWT the standard uses a
one-dimensional (1-D) subband decomposition of a 1-D
set of samples into low-pass and high-pass samples.
Low-pass samples represent a down-sampled, low-resolution version of the original set. High-pass samples repTable 2. The effect of component transformation on
the compression efficiency for the ski image. RCT is
employed in the lossless case and ICT in the lossy
case. No tiling is used.
Without Color
With Color
16.88 b/p
14.78 b/p
compression at
0.25 b/p
25.67 dB
26.49 dB
▲ 6. Three-level dyadic wavelet decomposition of the image “Lena.”
Table 3. Daubechies 9/7 Analysis
and Synthesis Filter Coefficients.
Analysis Filter Coefficients
Low-Pass Filter hL(i)
High-Pass Filter hH(i)
Synthesis Filter Coefficients
Low-Pass Filter gL(i)
High-Pass Filter gH(i)
Table 4. Le Gall 5/3 Analysis and
Synthesis Filter Coefficients.
Analysis Filter
Synthesis Filter
Filter hL(i)
Filter hH(i)
Filter gL(i)
Filter gH(i)
resent a down-sampled residual version of the original
set, needed for the perfect reconstruction of the original
set from the low-pass set. The DWT can be irreversible or
reversible. The default irreversible transform is implemented by means of the Daubechies 9-tap/7-tap filter [4].
The analysis and the corresponding synthesis filter coefficients are given in Table 3. The default reversible transformation is implemented by means of the Le Gall
5-tap/3-tap filter, the coefficients of which are given in
Table 4 [2], [12], [46].
The standard can support two filtering modes: convolution based and lifting based. For both modes to be implemented, the signal should first be extended
periodically as has been demonstrated in the previous article by Usevitch [70] (Fig. 7). This periodic symmetric
extension is used to ensure that for the filtering operations that take place at both boundaries of the signal, one
signal sample exists and spatially corresponds to each coefficient of the filter mask. The number of additional samples required at the boundaries of the signal is therefore
filter-length dependent. The symmetric extension of the
boundary is of type (1,1), i.e., the first and the last samples appear only once and are whole sample (WS) since
the length of the kernel is odd [3], [10], [11].
Convolution-based filtering consists in performing a
series of dot products between the two filter masks and
the extended 1-D signal. Lifting-based filtering consists
of a sequence of very simple filtering operations for which
alternately odd sample values of the signal are updated
with a weighted sum of even sample values, and even
sample values are updated with a weighted sum of odd
sample values (Fig. 8) [3], [45], [62], [64], [65]. For the
reversible (lossless) case the results are rounded to integer
values. The lifting-based filtering [2], [38] for the 5/3
analysis filter is achieved by means of
 x (2 n) + x ext (2 n + 2) 
y(2 n + 1) = x ext (2 n + 1) −  ext
 y(2 n − 1) + y(2 n + 1) + 2 
y(2 n) = x ext (2 n) + 
where x ext is the extended input signal and y is the output
signal. The 5/3 filter allows repetitive encoding and de-
coding of an image without any loss. Of course,
this is true when the decompressed image values
are not clipped when they fall outside the full dynamic range (i.e., 0-255 for an 8 b/p image) [48].
Traditional wavelet transform implementations require the whole image to be buffered and
the filtering operation to be performed in vertical
and horizontal directions. While filtering in the
horizontal direction is very simple, filtering in the
vertical direction is more cumbersome. Filtering
along a row requires one row to be read; filtering
▲ 7. Periodic symmetric extension of the finite length signal “ABCDEFG.”
along a column requires the whole
image to be read. The line-based
wavelet transform overcomes this
difficulty, providing exactly the same
transform coefficients as the traditional wavelet transform implemenXext(n)
tation [17], [18], [38]. However, the
line-based wavelet transform alone
does not p r ovi de a c omplete
line-based encoding paradigm for
JPEG 2000. A complete row-based
y(2n+1) Highpass
coder has to take also into account all
the subsequent coding stages up to
the entropy coding. Such an algo- ▲ 8. The forward (analysis) wavelet transform using lifting. P and U stand for prediction
and update, respectively.
rithm is described in [17] and [18].
After transformation, all coefficients are quantized. Uniform scalar quantization with dead-zone about the origin is used in Part I and trellis coded quantization
(TCQ) in Part II of the standard [38], [39], [44].
Quantization is the process by which the coefficients are
reduced in precision. This operation is lossy, unless the
quantization step is 1 and the coefficients are integers, as
produced by the reversible integer 5/3 wavelet. Each of
the transform coefficients a b (u , v)) of the subband b is
quantized to the value q b (u , v) according to the formula
[38], [39]
 a (u , v) 
q b (u , v) = sign(a b (u , v)) b
 ∆ b 
The quantization step-size ∆ b is represented relative to
the dynamic range of subband b. In other words, the
JPEG 2000 standard supports separate quantization
step-sizes for each subband [70]. However, one
quantization step-size is allowed per subband. The dynamic range depends on the number of bits used to represent the original image tile component and on the choice
of the wavelet transform. All quantized transform coefficients are signed values even when the original components are unsigned. These coefficients are expressed in a
sign-magnitude representation prior to coding. For reversible compression, the quantization step-size is required to be one.
Entropy Coding
Entropy coding is achieved by means of an arithmetic coding system that compresses binary symbols relative to an
adaptive probability model associated with each of 18 different coding contexts. The MQ coding algorithm is used
to perform this task and to manage the adaptation of the
conditional probability models [39], [55]. This algorithm
has been selected in part for compatibility reasons with the
arithmetic coding engine used by the JBIG2 compression
standard and every effort has been made to ensure commonality between implementations and surrounding intellectual property issues for JBIG2 and JPEG 2000 [39].
The recursive probability interval subdivision of Elias coding is the basis for the binary arithmetic coding process.
With each binary decision, the current probability interval
is subdivided into two subintervals, and the code stream is
modified (if necessary) so that it points to the base (the
lower bound) of the probability subinterval assigned to the
symbol, which occurred. Since the coding process involves
addition of binary fractions rather than concatenation of
integer code words, the more probable binary decisions
can often be coded at a cost of much less than one bit per
decision [7], [55], [56].
As mentioned above, JPEG 2000 uses a very restricted number of contexts for any given type of bit.
This allows rapid probability adaptation and decreases
the cost of independently coded segments. The context
models are always reinitialized at the beginning of each
code block and the arithmetic coder is always terminated
rectangles, called code blocks, which form the input to the
entropy coder (Fig. 9). The size of the code block is typically 64 × 64 and no less than 32 × 32.
Within each subband the code blocks are visited in raster order. These code blocks are then coded a bit plane at a
time starting with the most significant bit plane with a
nonzero element to the least significant bit plane
[67]-[69]. Each code block is coded entirely independently, without reference to other blocks in the same or
other subbands, something that is in contrary to the approach adopted by the zero-tree coder described by
Usevitch [70]. This independent embedded block coding
offers significant benefits, such as spatial random access
to the image content, efficient geometric manipulations,
error resilience, parallel computations during coding or
decoding, etc. The individual bit planes of the coefficients
in a code block are coded within three coding passes.
Bit-Stream Formation
Each of these coding passes collects contextual informaPrecincts and code blocks
tion about the bit plane data. The arithmetic coder uses
After quantization, each subband is divided into rectanguthis contextual information and its internal state to generlar blocks, i.e., nonoverlapping rectangles. Three spatially
ate a compressed bit stream. Different termination mechconsistent rectangles (one from each subband at each resoanisms allow different levels of independent extraction of
lution level) comprise a packet partition location or prethis coding pass data.
cinct. Each precinct is further divided into nonoverlapping
Each bit plane of a code block is scanned in a particular
order (Fig. 10). Starting from the top left, the first
four bits of the first column are scanned. Then the
first four bits of the second column, until the
width of the code block is covered. Then the second four bits of the first column are scanned and
0 1
so on. A similar vertical scan is continued for any
2 3
leftover rows on the lowest code blocks in the
subband [38]. This stripe height of 4 has been
Code Block
carefully selected to facilitate efficient hardware
Tile Component
and software implementations [68], [69].
Each coefficient bit in the bit plane is coded in
8 9
4 5
6 7
10 11
only one of the three coding passes, namely the
significance propagation, the magnitude refinement, and the cleanup pass. For each pass, contexts are created which are provided to the
arithmetic coder [48], [69].
During the significance propagation pass, a
▲ 9. Partition of a tile component into code blocks and precincts.
bit is coded if its location is not significant, but at
least one of its eight-connect neighbors is significant. Nine context bins are created based on how many
and which ones are significant. If a coefficient is signifi0
cant then it is given a value of 1 for the creation of the con2
text, otherwise it is given a value of 0. The mapping of the
contexts also depends on which subband (at a given decomposition level) the code block is in. The significance
propagation pass includes only bits of coefficients that
were insignificant (the significance bit has yet to be encountered) and have a nonzero context. All other coefficients are skipped. The context is delivered to the
arithmetic decoder (along with the bit stream) and the
decoded coefficient bit is returned.
The second pass is the magnitude refinement pass.
▲ 10. Scan pattern of each bit plane of each code block.
During this pass, all bits that became significant in a preat the end of each block (i.e., once, at the end of the last
subbit plane). This is useful for error resilience also. (A
code block is the fundamental entity for entropy coding—see also next section.)
In addition to the above, a lazy coding mode is used to
reduce the number of symbols that are arithmetically
coded [39], [48]. According to this mode, after the
fourth bit plane is coded, the first and second pass are included as raw (uncompressed data), i.e., the MQ coder is
bypassed, while only the third coding pass of each bit
plane employs arithmetic coding. This results in significant speedup for software implementations at high bit
rates. Lazy coding has a negligible effect on compression
efficiency for most natural images, but not for compound
imagery [69].
vious biplane are coded. The magnitude refinement pass
includes the bits from coefficients that are already significant (except those that have just become significant in the
immediately preceding significance propagation pass).
The context used is determined by the summation of the
significance state of the horizontal, vertical, and diagonal
neighbors. These are the states as currently known to the
decoder, not the states used before the significance decoding pass. Further, it is dependent on whether this is
the first refinement bit (the bit immediately after the significance and sign bits) or not.
The final pass is the clean-up pass in which all bits not
encoded during the previous passes are encoded (i.e., coefficients that are insignificant and had the context value
of zero during the significance propagation pass). The
cleanup pass not only uses the neighbor context, like that
of the significance propagation pass, but also a run-length
context. Run coding occurs when all four locations in the
column of the scan are insignificant and each has only insignificant neighbors [38], [39], [48], [69].
Packets and Layers
For each code block, a separate bit stream is generated. No
information from other blocks is utilized during the genera-
Comparative results have shown
that JPEG 2000 is indeed
superior to established still
image compression standards.
tion of the bit stream for a particular block. Rate distortion
optimization is used to allocate truncation points to each
code block. The bit stream has the property that it can be
truncated to a variety of discrete lengths, and the distortion
incurred, when reconstructing from each of these truncated
subsets, is estimated and denoted by the mean squared error. During the encoding process, the lengths and the distortions are computed and temporarily stored with the
compressed bit stream itself. The compressed bit streams
from each code block in a precinct comprise the body of a
packet. A collection of packets, one from each precinct of
each resolution level, comprises the layer (Fig. 11). A packet
could be interpreted as one quality increment for one resolution level at one spatial location, since precincts correspond
roughly to spatial locations. Similarly, a layer could be interpreted as one quality increment for the entire full resolution
Image Component
Code Stream
Coded Code Block
Code Block
Note: H Stands for Header
▲ 11. Conceptual correspondence between the spatial and the bit stream representations.
image [48]. Each layer successively and monotonically improves the image quality, so that the decoder is able to decode the code block contributions contained in each layer in
sequence. The final bit stream is organized as a succession of
layers. Each component is coded independently, and the
coded data are interleaved on a layer basis. There are four
types of progression in the JPEG 2000 bit stream, namely
resolution, quality, spatial location and component. Different types of progression are achieved by the appropriate ordering of the packets within the bit stream (assuming that
the image consists of a single tile).
Once the entire image has been compressed, a
post-processing operation passes over all the compressed
code blocks. This operation determines the extent to
which each code block’s embedded bit stream should be
truncated to achieve a particular target bitrate or distortion. The first, lowest layer (of lowest quality), is formed
from the optimally truncated code block bit streams in the
manner described above. Each subsequent layer is formed
by optimally truncating the code block bit streams to
achieve successively higher target bit rates [66]-[68].
image, the spatial and SNR (quality) scalability, the error
resilience, and the possibility of intellectual property
rights protection. All these features are incorporated
within a unified algorithm. An overview of these features
The functionality of ROI is important in applications
where certain parts of the image are of higher importance
than others [5]. In such a case, these regions need to be
encoded at higher quality than the background. During
the transmission of the image, these regions need to be
transmitted first or at a higher priority, as for example in
the case of progressive transmission.
The ROI coding scheme in Part I of the standard is
based on the so-called MAXSHIFT method of Christopoulos et al. [14]-[16]. The MAXSHIFT method is an
extension of the general ROI scaling-based coding
method of [6]. The principle of the general ROI scaling-based method is to scale (shift) coefficients so that
the bits associated with the ROI are placed in higher bit
planes than the bits associated with the background as
depicted in Fig. 12. Then, during the embedded coding
Remarkable Features of the Standard
process, the most significant ROI bit planes are placed in
The JPEG 2000 standard exhibits many nice features, the
the bit stream before any background bit planes of the
most significant being the possibility to define ROI in an
image. Depending on the scaling value, some bits of the
ROI coefficients might be
encoded together with
Scaling Based
nonROI coefficients. Thus,
the ROI will be decoded, or
refined, before the rest of
the image. Regardless of
the scaling, a full decoding
of the bit stream results in a
reconstruction of the whole
image with the highest fiBG
delity available. If the bit
stream is truncated, or the
encoding process is terminated before the whole im▲ 12. Scaling of the ROI coefficients.
age is fully encoded, the
ROI will be of higher fidelity than the rest of the image.
In JPEG 2000, the general scaling-based method is
implemented as follows:
▲ 1) The wavelet transform is calculated.
▲ 2) If an ROI has been defined, then an ROI mask is derived, indicating the set of coefficients that are required
for up to lossless ROI reconstruction (Fig. 13).
▲ 3) The wavelet coefficients are quantized. Quantized
coefficients are stored in a sign magnitude representation. Magnitude bits comprise the most significant part
of the implementation precision used (one of the reasons for this is to allow for downscaling of the background coefficients).
▲ 4) The coefficients that lay out of the ROI are
▲ 13. Wavelet domain ROI mask generation.
downscaled by a specified scaling value.
▲ 5) The resulting coefficients are progressively entropy
encoded (with the most significant bit planes first).
The decoder reverses these steps to reconstruct the image (Step 2 is still performed before Step 3). As overhead
information, the scaling value assigned to the ROI and
the coordinates of the ROI are added to the bit stream.
The decoder performs also the ROI mask generation but
scales up the background coefficients in order to recreate
the original coefficients.
According to the MAXSHIFT method, which is used
in Part I of the JPEG 2000 standard, the scaling value is
computed in such a way that it makes possible to have arbitrary shaped ROIs without the need for transmitting
shape information to the decoder. This means also that
the decoder does not have to perform ROI mask generation either (this might still be needed at the encoder). The
encoder scans the quantized coefficients and chooses a
scaling value S such that the minimum coefficient belonging to the ROI is larger than the maximum coefficient of
the background (non-ROI area). The decoder receives
the bit stream and starts the decoding process. Every coefficient that is smaller than S belongs to the background
and is therefore scaled up. The decoder needs only to upscale the received background coefficients.
The advantages of the MAXSHIFT method, as compared to the general scaling-based method, is that encoding of arbitrary shaped ROIs is now possible without the
need for shape information at the decoder (i.e., no shape
Target − Approx. 25% Circular ROI − Relative Sizes
decoder is required) and without the need for calculating
the ROI mask. The encoder is also simpler, since no shape
encoding is required. The decoder is almost as simple as a
nonROI capable decoder, while it can still handle ROIs
of arbitrary shape.
In the MAXSHIFT method, since the bit planes with
information belonging to the ROI are completely separated from those belonging to the background, the number of bit planes for the ROI and for the background can
be chosen independently. This gives the possibility to
choose different bit rates for the ROI and for the background. To do this, it is sufficient to discard the least significant bit planes of the ROI and background. With the
general scaling-based mode, it is not possible to control
these numbers independently.
Experiments have shown that for the lossless coding of
images with ROIs, the MAXSHIFT method increases
the bit rate by approximately 1% in comparison with the
lossless coding of the image without ROI [14]-[16]. This
figure is even smaller compared to the general scalingbased method, depending on the scaling value used. This
is true for large images (larger than 2K × 2K) and for ROI
sizes of about 25% of the image. Such an overhead is indeed small, given the fact that the general scaling-based
method for arbitrary shaped ROI would require shape information to be transmitted to the decoder, thus increasing the bit rate (in addition to the need of shape
encoder/decoder and ROI mask generation at the deWoman − 25% ROI − Relative Size
Max Shift
Aerial2 − Approx. 25% circular ROI − Relative Sizes
Gold − 25% Rect ROI − Relative Sizes
Max Shift
Max Shift
Max Shift
▲ 14. Lossless coding of images for different scaling factors and different ROI shapes as compared to the non-ROI case. (Image sizes in
pixels: aerial 2048 × 2048, woman 2048 × 2560, target 512 × 512, gold 720 × 576).
in a progressive by layer manner, not all of the layers of
the wavelet coefficients belonging to the background
need to be encoded. This corresponds to using different
quantization steps for the ROI and the background. Fig.
15 shows an example of ROI coding with the
MAXSHIFT method. The ROI is used in all subbands,
that is why at the early stages of the transmission, not
enough information is used for the background. For
comparison purposes, the same result is shown in Fig.
16 for the general scaling-based method, with the scaling value set to six. Similar results can be obtained with
the MAXSHIFT method if the few low-resolution
subbands are considered as full ROIs. The results show
that the MAXSHIFT method can give similar results to
the general scaling method, without the need of shape
information and mask generation at the decoder.
coder side). The performance of the MAXSHIFT
method and of the general scaling-based method for different scaling factors, as compared to the lossless coding
of an image without ROI, are depicted in Fig. 14. The
ROI shape is circular for the aerial and target images and
rectangular for the woman and the gold images. It is seen
that the MAXSHIFT method results in a very small increase in the bit rate, compared to the general scaling-based method. In fact, for arbitrary shaped regions,
where shape information needs to be included in the bit
stream, the general scaling-based method and the
MAXSHIFT method achieve similar bit rates.
The MAXSHIFT method allows the implementers of
an encoder to exploit a number of functionalities that are
supported by a compliant decoder. For example, it is
possible to use the MAXSHIFT method to encode an
image with different quality for the ROI and the background. The image is quantized so that the ROI gets the
desired quality (lossy or lossless) and then the
MAXSHIFT method is applied. If the image is encoded
More about ROI coding
ROI coding is a process performed at the encoder. The
encoder decides which is the ROI to be coded in better
▲ 15. ROI encoding results by means of the MAXSHIFT method. Part of the decompressed image “woman” at: (a) 0.5 b/p, (b) 1 b/p,
and (c) 2 b/p.
▲ 16. ROI encoding results by means of the general scaling method. Part of the decompressed image “woman” at: (a) 0.5 b/p, (b) 1
b/p, and (c) 2 b/p.
quality than the background. If the ROIs, though, are not
known to the encoder in advance, there is still possibility
for the decoder to receive only the data that is requested.
(A method for interactive ROI selection is described in
[60].) Although the simplest method is tiling, this requires that the image be encoded in tiling mode. Another
way is to extract packet partitions from the bit stream.
This can be done easily, since the length information is
stored in the header. Due to the filter impulse response
lengths, care has to be taken to extract all data required to
decode the ROI. Fine grain access can be achieved by
parsing individual code blocks. As in the case of packet
partition precincts, it is necessary to determine which
code blocks affect which pixel locations (since a single
pixel can affect four different code blocks within each
subband and each resolution and each component). We
can determine the correct packet affecting these code
blocks from the progression order information. The location of the compressed data for the code blocks can be determined by decoding the packet headers. This might be
easier than operating the arithmetic coder and context
modeling to decode the data [48].
The procedure of coefficient scaling might, in some
cases, cause overflow problems due to the finite implementation precision. In JPEG 2000 this problem is minimized since the background coefficients are scaled down,
rather than scaling up the ROI coefficients. Thus, if the
implementation precision is exceeded only the least significant bit planes of the background are lost (the decoder
or the encoder will ignore this part). The advantage is that
the ROI, which is considered to be the most important
part of the image, is still optimally treated, while the qual-
▲ 17. Example of SNR scalability. Part of the decompressed image “bike” at (a) 0.125 b/p, (b) 0.25 b/p, and (c) 0.5 b/p.
▲ 18. Example of the progressive-by-resolution decoding for the color image “bike.”
ity of the background is allowed to be degraded, as it is
considered visually less important.
The ROI general scaling-based method can be applied
to any embedded coding scheme, as for example, the embedded DCT based coders [52], the various wavelet filters [31] and the zero-tree coders [57], [58], [61].
Scalable coding of still images means the ability to achieve
coding of more than one qualities and/or resolutions simultaneously. Scalable image coding involves generating a coded
representation (bit stream) in a manner which facilitates the
derivation of images of more than one quality and/or resolution by scalable decoding. Bit-stream scalability is the property of a bit stream that allows decoding of appropriate
subsets of the bit stream to generate complete pictures of
quality and/or resolution commensurate with the proportion
of the bit stream decoded. Decoders of different complexities
(from low performance to high performance) can coexist for
a scalable bit stream. While low performance decoders may
decode only small portions of the bit stream producing basic
quality, high performance decoders may decode much more
and produce significantly higher quality. The most important
types of scalability are SNR scalability and spatial or resolution scalability. The JPEG 2000 compression system supports scalability. A key advantage of scalable compression
is that the target bit rate or reconstruction resolution need
not be known at the time of compression. A related advantage of practical significance is that the image need not be
compressed multiple times to achieve a target bit rate, as is
common with the JPEG compression standard. An additional advantage of scalability is its ability to provide resilience to transmission errors, as the most important data of
the lower layer can be sent over the channel with better error performance, while the less critical enhancement layer
data can be sent over the channel with poor error performance. Both types of scalability are very important for
Internet and database access applications and bandwidth
scaling for robust delivery. The SNR and spatial scalability
types include the progressive and hierarchical coding
modes defined in the JPEG, but they are more general.
SNR Scalability
SNR scalability is intended for use in systems with the primary common feature that a minimum of two layers of image quality is necessary. SNR scalability involves generating
at least two image layers of the same spatial resolution, but
different qualities, from a single image source. The lower
layer is coded by itself to provide the basic image quality and
the enhancement layers are coded to enhance the lower
layer. An enhancement layer, when added back to the lower
layer, regenerates a higher quality reproduction of the input
image. Fig. 17 illustrates an example of SNR scalability. The
image is first losslessly compressed and decompressed at
0.125 b/p, 0.25 b/p, and 0.5 b/p.
Spatial Scalability
Spatial scalability is intended for use in systems with the
primary common feature that a minimum of two layers of
spatial resolution is necessary. Spatial scalability involves
generating at least two spatial resolution layers from a single source such that the lower layer is coded by itself to
provide the basic spatial resolution and the enhancement
layer employs the spatially interpolated lower layer and
carries the full spatial resolution of the input image
source. Fig. 18 shows an example of three levels of progressive-by-resolution decoding for the test image “bike.”
Spatial scalability is useful for fast database access as well
as for delivering different resolutions to terminals with
different capabilities in terms of display and bandwidth.
JPEG 2000 supports also a combination of spatial and
SNR scalability. It is possible therefore to progress by
spatial scalability at a given (resolution) level and then
change the progression by SNR at a higher level. This order in progression allows a thumbnail to be displayed
first, then a screen resolution image and then an image
suitable for the resolution of the printer. It is evident that
SNR scalability at each resolution allows the best possible
image to be displayed at each resolution.
Notice that the bit stream contains markers that identify
the progression type of the bit stream. The data stored in
packets are identical regardless of the type of scalability
used. Therefore it is trivial to change the progression type
or to extract any required data from the bit stream. To
change the progression from SNR to progressive by resolution, a parser can read the markers, change the type of
progression in the markers, and then write the new markers in the new order. In this manner, fast transcoding of the
bit stream can be achieved by a server or gateway, without
requiring the use of image decoding and re-encoding, not
even the employment of the MQ coder. The required complexity corresponds to that of a copy operation.
In a similar fashion, applications that require the use of
a gray scale version of a color compressed image, as for example printing a color image to a gray-scale printer, do
not need to receive all color components. A parser can
read the markers from the color components and write
the markers for one of the components (discarding the
packets that contain the color information) [24], [48].
Error Resilience
Error resilience is one of the most desirable properties in
mobile and Internet applications [25]. JPEG 2000 uses a
variable-length coder (arithmetic coder) to compress the
quantized wavelet coefficients. Variable-length coding is
known to be prone to channel or transmission errors. A bit
error results in loss of synchronization at the entropy decoder and the reconstructed image can be severely damaged.
To improve the performance of transmitting compressed
images over error prone channels, error resilient b/p and
tools are included in the standard. The error resilience tools
deal with channel errors using the following approaches:
data partitioning and resynchronization, error detection
and concealment, and quality of service (QoS) transmission
based on priority [38], [47], [49]. Error resilience is
achieved at the entropy coding level and at the packet level.
Table 5 summarizes the various ways this is achieved [38].
Entropy coding of the quantized coefficients is performed within code blocks. Since encoding and decoding
of the code blocks are independent processes, bit errors in
the bit stream of a code block will be restricted within that
code block. To increase error resilience, termination of
the arithmetic coder is allowed after every coding pass
and the contexts may be reset after each coding pass. This
allows the arithmetic decoder to continue the decoding
process even if an error has occurred.
The lazy coding mode is also useful for error resilience.
This relates to the optional arithmetic coding bypass in
which bits are fed as raw bits into the bit stream without
arithmetic coding. This prevents the error propagation
types to which variable length coding is susceptible.
At the packet level, a packet with a resynchronization
marker allows spatial partitioning and resynchronization.
This is placed in front of every packet in a tile with a sequence
number starting at zero and incremented with each packet.
Visual Frequency Weighting
The human visual system plays an important role in the
perceived image quality of compressed images [54], [74].
System designers and users should be able to take advantage of the current knowledge in visual perception, i.e., to
utilize models of the visual system’s varying sensitivity to
spatial frequencies, as measured in the contrast sensitivity
function (CSF). Since the CSF weight is determined by
the visual frequency of the transform coefficient, there
will be one CSF weight per subband in the wavelet transform [76], [77]. The design of the CSF weights is an encoder issue and depends on the specific viewing condition
under which the decoded image is to be viewed.
Two types of visual frequency weighting are supported by the JPEG 2000 standard. The fixed visual
weighting (FVW) and the visual progressive coding or
visual progressive weighting (VPW). In the FVW, only
one set of CSF weights is chosen and applied in accordance with the viewing conditions. In the VPW, different
sets of CSF weights are used at the various stages of the
embedded coding. This is because during a progressive
transmission stage, the image is viewed at various distances. For example, at low bit rates, the image is viewed
from a relatively large distance, while as more and more
bits are received and the quality is improved, the viewing
distance is decreased (the user is more interested in details
and the viewing distance is decreased or the image is magnified, which is equivalent to reducing the viewing distance). FVW can be considered as a special case of VPW.
New File Format with IPR Capabilities
The JPEG 2000 (JP2) file format commences with a
unique signature and contains the size of the image, the
bit depth of the components in the file in cases where the
bit depth is not constant across all components, the color
space of the image, the palette which maps a single component in index space to a multiple-component image,
the type and ordering of the components within the code
stream, the resolution of the image, the resolution at
which the image was captured, the default resolution at
which the image should be displayed, the code stream, intellectual property information about the image, a tool by
which vendors can add XML formatted information to a
JP2 file, and other information [8], [38], [39].
The JP2 file format is optional in the standard. The JP2
format provides a foundation for storing application specific data (metadata) in association with a JPEG 2000
code stream, such as information required to display the
image. This format has got provisions for both image and
metadata and specifies mechanisms to indicate image
properties, such as the tone-scale or color-space of the image, to recognize the existence of intellectual property
rights (IPR) information in the file and to include
metadata (as for example vendor specific information).
Metadata give the opportunity to the reader to extract information about the image, without having to decode it,
thus allowing fast text based search in a database.
In addition to specifying the color space, the standard
allows for the decoding of single component images,
where the value of that single component represents an
index into a palette of colors. An input of a decompressed
sample to the palette converts the single value to a multiple-component tuple. The value of that tuple represents
the color of the sample.
Performance Comparisons
To judge the efficiency of the JPEG 2000 as compared to
other standards, extensive comparisons have been conducted
with regard to lossy performance, lossless performance, resilience to errors, and other characteristics. The algorithms have
been evaluated on several images, mainly from the JPEG
2000 test set, which cover a wide range of imagery types.
The Software
The software used for the experiments is reported in Table 6. The Verification Model 8.6 was used for the JPEG
Table 5. Tools for Error Resilience.
Type of Tool
coding level
—code blocks
—termination of the arithmetic coder for
each pass
—reset of contexts after each coding pass
—selective arithmetic coding bypass
—segmentation symbols
Packet level
—short packet format
—packet with resynchronization marker
Table 6. Software Implementations.
Name of the
JPEG 2000
JPEG 2000
JPEG 2000
Verification Model 8.6
Lossless JPEG
Available at (*)
Image Power
Univ. of British Columbia
Software Developers
Cannon Research
Independent JPEG Group
Univ. of British Columbia
Cornell University
(*) Available to members only
2000 compression and decompression [39]. The other
implementation software was the Independent JPEG
Group (version 6b), the SPMG JPEG-LS (version 2.2),
the lossless JPEG (version 1.0) [30], the libpng of PNG
(version 1.03) [72], and the MPEG-4 MoMuSys Verification Model of August 1999 [42].
Lossy Compression Results
The objective (PSNR) comparison results between the
different standards are graphically shown in Fig. 19. It is
seen that the lossy (nonreversible) JPEG 2000 (J2KNR)
outperforms all other standards. Compared to the JPEG
standard, it is realized that JPEG 2000 performs better by
approximately 2 dB for all compression ratios. As expected, the lossless (reversible) JPEG 2000 (J2KR) does
▲ 19. PSNR results for the lossy compression of a natural image
by means of different compression standards.
not perform as well, due to the use of reversible wavelet
filters, i.e., the 5-tap/3-tap filters.
The superiority of the JPEG 2000 over the existing
JPEG can be subjectively judged with the help of Fig. 20
and Fig. 21, where the reconstructed grayscale image
“watch” and the color image “ski” are shown after compression at 0.2 b/p and 0.25 b/p, respectively. Results of
the image cmpnd1 compressed by means of JPEG baseline and JPEG 2000 at 0.5 b/p, are shown in Fig. 22. It
becomes evident that JPEG 2000 outperforms JPEG
baseline. This performance superiority of the JPEG 2000
decreases as the bit rate increases. In fact, from the compression point of view, JPEG 2000 gives about 10 to
20% better compression factors compared to JPEG baseline, for a bit rate of about 1 b/p. Visual comparisons of
JPEG compressed images (baseline JPEG with optimized
Huffman tables) and JPEG 2000 compressed images
showed that for a large category of images, JPEG 2000
file sizes were on average 11% smaller than JPEG at 1.0
b/p, 18% smaller at 0.75 b/p, 36% smaller at 0.5 b/p and
53% smaller at 0.25 b/p [37]. In general, we can say that
for high quality imaging applications (i.e., 0.5-1.0 b/p)
JPEG 2000 is 10-20% better than JPEG.
Lossless Compression Results
The lossless compression efficiency of the reversible
JPEG 2000 (J2KR), JPEG-LS, lossless JPEG (L-JPEG),
and PNG is reported in Table 7 [59]. It is seen that JPEG
2000 performs equivalently to JPEG-LS in the case of the
natural images, with the added benefit of scalability.
JPEG-LS, however, is advantageous in the case of the
compound image. Taking into account that JPEG-LS is
In both cases the bit-stream header was transmitted without errors. As can be deduced from Table 8, the reconstructed image quality under transmission errors is higher
for JPEG 2000 compared to that of JPEG. However, at
low bit rates (0.25 and 0.5 b/p), the quality of JPEG 2000
decreases more rapidly than JPEG as the error rate increases. It is interesting to observe that at higher error
rates (i.e., 1e-4), the reconstructed image quality in JPEG
2000 is almost constant across all bit rates. This is due to
the fact that in JPEG 2000 each subband block is coded
by bit planes. When the error rate is high enough almost
all blocks are affected in the most significant bit planes,
which are transmitted first. When a particular bit plane is
affected in a block, lower bit planes cannot be decoded
and are therefore of no use. In the case of JPEG the problem is even worse: the higher the encoding bit-rate the
lower the decoded quality. This can be explained by the
fact that when an 8 × 8 block is affected by a transmission
error the entire block is basically lost. The higher the en-
significantly less complex than JPEG 2000, it is reasonable to use JPEG-LS for lossless compression. In such a
case though, the generality of JPEG 2000 is sacrificed.
Error Resilience Results
The evaluation results of JPEG 2000 in error resilience
are shown in Table 8 [19], [59]. A transmission channel
with random errors has been simulated, and the average
reconstructed image quality after decompression has
been measured. In the case of JPEG, the results were obtained by using the maximum amount of restart markers,
which amounts to an overhead of less than 1%. In the case
of JPEG 2000, the sensitive packet information was
moved to the bit-stream header (using the PPM marker
[38]) and the entropy coded data had been protected by
the regular termination of the arithmetic coder combined
with the error resilience termination and segment symbols (the overhead for these protections is less than 1%).
▲ 20. Image “watch” of size 512 × 512 (courtesy of Kevin Odhner): (a) original, and reconstructed after compression at 0.2 b/p by
means of (b) JPEG and (c) JPEG 2000.
▲ 21. Reconstructed image “ski” after compression at 0.25 b/p by means of (a) JPEG and (b) JPEG 2000.
▲ 22. Part of the reconstructed image “cmpnd1” after compression at 0.5 b/p by means of (a) JPEG and (b) JPEG 2000.
coding bit rate, the more the bits needed to code a block,
and therefore the probability of a block being hit by an error and lost is higher for the same bit error rate. In other
words, in JPEG the density of error protection decreases
with an increase in bit rate.
Features and Functionalities
Table 9 summarizes the comparative results of JPEG,
JPEG-LS, and JPEG 2000 from the functionality point of
view. The number of the bullet symbols indicates the
strength by which each functionality is supported. The
Verification Model 8.6 of JPEG 2000 has been used in
this case.
MPEG-4 VTC and JPEG 2000 produce progressive
bit streams. JPEG 2000 provides bit streams that are
parseable and can easily be reorganized by a transcoder on
the fly. JPEG 2000 also allows random access (with minimal decoding) to the block-level of each subband, thus
making possible to decode a region of an image without
Table 7. Lossless Compression Ratios.
decoding the whole image. Notice that MPEG-4 supports coding of arbitrary shape objects by means of an
adaptive DWT, but it does not support lossless coding
[27], [75].
Notice that DCT based algorithms can also achieve
many of the features of JPEG 2000, as ROI, embedded
bit stream, etc. [6], [14]-[16], [31], [51], [52], [53],
[60]. DCT based coding schemes, however, due to the
block-based coding nature, cannot perform well at low
rates, unless post-processing operations are involved [7],
[32]. The complexity of those schemes is increased compared to baseline JPEG and their compression performance is not better than wavelet based coding schemes
(although very close). Additionally, although JPEG 2000
offers better performance than JPEG, different types of
artifacts appear in wavelet based coders. Some results on
post-processing of JPEG 2000 compressed images for tiling and ringing artifact reduction, have already been reported [34], [36].
Overall, the JPEG 2000 standard offers the richest set
of features in a very efficient way and within a unified algorithm [19], [59], [63]. However, this comes at a price
of additional complexity in comparison to JPEG and
JPEG-LS. This might be perceived as a disadvantage for
some applications, as was the case with JPEG when it was
first introduced.
A Note on the Future Parts of the Standard
Part I of the JPEG 2000 standard was studied here. Part I
describes the core coding system, which should be used to
provide maximum interchange. Part II consists of optional technologies not required for all implementations.
Evidently, images encoded with Part II technology will
not be able to be decoded with Part I decoders. As an example, Part II will include trellis coded quantization [38],
[44], user defined wavelets, wavelet packets and other decompositions, general scaling-based ROI coding
functionalities for still image applications, like Internet,
method, etc. Part III will define motion JPEG 2000
color facsimile, printing, scanning, digital photography,
(MJP2) and will be based on Part I of JPEG 2000 [21].
remote sensing, mobile applications, medical imagery,
MJP2 will be used in many different areas, as for example
digital library, and E-commerce. Lossless and lossy codin applications where it is desired to have a single codec
ing, embedded lossy to lossless, progressive by resolution
for both still pictures and motion sequences (which is a
and quality, high compression efficiency, error resilience,
common feature of digital still cameras), or in applicaand lossless color transformations are some of its charactions where very high quality motion pictures are reteristics. Comparative results have shown that JPEG
quired (e.g., medical imaging and motion picture
2000 is indeed superior to established still image comproduction), or in video applications in error prone environments (e.g., wireless and the Internet). The
Table 8. PSNR results in dB for 200 runs of the decoded cafe
standard will allow one or more JPEG 2000
image transmitted over a noisy channel for various bit error
compressed image sequences, synchronized aurates (ber) and compression rates for the JPEG baseline and the
dio and metadata to be stored in the Motion
JPEG 2000.
JPEG 2000 file format (MJ2). Finally, Motion
JPEG 2000 is targeting interoperability with
the JPEG 2000 file format (JP2) and the
MPEG-4 file format (MP4). Part IV of the
JPEG 2000
standard will define the conformance testing.
Part V will define the reference software as high
quality free software. Currently, two reference
JPEG 2000
software implementations do exist, the JJ2000
software in Java and the JasPer software in C
(see Table 6) [1]. Part VI will define a compound image file format. The work plan of
JPEG 2000
these parts is shown in Table 10 [24], [40],
As of this writing several new parts are under
consideration by the JPEG 2000 committee to
JPEG 2000
enhance its features. These are Part VIII for a
reference hardware implementation, Part IX
for mobile wireless specific application features
(JP3G), Part X for secure
JPEG 2000 bit streams
Table 9. Functionality Evaluation Results.
(JP SEC ), P art X I for
interactivity tools: APIs
JPEG 2000
and protocols (JPIP), and
Part XII for volumetric data
Lossless compression performance •••
(JP3D). In addition to the
above, various amendLossy compression performance
ments to different parts of
JPEG 2000 standard are in
Progressive bitstreams
preparation. These curRegion of Interest (ROI) coding
rently include two amendments to p art I, one
Arbitrary shaped objects
amendment to part II, and
one amendment to part V.
Random access
Low complexity
JPEG 2000 is the new standard for still image compression. It provides a new
framework and an integrated toolbox to better address increasing needs for
compression. It also provides a wide range of
Error resilience
noniterative rate control
A bullet indicates that the corresponding functionality is supported. The number of bullet symbols
characterizes the degree of support.
Table 10. Schedule of JPEG 2000 Parts.
Call for
Final Committee
Final Draft
JPEG 2000 Image Coding
System: Core Coding
JPEG 2000 Image Coding
System: Extensions
Motion JPEG 2000
Conformance Testing
Reference Software
Compound Image File
pression standards. Work is still needed for optimizing its
implementation performance. At the time of writing several hardware and software products compliant with the
JPEG 2000 Part I standard have been announced. A list
of them can be obtained from as
well as other relevant and useful information.
The authors would like to thank J. Kovacević and the
anonymous reviewers for their constructive comments.
Various inputs from Diego Santa Cruz and Rafael
Grosbois for the performance comparison section are acknowledged.
Athanassios Skodras received the B.Sc. degree in physics
from Aristotle University of Thessaloniki, Greece, in
1980, the Diploma degree in Computer Engineering and
Informatics in 1985, and the Ph.D. degree in electronics
in 1986 from the University of Patras, Greece. Since
1986 he has been with the Computer Technology Institute, Patras, Greece, and the University of Patras, where
currently he is an Associate Professor at the Electronics
Laboratory. In 1988-1989 and 1996-1997 he was a
Visiting Researcher at Imperial College, London, UK.
His current research interests include image and video
coding, digital watermarking, fast transform algorithms,
real-time digital signal processing, and multimedia applications. He has authored or co-authored over 60 technical papers and four books and holds two international
patents. He is an Associate Editor for the IEEE Transactions on Circuits and Systems II and Pattern Recognition. He
is the Chair of the IEEE CAS & SSC Chapters in Greece
and the Technical Coordinator of the WG6 on image and
video coding of the Hellenic Organization for Standardization. He was the Technical Program Chair for the 13th
IEEE International Conference on Digital Signal Processing (DSP-97). He is the co-recepient of the 2000
Chester Sall Award in the IEEE Transactions on Consumer
Electronics. He is a Chartered Engineer, Senior Member
of the IEEE, and member of the IEE, EURASIP, and the
Technical Chamber of Greece
Charilaos Christopoulos obtained his B.Sc in physics from
the University of Patras in 1989 his M.Sc. in software engineering from the University of Liverpool, UK, in 1991
and his Ph.D. in video coding from the University of
Patras in 1996. From 1993 to 1995 he was a research fellow at the Free University of Brussels. He joined
Ericsson Research in 1995 where he is now Manager of
Ericsson’s MediaLab. He has been head of Swedish delegation in ISO/SC29/WG01 (JPEG/JBIG), editor of the
Verification Model of JPEG 2000, and co-editor of the
JPEG 2000 standard in ISO. He holds 15 Swedish
filed/granted patents. He received the 2000 Chester W.
Sall Award in the IEEE Transactions on Consumer Electronics and third place for the same award in 1998. His research interests include image and video processing,
security and watermarking technologies, mobile communications, 3-D and virtual/augmented reality. He is a Senior Member of the IEEE, Associate Editor of IEEE
Transactions on Circuits and Systems II, member of the editorial board of Signal Processing, and member of the European Association for Speech, Signal and Image
Processing (EURASIP) and the Hellenic Society of
Touradj Ebrahimi received his M.Sc. and Ph.D. degrees
in electrical engineering from the Swiss Federal Institute
of Technology (EPFL), Lausanne, Switzerland, in 1989
and 1992, respectively. In 1993, he was with Sony Corporation in Tokyo. In 1994, he served as a research con-
sultant at AT&T Bell Laboratories. He is currently a Professor at EPFL. He has been the recipient of various distinctions such as the IEEE and Swiss national ASE award,
the SNF-PROFILE grant for advanced researchers, two
ISO-Certificates for key contributions to MPEG-4 and
JPEG 2000, and the best paper award of IEEE Transactions on Consumer Electronics. He is head of Swiss delegation to MPEG JPEG and SC29. He is Associate Editor of
IEEE Transactions on Image Processing, SPIE Optical Engineering Magazine, and EURASIP Image Communication
Journal. His research interests include still, moving, and
three-dimensional image processing and coding, and visual information security. He is the author or the co-author of more than 100 research publications and holds ten
patents. He is a member of the IEEE, SPIE, and IS&T.
[1] M.D. Adams and F. Kossentini, “JasPer: A software-based JPEG-2000
Codec implementation,” in Proc. IEEE Int. Conf. Image Processing, Vancouver, Canada, Sept. 2000, vol. II, pp. 53-56.
[2] M.D. Adams and F. Kossentini, “Reversible integer-to-integer wavelet
transforms for image compression: Performance evaluation and analysis,”
IEEE Trans. Image Processing, vol. 9, pp. 1010-1024, June 2000.
[3] M.D. Adams, “Reversible wavelet transforms and their application to embedded image compression,” M.S. thesis, Univ. Victoria, Canada, 1998.
[4] M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies, “Image coding
using the wavelet transform,” IEEE Trans. Image Processing, pp. 205-220,
Apr. 1992.
[5] J. Askelof, C. Christopoulos, M. Larsson Carlander, and F. Oijer, “Wireless
image applications and next-generation imaging,” Ericsson Review, no. 2, pp.
54-61, 2001. Available review/2001_02/.
[6] E. Atsumi and N. Farvardin, “Lossy/lossless region-of-interest image coding based on set partitioning in hierarchical trees,” Proc. IEEE Int. Conf. Image Processing, Chicago, IL, Oct. 1998, pp. 87-91.
[7] V. Bhaskaran and K. Konstantinides, Image and Video Compression Standards: Algorithms and Applications, 2nd ed. Norwell, MA: Kluwer, 1997.
[8] M. Boliek, J. Scott Houchin, and G. Wu, “JPEG 2000 next generation image compression system features and syntax,” in Proc. IEEE Int. Conf. Image
Processing, Vancouver, Canada, Sept. 2000, vol. II, pp. 45-48.
[9] A. Bovik, Ed., Handbook of Image & Video Processing. San Diego, CA: Academic, 2000.
[10] C.M. Brislawn, “Classification of nonexpansive symmetric extension transforms for multirate filter banks,” Appl. Computational Harmonic Anal., vol.
3, pp. 337-357, 1996.
[11] C.M Brislawn, “Preservation of subband symmetry in multirate signal
coding,” IEEE Trans. Signal Processing, vol. 43, pp. 3046-3050, Dec. 1995.
[12] A.R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo, “Lossless
image compression using integer to integer wavelet transforms,” in Proc.
IEEE Int. Conf. Image Processing, vol. 1. Santa Barbara, CA, Oct. 1997, pp.
[13] C.A. Christopoulos, A.N. Skodras, and T. Ebrahimi, “The JPEG 2000
still image coding system: An overview,” IEEE Trans. Consumer Electron.,
vol. 46, pp. 1103-1127, Nov. 2000.
[14] C.A. Christopoulos, J. Askelof, and M. Larsson, “Efficient methods for
encoding regions of interest in the upcoming JPEG 2000 still image coding
standard,” IEEE Signal Processing Lett., vol. 7, pp. 247-249, Sept. 2000.
[15] C.A. Christopoulos, J. Askelof, and M. Larsson, “Efficient encoding and
reconstruction of regions of interest in JPEG 2000,” in Proc. X European
Signal Processing Conf. (EUSIPCO-2000), Tampere, Finland, Sept. 2000,
pp. 1133-1136.
[16] C.A. Christopoulos, J. Askelof, and M. Larsson, “Efficient region of interest encoding techniques in the upcoming JPEG 2000 still image coding
standard,” in Proc. IEEE Int. Conf. Image Processing (ICIP 2000), vol. II.
Vancouver, Canada, Sept. 2000, pp. 41-44.
[17] C. Chrysafis and A. Ortega, “An algorithm for low memory wavelet image compression,” in Proc. IEEE Int. Conf. Image Processing, vol. III. Kobe,
Japan, Oct. 1999, pp. 354-358.
[18] C. Chrysafis and A. Ortega, “Line-based, reduced memory, wavelet image
compression,” IEEE Trans. Image Processing, vol. 9, pp. 378-389, Mar.
[19] T. Ebrahimi, D. Santa Cruz, J. Askelöf, M. Larsson, and C.
Christopoulos, “JPEG 2000 still image coding versus other standards,” in
Proc. SPIE Int. Symp., San Diego CA, 30 July - 4 Aug. 2000, pp. 446-454.
[20] F. Frey and S. Suesstrunk, “Digital photography—How long will it last?,”
in IEEE Int. Symp. Circuits and Systems (ISCAS 2000), vol. V. Geneva,
Switzerland, 28-31 May 2000, pp. 113-116.
[21] T. Fukuhara, K. Katoh, K. Hosaka, and A. Leung, “Motion-JPEG 2000
standardization and target market,” in Proc. IEEE Int. Conf. Image Processing
(ICIP 2000), Vancouver, Canada, 10-13 Sept. 2000, vol. II, pp. 57-66.
[22] M. Ghanbari, Video Coding: An Introduction to Standard Coders. London,
UK: IEE, 1999.
[23] J.D. Gibson, T. Berger, T. Lookabaugh, D. Lindbergh, and R.L. Baner,
Digital Compression for Multimedia: Principles & Standards. San Mateo, CA:
Morgan Kaufmann, 1998.
[24] M.J. Gormish, D. Lee, and M.W. Marcellin, “JPEG 2000: Overview, architecture and applications,” in Proc. IEEE Int. Conf. Image Processing (ICIP
2000),Vancouver, Canada, 10-13 Sept. 2000, vol. II, pp. 29-32.
[25] V.K. Goyal, “Multiple description coding: Compression meets the network,” IEEE Signal Processing Mag., vol. 18, pp. 74-93, Sept. 2001.
[26] V.K. Goyal, “Theoretical foundations of transform coding,” IEEE Sig.
Proc. Magazine, vol. 18, pp. 9-21, Sept. 2001.
[27] Information Technology—Coding of Audio-Visual Objects—Part 2: Visual,
ISO/IEC 14496-2:1999, Dec. 1999.
[28] JPEG 2000 Part I: Final Draft International Standard (ISO/IEC
FDIS15444-1), ISO/IEC JTC1/SC29/WG1 N1855, Aug. 2000.
[29] JPEG 2000 Requirements and Profiles, ISO/IEC JTC1/SC29/WG1 N1271,
Mar. 1999.
[30] Report on CEV2: Postprocessing for Ringing Artifact Reduction, ISO/IEC
JTC1/SC29/WG1 N1354, July 1999.
[31] Information Technology—Coded Representation of Picture and Audio Information—Lossy/Lossless Coding of Bi-level Images, ISO/IEC JTC1/SC29/WG1
N1359, 14492 Final Committee Draft, July 1999.
[32] Post-Processing Approach to Tile Boundary Removal, ISO/IEC
JTC1/SC29/WG1 N1487, Dec. 1999.
[33] Visual Evaluation of JPEG 2000 Color Image Compression Performance,
ISO/IEC JTC1/SC29/WG1 N1583, Mar. 2000.
[34] JPEG 2000 Verification Model 8.6, ISO/IEC JTC1/SC29/WG1 N1894,
[35] New Work Item: JPEG 2000 Image Coding System, ISO/IEC
JTC1/SC29/WG1 N390R, Mar. 1997.
[36] Call for Contributions for JPEG 2000 (JTC 1.29.14, 15444): Image Coding
System, ISO/IEC JTC1/SC29/WG1 N505, Mar. 1997.
[37] JPEG-LS (14495) Final CD, ISO/IEC JTC1/SC29/WG1 N575, July
[38] Progressive Lossy To Lossless Core Experiment with a Region of Interest: Results
with the S, S+P, Two-Ten Integer Wavelets and with the Difference Coding
Method, ISO/IEC JTC1/SC29/WG1 N741, Mar. 1998.
[39] Core Experiment on Improving the Performance of the DCT: Results with the
Visual Quantization Method, Deblocking Filter and Pre/Post Processing,
ISO/IEC JTC1/SC29/WG1 N742, Mar. 1998.
[59] D. Santa Cruz and T. Ebrahimi, “An analytical study of the JPEG 2000
functionalities,” in Proc. IEEE Int. Conf. Image Processing (ICIP 2000), Vancouver, Canada, 10-13 Sept. 2000, vol. II, pp. 49-52.
[40] Resolutions of 22nd WG1 New Orleans Meeting, ISO/IEC
JTC1/SC29/WG1 N1959, 8 Dec. 2000.
[60] D. Santa Cruz, M. Larsson, J. Askelof, T. Ebrahimi, and C.
Christopoulos, “Region of interest coding in JPEG 2000 for interactive client/server applications,” in Proc. IEEE Int. Workshop Multimedia Signal Processing, Copenhagen, Denmark, Sept. 1999, p. 389-384.
[41] Press Release of the 23rd WG1 Singapore Meeting, ISO/IEC
JTC1/SC29/WG1 N2058, 9 Mar. 2001.
[42] MoMuSys VM, ISO/IEC JTC1/SC29/WG11 N2805, Aug. 1999.
[43] K. Parulski and M. Rabbani, “The continuing evolution of digital cameras
and digital photography systems,” in IEEE Int. Symp. Circuits Syst. (ISCAS
2000), vol. V. Geneva, Switzerland, 28-31 May 2000, pp. 101-104.
[44] J.H. Kasner, M.W. Marcellin, and B.R. Hunt, “Universal trellis coded
quantization,” IEEE Trans. Image Processing, vol. 8, pp. 1677-1687, Dec.
[45] J. Kovacević and W. Sweldens, “Wavelet families of increasing order in arbitrary dimensions,” IEEE Trans. Image Processing, vol. 9, pp. 480-496,
Mar. 2000.
[46] L. Gall and A. Tabatabai, “Subband coding of digital images using symmetric short kernel filters and arithmetic coding techniques,” in Proc. IEEE
Int. Conf. ASSP, NY, 1988, pp. 761-765.
[61] J.M. Shapiro, “Embedded imaging coding using zerotrees of wavelet coefficients,” IEEE Trans. Signal Processing, vol. 41, pp. 3445-3462, Dec. 1993.
[62] A. Bilgin, P.J. Sementilli, F. Sheng, and M.W. Marcellin, "Scalable image
coding using reversible integer wavelet transforms," IEEE Trans. Image Processing, vol. 9, pp. 1972-1977, Nov. 2000.
[63] A.N. Skodras, C. Christopoulos, and T. Ebrahimi, “JPEG 2000: The upcoming still image compression standard,” Pattern Recognition Lett., vol.
22, pp. 1337-1345, Oct. 2001.
[64] W. Sweldens, “The lifting scheme, A custom-design construction of
biorthogonal wavelets,” Appl. Comput. Harmonic Anal., vol. 3, no. 2, pp.
186-200, 1996.
[65] W. Sweldens, “The lifting scheme: Construction of second generation
wavelets,” SIAM J. Math. Anal., vol. 29, no. 2, pp. 511-546, 1997.
[47] J. Liang and R. Talluri, “Tools for robust image and video coding in
JPEG 2000 and MPEG-4 Standards,” in Proc. SPIE Visual Communications
and Image Processing Conf. (VCIP), San Jose, CA, Jan. 1999.
[66] D. Taubman and A. Zalkor, “Multirate 3-D subband coding of video,”
IEEE Trans. Image Processing, vol. 3, pp. 572-578, Sept. 1994.
[48] M.W. Marcellin, M. Gormish, A. Bilgin, and M. Boliek, “An overview of
JPEG 2000,” in Proc. IEEE Data Compression Conf., Snowbird, UT, Mar.
2000, pp. 523-541.
[67] D. Taubman, “High performance scalable image compression with
EBCOT,” in Proc. IEEE Int. Conf. Image Processing, vol. III. Kobe, Japan,
Oct. 1999, pp. 344-348.
[49] I. Moccagata, S. Sodagar, J. Liang, and H. Chen, “Error resilient coding
in JPEG-2000 and MPEG-4,” IEEE J. Select. Areas Commun (JSAC), vol.
18, pp. 899-914, June 2000.
[68] D. Taubman, “High performance scalable image compression with
EBCOT,” IEEE Trans. Image Processing, vol. 9, pp. 1158-1170, July 2000.
[50] M.J. Nadenau and J. Reichel, “Opponent color, human vision and wavelets for image compression,” in Proc. 7th Color Imaging Conf., Scottsdale,
AZ, 16-19 Nov. 1999, pp. 237-242.
[51] D. Nister and C. Christopoulos, “Lossless region of interest with a naturally progressive still image coding algorithm,” in Proc. IEEE Int. Conf. Image Processing, Chicago, IL, Oct. 1998, pp. 856-860.
[52] D. Nister and C. Christopoulos, “An embedded DCT-based still image
coding algorithm,” IEEE Signal Processing Lett., vol. 5, pp. 135-137, June
[53] D. Nister and C. Christopoulos, “Lossless region of interest with embedded wavelet image coding,” Signal Processing, vol. 78, no. 1, pp. 1-17,
[54] T. O’Rourke and R. Stevenson, “Human visual system based wavelet decomposition for image compression,” J. VCIP V. 6, pp. 109-121, 1995.
[55] W.B. Pennebaker and J. L. Mitcell, JPEG: Still Image Data Compression
Standard. New York: Van Nostrand Reinhold, 1993.
[69] D. Taubman, E. Ordentlich, M. Weinberger, G. Seroussi, I. Ueno, and F.
Ono, “Embedded block coding in JPEG 2000,” in Proc. IEEE Int. Conf. Image Processing, Vancouver, Canada, Sept. 2000, vol. II, pp. 33-36.
[70] B.E. Usevitch, “A tutorial on modern lossy wavelet image compression:
Foundations of JPEG 2000,” IEEE Signal Processing Mag., vol. 18, pp.
22-35, Sept. 2001.
[71] M. Vetterli, “Wavelets, approximation and compression,” IEEE Signal
Processing Mag., vol. 18, pp. 59-73, Sept. 2001
[72] W3C, PNG (Portable Network Graphics) Specification, Oct. 1996. Available:
[73] G.K. Wallace, “The JPEG still picture compression standard,” IEEE Trans.
Consumer Electron., vol. 38, pp. xviii-xxxiv, Feb. 1992.
[74] A.B. Watson, G. Yang, J. Solomon, and J. Villasenor, “Visibility of wavelet quantization noise,” IEEE Trans. Image Processing, vol. 6, pp.
1164-1175, 1997.
[56] K.R. Rao and J.J. Hwang, Techniques and Standards for Image, Video and
Audio Coding. Englewood Cliffs, NJ: Prentice-Hall, 1996.
[75] G. Xing, J. Li, S. Li, and Y.-Q. Zhang, “Arbitrarily shaped video object
coding by wavelet,” in IEEE Int. Symp. Circuits and Systems (ISCAS 2000),
vol. III. Geneva, Switzerland, 28-31 May 2000, pp. 535-538.
[57] A. Said and W.A. Pearlam, “A new fast and efficient image codec based on
set partitioning in hierarchical trees,” IEEE Trans. Circuits Syst. Video
Technol., vol. 6, pp. 243-250, June 1996.
[76] W. Zeng, S. Daly, and S. Lei, “Visual optimization tools in JPEG 2000,”
in Proc. IEEE Int. Conf. Image Processing (ICIP 2000), Vancouver, Canada,
10-13 Sept. 2000, vo. II, pp. 37-40.
[58] A. Said and W.A. Pearlam, “An image multiresolution representation for
lossless and lossy compression,” IEEE Trans. Image Processing, vol. 5, pp.
1303-1310, Sept. 1996.
[77] W. Zeng, S. Daly, and S. Lei, “Point-wise extended visual marking for
JPEG-2000 image compression,” in Proc. IEEE Int. Conf. Image Processing
(ICIP 2000), Vancouver, Canada, 10-13 Sept. 2000, vol. I, pp. 657-660.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF