Efficient high-performance implementation of JPEG-LS encoder.

Full text

Page 1

Efficient High-Performance ASIC Implementation of JPEG-LS Encoder


Markos Papadonikolakis1, Vasilleios Pantazis2 and Athanasios P. Kakarountas3, 4

1 Alma Technologies, Pikermi Attica, Greece
2 Desics Department, IMEC, Leuven, Belgium
3 VLSI Design Lab., Electrical & Computer Engineering Department, University of Patras, Rio, Greece
4 Department of Informatics on Biomedicine, University of Central Greece, Lamia, Greece


Abstract - This paper introduces an innovative design
which implements a high-performance JPEG-LS encoder. The
encoding process follows the principles of the JPEG-LS
lossless mode. The proposed implementation consists of an
efficient pipelined JPEG-LS encoder, which operates at a
significantly higher encoding rate than any other JPEG-LS
hardware or software implementation while keeping area
small.

Index Terms - Image processing, lossless compression, JPEG-
LS, LOCO-I, VLSI implementation.

I. INTRODUCTION
There are two main categories of image compression
algorithms, the lossy and the lossless compression algorithms.
The first category, which is more widely used, defines
encoding processes that omit parts of the data to be encoded,
exploiting the redundancy of the visual information. Hence the
lossy algorithms present much higher ratios of compression
than the lossless ones, which, as implied by their name, keep
the data to be encoded intact. However the latter group of
algorithms allows the complete reconstruction of the encoded
image without loss of information, a property which lossy
algorithms lack. Baseline JPEG is a very commonly used
lossy compression algorithm and JPEG-LS, which is the
targeted algorithm of this paper, is a typical lossless algorithm.
In general, lossless algorithms are used in many specialized
applications that emphasize more on the attainment of high
fidelity and less on the compression ratio. These applications
need to process images that may be satellite captured, medical
or biometric. Furthermore, professional photography is a field
in which lossless algorithms are able to thrive successfully.
The aforementioned group of applications presents a
fundamental requirement. Large volumes of image data have
to be compressed and transferred in real time. The latter
disallows the use of software implementations, while in the
same time renders the need for the development of hardware-
implemented solutions imperative, as their advantages in
terms of throughput, especially when considering an ASIC
implementation, are widely known. Furthermore their
operation is not impeded by other processes, something very
common in the field of software implementations, which have
to share their CPU time with other applications.
The JPEG-LS is an established standard for lossless and
near-lossless compression of grayscale and color continuous-
tone images [1]. It provides better compression ratios than the
lossless JPEG standard, while maintaining a fairly low level of
complexity. In the heart of the JPEG-LS lies the algorithm
LOCO-I (Low Complexity Lossless Compression for Images)
[2] and [3], a product of the collaboration of M. J. Weinberger
and G. Seroussi (Hewlett Packard laboratories). The main
feature of LOCO-I is that it combines the effective
compression models proposed by context modeling with a low
complexity level. This is achieved by the use of a
combinational method that matches a simple coding unit to its
modeling unit, while avoiding composite operations. Thus a
low complexity instance of the universal context modeling is
obtained. These attributes allow one of the highest
compression ratios when compared to other lossless
compression algorithms. Needless to say that most of those
algorithms are also significantly more complex. However the
potential of the JPEG-LS while in near-lossless mode is
limited, and its performance in aspects of compression is not
as good as that of other lossy compression algorithms (lossy
JPEG). This does not affect the range of applicability of
JPEG-LS in the fields of satellite or medical image
compression, which require the reconstruction of the original
information without alterations after the decoding process.
During the development process it was decided to
propose a hardware implementation of JPEG-LS which
implements the highly effective LOCO-I lossless compression
mode and not the near lossless one. That decision presented
two more advantages. First of all it reduces further the
complexity of the algorithm, which leads to a smaller in terms
of area and less energy-consuming integrated circuit. Apart
from that by removing the elements of the LOCO-I algorithm
that are used only in the near-lossless mode, the algorithm
itself is simplified and its stages can be lessened. This can be
exploited by applying various design techniques, leading to
the reduction of the overall maximum delay that is necessary
for the execution of the operations of the algorithm.
The designing process resulted in the ASIC implementation
of an encoding unit which increases the throughput about five
times when compared to
implementations, while maintaining low area requirements in
such a way that renders it suitable for use in portable
communication devices. The main achievement of this work is
the introduction of a real-time, small-sized, high-performance
JPEG-LS lossless encoder, which maintains a wide range of
other proposed ASIC
978-3-9810801-2-4/DATE07 © 2007 EDAA

Page 2

applicability.
In the following sections of this paper we will delve into
the operations of the algorithm and the architecture of the
proposed hardware solution. The second section contains an
overview of the JPEG-LS algorithm and a presentation of its
most important functions. Section III analyzes the proposed
JPEG-LS implementation in depth, providing details regarding
its architecture, logic and any modifications made in order to
decrease its critical path, as well as the prospective features of
the integrated circuit. In Section IV the proposed JPEG-LS is
implemented using ASIC technology. Its performance is
compared to that of other implementations. Finally, in Section
V the paper concludes.
II. THE JPEG-LS ALGORITHM
The JPEG-LS algorithm consists of two main units, a
context modeler and an encoder. The context modeler uses the
values of the adjacent to the current (the pixel to encode)
pixels, correlates it to a specific context and predicts the
current pixel value. The statistics of the selected context,
which are based on past data, are used for the calculation of
the prediction error and the coding variables. The statistic
parameters are updated with the new information derived from
the current pixel, and the coding unit encodes the prediction
error using a Golomb-Rice code.
LOCO-I also uses a run mode, in order to compress
constant regions. Run mode consists of a run scanner, which,
when operation is set to run mode, checks if the current pixel
is equal to the previous one; if so, run mode continues, else the
modeler enters run interruption mode. Run coder either
encodes the run length or the pixel that caused the run
interruption. A closer look at the operations which comprise
JPEG-LS follows.
A. Context Modeler
When in normal mode, the context modeler uses the values
of the neighboring to the current pixels, in order to compute

Rd RbRc
x Ra

Fig. 1. Causal template.

the local gradients D1= Rd - Rb D2 = Rb - Rc and D3= Rc –
Ra. The neighboring pixels Ra, Rb, Rc and Rd compose the
“causal template” of the current pixel x, as illustrated in fig.1.
The local gradients are quantized and the modeler, using
these three quantized vectors, selects one of 365 available
contexts for the current pixel. Four statistic parameters are
kept for each context. These consist of the accumulator of the
magnitudes of previous prediction errors Α[Q], the bias
variable B[Q], which accumulates the errors’ values, the
prediction error correction parameter C[Q] and the
occurrences counter N[Q].
The modeler uses a median edge detector to predict the
value of the current pixel Px:
min(,) if
max(,)f
otherwise.
RaRb Rc
+− ,

This operation is followed by the computation of the
prediction error, which is corrected using the bias cancellation
variable C[Q], and clamped in the appropriate range for the
coding unit.
The final step before the encoding of the prediction error
involves the calculation of the value of the Golomb variable k
and the update of the current context statistic parameters. The
derivation of k is performed using the parameters Α[Q] and
N[Q]:
for (k=0; (Ν[Q]<< k)< Α[Q]; k++);
The prediction error is mapped to a non-negative value,
with the use of variable k and the parameters B[Q] and N[Q],
as Golomb codes can only be applied on positive values.
After the computation of Golomb variable k, the statistics
are updated with the new prediction error. The bias correction
parameter C[Q] is updated with at most one unit per iteration,
according to the updated value of the bias variable B[Q]. This
preserves the low complexity of LOCO-I.
max(
min(
≤
,
,
)
)
Ra Rb
Ra Rb
Rc Ra Rb
Ra Rb

Px i Rc
,
,
≥




=

B. Golomb Code
The operations of the context modeler are followed by the
encoding of the mapped prediction error. This is performed
using the optimal codes for a Two Sided Geometric
Distribution (TSGD) [4], based on Golomb codes [3]. The
encoding procedure is carried out as follows:
• If the number formed by all but the k least significant bits
of the mapped error value is less than LIMIT – qbpp – 1, this
number is encoded in the output bit stream in unary
representation. This means that the number is represented with
a number of zeros equal to its value, which are followed by a
single ‘1’. Then the k least significant bits are appended to the
output bit stream unaltered.
• If the aforementioned condition does not apply, a number
of LIMIT – qbpp – 1 zeros and a single binary one is encoded
in the output bit stream to represent a maximum length code.
The entire mapped error value reduced by 1 is then encoded in
the output bit stream unchanged, using qbpp bits. LIMIT and
qbpp stand for the maximum number of bits of the Golomb
code and the number of bits necessary to represent the mapped
error value respectively. For 8-bit images, LIMIT is equal to
32 and qbpp to 8.
C. Overview
The LOCO-I block diagram (based on [2]) is illustrated in
fig. 2. It is not difficult to notice the existence of a data-
dependency loop in the algorithm flow. Every time that the
context modeler determines that two subsequent samples
belong in the same context, the previously updated parameter

Page 3

RdRb Rc
Rax
Local
Gradients
Flat
Region
?
Median
Edge
Detector
Bias
Cancellation
Context
Updater
Coder
variables
computer
Run
Scanner
Error
Mapper
Golomb
Coder
Run
Golomb
Coder
Coder
Context
Modeler
regular mode
run mode
regular mode
run mode
bitstream
image pixels
-
+

Fig. 2. LOCO-I block diagram (based on [2]).
C[Q] has to be used in order to perform the bias cancellation
of the prediction error of the next sample. Thus the
aforementioned loop includes the functions of the prediction
error calculation and the statistic parameters’ update. The
statistic parameters can not be updated before the execution
of the computations necessary to calculate the value of
Golomb variable k and the mapping of the prediction error,
operations that require the current values of Α[Q], B[Q] and
N[Q].
III. PROPOSED JPEG-LS IMPLEMENTATION
The main design technique used in this implementation is
pipelining. It leads to a significant reduction of the critical
path and thus to a high maximum throughput. It has been
shown that the execution of the LOCO-I involves several
functions and processes, from the context determination to
the encoding of the mapped error. If an attempt to implement
the algorithm’s flow in a single stage was made, a critical
path of unacceptable delay would be produced. This would
limit extremely the maximum operating frequency. However,
the distribution of the processes of the algorithm into more
than one stage, via pipelining, minimizes the critical path and
reduces the overall maximum delay significantly.
A. JPEG-LS Data Path
The most important aspect while breaking down the
algorithm into stages is the placement of the pipeline
registers. It is fundamental for delay-equivalent algorithmic
phases to be produced. The placement of such registers after
the context determination for the current sample, as well as
before the encoding of the mapped error, is apparent as these
processes are data-independent. The difficulty lies in
separating the data-dependent processes such as the
calculation of the prediction error and the update of the
context statistic parameters. The calculation of the following
prediction error needs the previously updated bias
cancellation parameter C[Q] so pipelining can not be applied
in this data-dependent loop. The critical path of the design is
then observed in the data-loop circuit, including the
prediction error calculation and the update of the statistic
parameters. These functions are complex, resulting in a
critical path whose delay would be higher than the desirable
for a high-performance core.
Nevertheless, the low-complexity scheme of LOCO-I
indicates that the update of the bias cancellation parameter
C[Q] is limited to at most one unit per iteration. Thus, the
next value of the bias parameter for the same context will be
the same, reduced, or incremented by one. Since there are
only three possible values for the updated parameter C[Q],
the look-ahead technique can be applied to the computation
of the error residual. The prediction error can be estimated
for these three possible values and then stored in a pipeline
register. At the next stage it will be possible to decide which
one of the three estimated errors has to be used, by
comparing the previous value of C[Q] with the updated one,
as demonstrated in fig. 3.
The pipeline register’s context update feeds the bias
cancellation value to the prediction estimation only when two
subsequent samples occur in the same context. Otherwise,
C[Q] is provided by the context parameters’ memory. It is
necessary for that reason to have a double sized memory in
order to store the values for bias cancellation C[Q], since two
different values of bias cancellation have to be read from the
memory in the same clock cycle: one for the error estimation
and one for the parameters’ update.
Apart from that, due to the fact that an extra clock cycle is
required after the context determination in order to access the

Page 4

reg
error
estimation
1
error
estimation
2
error
estimation
3
C[Q] pro
error est 2
error est 1
context
variables
update
+
-
'1'
'1'
<=>
reg
error est 3
C[Q] upd

Fig. 3. Prediction error calculation with estimated bias cancellation.

selected context’s parameters from the memory, it is possible
to apply an additional pipeline stage before the error
estimation phase. During the process of addressing the
context memory several pre-computations which will
speedup the error estimation of the next phase can be
performed. For instance, Px can be subtracted from Ix
beforehand, since both their values are already available (Ix
and Px stand for the error and the Golomb variable k). Two
pipeline stages have been used to implement the encoding
process. The first stage produces the code for the current
mapped error, while the second concatenates the code and
produces the final encoded bitstream as its output. The
calculations that need to be executed for the run mode scan
can be implemented in the first pipeline stage, while
determining the context of the current sample. The technique
of pipelining has been used also in the implementation of the
run interruption encoding procedures, even though those
operations are simpler than when in normal mode, which
makes their effect on the overall delay of the critical path
negligible.
The data path’s block diagram is illustrated in fig. 4. The
first stage involves the determination of the context and
computation of the prediction of the median edge detector.
necessary to access the desired context parameters. The The
pre-computations of the errors take place in the second stage,
while the memory’s decoding unit produces the address
estimation of the prediction error is carried out in the third
stage. It is done using the last updated value of C[Q]. In the
fourth stage the selection of the correct estimated value of the
prediction error is made. The next step involves the
computation of the value of the Golomb variable and the
mapping of the errors, while the context updater calculates
the new values of the context statistic parameters. The
mapped error is encoded in the two final pipeline stages.
The operations that take place in stage 4 are the ones that
present the critical path of the design. In that stage the
comparison between the previous and the updated value of
C[Q] is carried out. It includes the multiplexer for the choice
of the correct error estimation and also the parameters’
update. The update is performed concurrently to the Golomb
variable calculation and the error mapping, but its operations
are slower.
Context
Selector
Predictor
Error
Precomputation
Context
Memory
Address
Decoding
Bias
Cancellation
Error
Estimator
Context Modeler
Error
Mapper
Coder
Variables
Computation
Context
Updater
Code
Producer
Code
Appender
Golomb Coder
Stage V Stage VIStage IV Stage IIIStage II Stage I
bitstream
image
samples

Fig. 4. LOCO-I’s modified data path.

db
x
c
a
t-1
a
t-4
a
t-2
a
t-1
a
Previous line
in Memory
Current line in
Encoder
t-1
c

Fig. 5. Use of image line memory.

Page 5

Context
Parameters
Memory
Image
Line
Memory
Context Memory Interface
Input
Buffer
&
Line
Memory
Interface
bitstream
image
samples
Context Modeler Golomb Coder
JPEG-LS Encoder
Stage
I
Stage
II
Stage
III
Stage
IV
Stage
V
Stage
VI

Fig. 6. Proposed JPEG-LS architecture.


B. Memory Requirements
The size of the required memory for the JPEG-LS
encoding process depends on the storage of the context
parameters and of the previously encoded samples. The C [Q]
memory is double-sized, since two different addresses need
to be accessed concurrently, in order for the operations of
stage three and four to be carried out. The bit lengths of
Α[Q], B[Q], C[Q] and N[Q] for 8-bit images are 13, 7, 8 and
7, respectively. They are deduced from the possible range of
their values. Thus the total size of the context memory is
about 1.9 KB.
The minimum required memory needed to store the
neighbor samples’ values is one image line as demonstrated
in fig. 5. Only the values of the neighbors a, b, c and d have
to be accessed for the context determination of the current
pixel x. A technique which allowed us to lessen the overall
size of the memory is presented next. Sample x of instance t
will become the neighbor a for the following image sample,
thus it can be stored in a single register. The pixel c of t-1
instance will be no longer necessary as a neighbor so it can
be replaced. Thus the memory size is minimized. Only one
image line is required, while the works in [6] and [8] use a
memory size of two image lines. This minimizes the total
area of the ASIC design, since the on-chip memory consumes
most of the area in such designs. The memories used are
DUAL-RAMs, making it feasible to perform both read and
write operations in the same clock cycle.
C. JPEG-LS Architecture
The architecture of the proposed JPEG-LS encoder is
illustrated in fig. 6. It consists of the LOCO-I pipelined data,
the context and the image line memory and two memory
control interfaces. The line memory interface addresses the
line memory, stores the new neighbor samples and feeds the
LOCO-I data path with the values of the neighbors of the
current sample. The context memory interface addresses the
determined context parameters and feeds the error estimation
stage with the last updated value of C[Q] as well as the fourth
stage with the parameters of the context which was
previously addressed. Finally the updated values are stored
back in the context memory.
IV. IMPLEMENTATION AND RESULTS
The design was captured in VHDL and was fully
simulated and verified using the Model Technology’s
ModelSim Simulator. The goal of this work was to propose a
competitive hardware solution, so the design was synthesized
for ASIC technology. Synthesis was performed in a Synopsys
environment using several process libraries. The back-
annotated information that was extracted from the synthesis
tool was used to carry out the simulation of the design. The
verification of the design’s functionality and compatibility
was based on the collection of test images [1] and on various
popular testing images with the use of HP LOCO-I
executable [5] as software benchmark.
The characteristics of the JPEG-LS implementation are
demonstrated in table I, while the corresponding properties of
other implementations available
demonstrated in table II.
The authors have tried their best to be fair to the
comparison with other works. Surprisingly, none of the
available works gives a full report of the implementation’s
characteristics, in terms of area, operation frequency and
throughput. Thus, the authors have assumed for some
implementations that throughput is always the theoretic
maximum and this is how table II was derived. At this point,
it should be mentioned that implementation suggested by [7],
also available in the scientific literacy, does not state its
overall characteristics specifically.
Table III contains a comparison of the proposed TSMC
0.18µm implementation to other implementations available in
academia. From the following table, it is easily observed that
the proposed JPEG-LS implementation presents significantly
higher throughput than any other available implementation,
while minimizing the area requirements. This was expected
due to the simplicity of the proposed implementation’s
algorithm and the application of pipeline stages at critical
points of the encoder. It has to be remarked that the
manipulation of the LOCO-I functions in a way that allowed
us to produce a simpler design, of course with no effect on
in academia are

Page 6

the functionality of the algorithm, has offered significant
gains in terms of throughput.

TABLE I
CHARACTERISTICS OF THE PROPOSED JPEG-LS IMPLEMENTATION FOR
THE TARGETED ASIC TECHNOLOGIES
LOGIC
AREA
(EQ.
GATES
)
ASIC
TECH.
MEM
USAGE
(BITS)
OPERATING
FREQUENCY
(MHZ)
THROUGHPUT
(MPIXELS/S)
TSMC
0.18 µm
27681
23887 *
20815 **
183 183
UMC
0.18 µm
28127
23887 *
20815 **
160 160
TSMC
0.09 µm
Sage-X
Artisan
* For 1024 pixels per image line
** For 640 pixels per image line

25709
23887 *
20815 **
265 265
TABLE II
CHARACTERISTICS OF OTHER JPEG-LS IMPLEMENTATIONS
LOGIC
AREA
(EQ.
GATES)
WORK TECH.
MEM
USAGE
(BITS)
OPERATING
FREQUENCY
(MHZ)
THROUGHPUT
(MPIXELS/S)
[6] * -
49,457
gates
32768
(1024)
66
66 upper limit
(theoretic)
[8] *
0.18
µm
70,000
gates
24000
(640)
40
40 upper limit
(theoretic)
* Implementing lossless and near-lossless mode


TABLE III
COMPARISONS WITH THE OTHER JPEG-LS IMPLEMENTATIONS
COMPARED
TO THE
PROPOSED
AREA
(%)
[6] 178.7 %
[8] 252.9 %
WORK
COMPARED
TO THE
PROPOSED
MEMORY
(%)
137 %
115 %
COMPARED
TO THE
PROPOSED
THROUGHPUT
(%)
36.1%
21.8%

Table III reveals that the proposed implementation is by
far the most efficient in terms of throughput as it presents
almost five times higher throughput than the other ASIC
implementations. Also its requirements in terms of area are
significantly lower, being at least 1.7 times less than those of
other ASIC implementations.
V. CONCLUSION
This paper presented a novel implementation of a lossless
JPEG-LS hardware encoder. The pipelining design
techniques used allowed the full exploitation of the low-
complexity features of the LOCO-I algorithmic functions
leading to a very high operating frequency, the highest ever
reported for a JPEG-LS encoder in academia or industry. The
proposed implementation presents up to five times higher
throughput than any other available ASIC JPEG-LS
implementation.
Apart from its outstanding performance in speed, the
proposed design approach is quite efficient in terms of area,
where its requirements are very low. Finally, as it was
manifested in the section containing the results of the
implementation, the proposed encoder is very cost-effective
while being fully defined in terms of area, operation
frequency and throughput.
ACKNOWLEDGMENTS
We would like to thank the European Social Fund (ESF),
Operational Program for Educational and Vocational
Training II (EPEAEK II) and particularly the program
PYTHAGORAS, for funding the above work.
REFERENCES
Technology-Lossless [1] Information
compression of continuous-tone images-Baseline. International
Telecommunication Union (ITU-T Recommendation T.87).
ISO/IEC 14495-1, 1998.
[2] M.J. Weinberger, G. Sapiro and G. Seroussi, “The LOCO-I
Lossless image compression algorithm: Principle and
standardization into JPEG-LS”, IEEE Trans. on Image
Processing, vol. 9, Aug. 2000, pp. 1309-1324.
[3] S. W. Golomb, “Run-length encodings”, IEEE Trans. Inform.
Theory, vol. IT-12, pp. 399–401, July 1966.
[4] M. J. Weinberger, and G. Seroussi, “Optimal prefix codes for
sources with two-sided geometric distributions,” IEEE Trans.
Inform. Theory, vol. 46,pp.121–135, Jan.2000.
[5] M. J. Weinberger, and G. Seroussi, “From LOCO-I to the
JPEG-LS Standard”, Computer Systems Laboratory, HP
Laboratories Palo Alto,
http://www.hpl.hp.com/loco/
[6] A. Savakis and M. Pioriun, "Benchmarking and Hardware
Implementation of JPEG-LS", Rochester, NY, Sept. 2002,
International Conference on Image Processing Proceedings
(ICIP ’02), Volume 2 pp II-949- II-952.
[7] M Ferretti, M. Boffadossi,
Implementation for JPEG-LS”,
Systemistica, Univ. Pavia, Italy, 17th International Conference
on Pattern Recognition (ICPR'04) - Volume 1 pp. 769-772.
[8] Xiang Xie, GuoLin Li and ZhiHua Wang, “A Near-lossless
Image Compression Algorithm Suitable for Hardware Design
in Wireless Endoscopy System”, Department of Electronic
Engineering, Tsinghua University, Beijing, P. R. China,
100084, ASICON 2005. 6th International Conference On
ASIC, Volume 1, 24-27 Oct. 2005 pp. 37 – 40.

and near-lossless
HPL-1999-3, January 1999.
“A Parallel
Dip.
Pipelined
Informatica e