Design of an Error-Tolerance Scheme

for Discrete Wavelet Transform

in JPEG 2000 Encoder
Chun-Lung Hsu, Member, IEEE, Yu-Sheng Huang, Ming-Da Chang, and Hung-Yen Huang
AbstractThe JPEG 2000 image compression standard is designed for a broad range of data compression applications. The discrete
wavelet transformation (DWT), central to the signal analysis and important in the JPEG 2000, is quite susceptible to computer-induced
errors. The errors can be spread to many output transform coefficients if the DWT is implemented by using lifting scheme. This paper
proposes an efficient error tolerance scheme (ETS) to detect errors occurring in DWT. A pipeline-based DWT structure is also
developed in this paper to speed up the error detection process. The proposed ETS design uses weighting sums of the DWT
coefficients at the output compared with an equivalent check value derived from the input-data. With the proposed ETS design, the
errors introduced at DWT can be effectively detected. Additionally, the results of error detection can be further analyzed and evaluated
to show the capability of error tolerance. Some standard images are used as test samples to verify the feasibility of the proposed ETS
design. Experimental results and comparisons show that the proposed ETS has good performance in error detection time and error
tolerance capability.
Index TermsJPEG 2000, DWT, error detection, error tolerance.

1 INTRODUCTION
J
PEG 2000 has recently been approved as an international
standard for the compression of still digital pictures [1].
Unlike JPEG which uses the discrete cosine transform
(DCT), the new standard is entirely wavelet-based to
provide better compression performance and some rich
features. For example, JPEG 2000 allows efficient lossy and
lossless compressions within a single unified coding
framework, provides superior image quality at low bit
rates, supports a more flexible file format, and avoids
excessive computational and memory complexity.
The block diagram of JPEG 2000 encoder is illustrated in
Fig. 1 [2]. The discrete wavelet transformation (DWT) is first
applied on the source image data. The transform coeffi-
cients are then quantized and entropy encoding is applied,
before forming the output codestream (bitstream). The
decoder is the reverse of the encoder. The codestream is
first entropy decoding, inverse quantized, and inverse DWT
(IDWT), thus, resulting in the reconstructed image data. The
DWT is central to the JPEG 2000 image compression
standard which includes lifting configurations for imple-
menting the forward and inverse transforms. The main
properties of DWT are the space-frequency localization and
inherent multiresolution structure. In other words, wavelets
allow efficient representation of a signal with a small
number of nonzero coefficients. Also, wavelets take
advantage of data correlation in space and frequency.
DWT is implemented with computer hardware ultimately,
the processing operations are susceptible to transient
failures, primarily single-event upsets, alternately termed
as soft errors. These factors will increase the influences as
VLSI feature sizes shrink [3], [4]. Although the implementa-
tion of DWT is susceptible by computer-induced errors, the
image quality is also kept within the application-specific or
range of acceptability if an error tolerance structure can be
effectively developed. Thus, design an effective structure
for error tolerance evaluation becomes an important issue of
DWT error detection in JPEG 2000 applications.
Error tolerance is a new design and test paradigm, which
takes into consideration whether erroneous outputs of
defective circuits still produce acceptable results [5], [6].
Error tolerance classifies a system as being acceptable/
unacceptable by estimating the performance degradation
due to errors, rather than relying solely on the conventional
perfect/imperfect classification. Error tolerance analyzes the
system-level effects of errors, and accepts circuits if the
performance degradation can meet the application-specific
or range of acceptability. A common characteristic of all
compression standards for images is that they rely on lossy
compression, that is, the decoded image is not an exact copy
of the original. Thus, in this paper, we view the effect of
errors as potential additional distortion suffered by the
decoded image. This added distortion will in some cases still
lead to an acceptable output. Therefore, image compression
systems are good applications for error tolerance techniques.
This paper proposes an ETS design that targeted for
detecting errors of the DWTsubsysteminJPEG2000, as DWT
is one of the most important subsystems in terms of both
computation and memory requirements. The remainder of
628 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
. The authors are with the Department of Electrical Engineering, National
Dong Hwa University, 1, Sec. 2, Da Hsueh Rd., Shou-Feng, Hualien 974,
Taiwan, ROC. E-mail: cch@mail.ndhu.edu.tw, {d9523006, m9523030,
m9723029}@ems.ndhu.edu.tw.
Manuscript received 6 June 2009; revised 3 Nov. 2009; accepted 16 Dec. 2009;
published online 6 Dec. 2010.
Recommended for acceptance by D. Gizopoulos.
For information on obtaining reprints of this article, please send e-mail to:
tc@computer.org, and reference IEEECS Log Number TC-2009-06-0260.
Digital Object Identifier no. 10.1109/TC.2010.239.
0018-9340/11/$26.00 2011 IEEE Published by the IEEE Computer Society
this paper is organized as follows: Section 2 reviews the basic
principles and key features of DWT in JPEG 2000. The
proposed ETS structure, pipeline-based DWT design, error
model definition, error detection, and tolerance strategy are
presented in Section 3. Section 4 shows the experimental
results and comparisons for performance evaluation and
discussion. Finally, Section 5 provides the final conclusions.
2 BACKGROUND
DWT is usually computed through convolution and
subsampling with a couple of filters to produce an
approximation low-pass filter result and a detail signal
high-pass filter result. The multiresolution decomposition is
obtained by iterating the convolution and subsampling of
these two filters over the approximation components. For
two-dimension (2D) signals, there exist separable wavelets
for which the computation can decompose into horizontal
processing (on the rows) followed by vertical processing (on
the columns), using the same one-dimension (1D) filter. At
each level of the wavelet decomposition, each row of an
image is first transformed using a 1D horizontal analysis
filter-bank. The same filter-bank is then applied vertical to
each column of the filtered and subsampled images,
referred to as subbands. Fig. 2 illustrates a 2D DWT
operation flow. The four subbands of Fig. 2 are denoted
as horizontally and vertically low-pass (LL), horizontally
high-pass and vertically low-pass (HL), horizontally low-
pass and vertically high-pass (LH), and horizontally and
vertically high-pass (HH) [7].
In order to construct filter banks in an efficient way, the
lifting scheme for both designing fast wavelets and
performing the discrete wavelet transform was presented
[7], [8], [9], [10]. The lifting scheme can decompose DWT
filter into several lifting steps. If
e
/: and e q: indicate the
low-pass and high-pass analysis filters, the polyphase
matrix
e
1: is defined as
e
1:
e
/
c
:
e
/
o
:
e q
c
: e q
o
:
!
. 1
Additionally, the polyphase matrix
e
1: can be factor-
ized into a sequence of alternating upper and lower
triangular matrices multiplied by a constant diagonal
matrix, as shown in (2).
e
1:
/ 0
0 /
1
!
Y
i
i1
1 e :
i
:
0 1
!
1 0
e
t
i
: 1
!
. 2
where / is a constant to scale the coefficient, e :
i
: and
e
t
i
:
indicate the predict and update operators, respectively
[11], [12], [13]. JPEG 2000 adopts lifting technique as the
wavelet transform method and, generally, uses biorthogo-
nal (9,7)/(5,3) filter for lossy/lossless compression [11]. The
general block diagram of the lifting technique is illustrated
in Fig. 3, which consists of four steps:
1. Split step. The original samples are separated into
two disjoint sets, named even part and odd part.
2. Predict step. The even samples are multiplied by the
time domain equivalent of e :: and are added to the
odd samples.
3. Update step. The updated odd samples are multiplied
bythe time domainequivalent of
e
t: andare addedto
the even samples.
4. Scaling step. The even and odd samples are multi-
plied by /
1
and /, respectively.
3 PROPOSED ERROR TOLERANCE SCHEME
Fig. 4 shows the proposed ETS design, which consists of an
input parity procedure (IPP), an output parity procedure
(OPP), and a parity analyzer. The main objective of the
proposed ETS is to compare the differences between C
ii
and
C
ont
values to find the errors that occurred in the DWT.
Then, the parity analyzer will further analyze whether the
errors can be tolerated or not. From Fig. 4, each row pixels of
an i i image will be divided into the even and odd
number of data samples input for DWT operation. Thus, the
size of input-data vector X shown in Fig. 4 is 1 i. A is a
i i matrix of wavelet transform via lifting. Additionally,
a 1 i tolerance weighting matrix W, has to be developed
to assist establish the IPP and OPP structures for error
detection. The error models, tolerance weighting matrix W,
pipeline-based DWT design for IPP and OPP structures,
and parity analyzer are explicitly described in the following.
3.1 Modeling Errors
DWT is one of the most important subsystems in the JPEG
2000 encoder. Generally, DWT is surrounded by sets of
multipliers and adders that determine how data flows
through them. Thus, the realistic error assumption has to
be addressed for error detection [14]. However, the
proposed ETS design mainly focuses on the effects of
computer-induced errors, which will be modeled through
transfer matrices related to the lifting sections. For a
HSU ET AL.: DESIGN OF AN ERROR-TOLERANCE SCHEME FOR DISCRETE WAVELET TRANSFORM IN JPEG 2000 ENCODER 629
Fig. 2. Operation flow of a forward 2D DWT.
Fig. 1. Block diagram of JPEG 2000 compression standard.
Fig. 3. General block diagram of the lifting technique.
www.cognitionlabs.in
forward 2D DWT, the numerical errors, caused by an
underlying computer-induced errors, will propagate their
corrupting influence to the output. The computer-induced
errors can be modeled in a way at the numerical processing
[15], [16], [17]. Especially, the data flow structure works in
the forward wavelet transform. Redinbo and Nguyen [17]
clearly indicate how even a single numerical error, caused
by underlying computer-induced errors, propagates its
corrupting influence to many outputs. Thus, any numerical
error effects can be spread by the data processing flow.
Fig. 5 shows a general error model in the numerical
processing. The error model in Fig. 5 does not imply that
errors are additive, but rather, whenever an error occurs, it
may be described by adding a numerical value to the actual
calculated output. Thus, the whole row data of the actual
outputs will be influenced if the multinumerical errors are
presented. Additionally, visibility of DWT basis functions
will depend upon display visual resolution [18]. In terms of
signal that reaches the eye, the visibility will be indistinct if
the display resolution is decreased by a factor of error.
Based on the above-mentioned description, the error
mechanisms can be defined by two types such as intensive
error model (IEM) and distributed error model (DEM). If
the numerical errors will influence the pixels in contig-
uous rows of an image then the error model, IEM is
defined (see Fig. 6a). On the other hand, the DEM (see
Fig. 6b) is presented if many single row pixels of an image
will be influenced by the numerical errors. According to
the error models, the proposed ETS design will be
demonstrated as an effective method to explore the error
impact in the DWT and further analyze the tolerance of
errors for JPEG 2000 encoder applications.
3.2 Tolerance Weighting
In order to meet the human visual system and increase the
flexibility of the proposed ETS design, a tolerance weight-
ing matrix, W, has to be developed. The weighting factors
of W are very important for error tolerance analysis, since
the most significant data are generally centralized in the
central parts of an image. The error influence in the central
parts is more serious than that in the boundaries of an
image [19]. Thus, a weighting matrix has to be built for
supporting the evaluation of error influence when the
proposed ETS is active. The schematic representations of
parity weighting of the benchmark image Lena is shown
in Fig. 7, which divides an image into some blocks to set the
different parity weighting factors. Generally, good parity
weighting factors should have gain factors whose ranges
span three to four orders of pixel magnitude [17]. Based on
the parity weighting factors description in [17], this paper
built a 1 i tolerance weighting matrix for error detection
and error tolerance evaluation purpose.
3.3 Pipeline-Based DWT Design
The main components of the proposed ETS design are the
IPP and OPP structures, whose objectives are generating
the input/output checking values C
ii
and C
ont
and then
deliver them to parity analyzer for error detection. From
Fig. 4, the checking values C
ii
and C
ont
are a weighted sum
of the DWT inputs and outputs. Thus, in order to speed up
the DWT operations and error detection processes, a
pipeline-based DWT structure is proposed for supporting
the IPP and OPP designs.
630 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
Fig. 4. Block diagram of the proposed ETS design.
Fig. 5. Errors in numerical processing. Fig. 6. Error models. (a) IEM. (b) DEM.
www.cognitionlabs.in

Based on the conventional structure for one level 9/7
DWT via lifting [20], Fig. 8 shows the pipeline-based
design for a DWT via lifting. Input-data samples are split
into two subsequences at the beginning of the pipeline-
based DWT structure. For an i i image, the input-data
vector is represented by
X r
0
r
1
r
2
r
3
. . . r
i2
r
i1
.
s
0
r
0
r
2
. . . r
i2

:
0
0
:
0
1
. . . :
0
i,21

.
d
0
r
1
r
3
. . . r
i1

d
0
0
d
0
1
. . . d
0
i,21

.
3
where :
|
i
and d
|
i
represent the ith even and odd samples in
the lth stage of lifting step.
Since the conventional DWT structure uses the parallel
technique to deal with the input samples, the operation time
of the proposed pipeline-based DWT is less than that of the
conventional one. For simplicity, the steps of the ith even
and odd samples shown in Fig. 8 are described as an
example and listed in (4-11).
1. Split step. The input sequences r
i
are split into even
and odd parts, :
0
i
and d
0
i
.
d
0
i
r
2i1
. 4
:
0
i
r
2i
. 5
2. Lifting step. The two splitting sequences (:
0
i
and d
0
i
)
are performed by two lifting steps.
d
1
i
d
0
i
c

:
0
i
:
0
i1

.
:
1
i
:
0
i
2u

d
1
i

.
)
1i)tiiq :tcj 1 6
7
d
2
i
d
1
i

:
1
i
:
1
i1

.
:
2
i
:
1
i
2c

d
2
i

.
)
1i)tiiq :tcj 1 8
9
3. Scaling step. Through the normalization factors /
1
and /, the low-pass and high-pass wavelet coeffi-
cients :
i
and d
i
can be obtained.
d
i
/ d
2
i
. 10
:
i
/
1
:
2
i
. 11
Fig. 8 clearly indicates that the final output of the ith odd
and even samples with two lifting steps can be obtained
after four and five pipeline operations, respectively. The
following samples (:
i1
. d
i1
. :
i2
. . . . ) will be captured in
turn based on the pipeline technique. Fig. 8 also shows that
the proposed pipeline-based DWT design has a good
performance in reducing the operation time. For example,
the operations of lifting step 1 of the ith even sample, :
1
i
and
the (i+1)th odd sample, d
1
i1
are working simultaneously
(see Fig. 8b). The similar case can also be found in Fig. 8c,
which shows that the lifting step 1 of the (i+1)th even
sample, :
1
i1
and the i 2th odd sample, d
1
i2
are working
at the same time. Figs. 8d and 8e illustrate the same cases
with the pipeline technique. According to the pipeline
approach, the final output of DWT can be consolidated in
the following matrix form:
d
s
!

/I 0
0 /
1
I
!
d
2
s
2
" #

/I 0
0 /
1
I
!
I M
a
cM
d
cM
c
!
d
1
s
1
" #

/I 0
0 /
1
I
!
I M
a
cM
d
cM
c
!
I cM
a
uM
d
cuM
b
!
d
0
s
0
" #
.
12
where M
a
, M
b
, M
c
, and M
d
are the i i matrices and the
parameters c. u. . c, and / represent floating point arith-
metic for a 9/7 wavelet transform via lifting [20].
Based on the pipeline-based DWT design, the IPP, and
OPP structures of the proposed ETS design shown in Fig. 4
can be realized by using multipliers and adders. The two
checking values C
ii
and C
ont
from IPP and OPP structures
have to be designed to equal initially for detecting the
errors. From Fig. 4, the checking values C
ii
and C
ont
can be
written as (13) and (14), respectively,
C
ii
XA
T
W
T
. 13
C
ont
WAX
T
. 14
where
W n
0
n
1
n
2
n
3
. . . n
i2
n
i1
.
Figs. 9a and 9b show the corresponding timing charts of C
ii
and C
ont
procedures to evaluate the timing consumption.
Assume that an image has 512 512 pixels using here for
an example. The values C
ii d
0
and C
ii :
0
of the first odd and
even samples of the input-data are obtained at clocks five
and six, respectively, (see Fig. 9a). Based on the input-data
vector shows in (3), the last odd and even input samples
(C
ii d
255
and C
ii :
255
) will be found at the 260th clock and the
261st clock, respectively. The same timing consumption is
also needed by determining the checking value C
ont
(see Fig. 9b). For the timing illustration in Fig. 9, the
HSU ET AL.: DESIGN OF AN ERROR-TOLERANCE SCHEME FOR DISCRETE WAVELET TRANSFORM IN JPEG 2000 ENCODER 631
Fig. 7. The schematic representations of parity weighting of an image.
www.cognitionlabs.in
632 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
Fig. 8. The processes of pipeline-based DWT design.
www.cognitionlabs.in
difference between the values C
ii
and C
ont
can be captured
after 261 clocks if the errors occurred in one row of an
image.
3.4 Parity Analyzer
The parity analyzer shown in Fig. 4 plays the role of a
comparator to check a syndrome (the difference between
C
ii
and C
ont
) and determine whether a syndrome is tolerant
with a chosen threshold or not. According to the error
models in Section 3.1, the thresholds of IEM and DEM are
defined as TH
11`
and TH
11`
, respectively, to detect the
intensive and distributed errors in an image. Equation (15)
indicates that the errors of IEM can be found if the
difference between the checking values C
ii
and C
ont
are
greater than a threshold TH
11`
within the contiguous rows
from jth to kth. However, the definition of IEM detection
in (15) cannot be sure to detect the DEM if some of the
errors are occurred in only one row or some distributed
rows. Thus, a threshold TH
11`
shown in (16) is used to
detect the errors of DEM further:
X
/,1
i,
C
ii i
C
ont i
j j ! TH
11`
. 15
X
i
i1
C
ii i
C
ont i
j j ! TH
11`
. 16
Although the errors of IEM and DEM can be captured by
the mathematical formulas shown in (15) and (16), the
values of thresholds TH
11`
and TH
11`
are hardly used to
determine the capability of error tolerance in a real image.
Generally, the human visual system is more sensitive about
HSU ET AL.: DESIGN OF AN ERROR-TOLERANCE SCHEME FOR DISCRETE WAVELET TRANSFORM IN JPEG 2000 ENCODER 633
Fig. 8. Continued.
www.cognitionlabs.in
the brightness variations than the changes in chrominance.
Thus, the brightness variation detection method [21], which
plays an important role for error detection in an image, is
adopted here to redefine the thresholds of error detection.
For the computational efficiency consideration, the criterion
of brightness variation is determined by using the histo-
gram-based method, which usually shows sensitivity to the
image changes within a similar brightness condition [22].
The histogram difference is defined by
1
H1o

X
H
,
H
i

. 17
where H
i
and H
,
signify the histograms in the ith and
jth rows of an image. By setting threshold of the histogram
difference D
HIS
, the image holding brightness variations can
be detected. Equations (15) and (16) match up the concept
of histogram difference. Thus, two normalization of bright-
ness variation rates (BVRs) derived from (15) and (16) can
be defined and shown in (18) and (19), respectively.
634 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
Fig. 9. Timing chart. (a) C
ii
procedure. (b) C
ont
procedure.
www.cognitionlabs.in
Significantly, in order to accurately evaluate the quality
impact of an image by brightness variations, the BVRs are
normalized as fractional variations. The overall working
flow of the proposed ETS design is shown in Fig. 10:
1\ 1
11`

P
/,1
i,
C
ii i
C
ont i
j j
P
/,1
i,
C
ii i
100%. 18
1\ 1
11`

P
i
i1
C
ii i
C
ont i
j j
P
i
i1
C
ii i
100%. 19
4 EXPERIMENT RESULTS
The 9/7 lifting DWT of the JPEG 2000 image compression
standard is used as a circuit under test (CUT) to
demonstrate the good performance in error detection time
and error tolerance capability of the proposed ETS design.
Consider the lifting DWT operating with error injection
values modeled by a Gaussian noise source [23]. The error
detection performance is strongly dependent on the noise
variation and the selected detection threshold. Based on
the description of detection thresholds shown in [20],
Fig. 11 illustrates the detection performance curves of the
9/7 lifting DWT corresponding to the errors with two
different variations: 10
4
o
2
0
and 10
2
o
2
0
, where o
2
0
repre-
sents the variation of input-data. Experimental results
show that the value, jC
ii
C
ont
j, from the parity analyzer
of the proposed ETS is on the order of 10
10
with round-
off errors. Thus, the necessary thresholds have to be
chosen well above this level for error detection. Fig. 11
clearly indicates that the system has high detection
performance when the thresholds of IEM and DEM
detection are less than 10
4
and 10
7
, respectively. To
examine the effectiveness of the proposed ETS design in
different experimental conditions, six benchmark se-
quences (Lena, Peppers, Baboon, Barbara, Goldhill, and
Cameraman) are selected in the experiments. The compar-
isons between the proposed ETS design and the work in
[17] are presented in this section to demonstrate that the
proposed ETS design has good performance in error
detection time. Additionally, quality observation and
objective analysis are discussed to evaluate the capability
of error tolerance.
4.1 Performance Evaluation
The error detection time significantly affects the circuit
performance when the circuit is under testing. However,
good performance in speeding up the error detection
generally depends on the quality of circuit design. A
pipeline-based DWT design is used of the proposed ETS to
accelerate the speed of error detection.
The computational complexity is the critical factor to
evaluate the performance of error detection time. Based on
Figs. 4 and 13, since the sizes of matrices X and W are both
1 i and A is a i i matrix, the computational complexity
of IPP structure is about 6i
2
i 1 (see Fig. 12a) with one
rowinput-data if the conventional DWT is used. However, a
great reduction of computational complexity can be obtained
by using the proposed pipeline-based DWT design. Fig. 9a
clearly indicates the actuality and shows that the first even
data will be exported after six clocks with 13 times of
multiplication and six times of addition. Thus, for one row
input-data, the computational complexity of the IPP struc-
ture is only 19i. For quickly approximating the operation
speed of a procedure, the notation Big-O is usually used to
estimate the running time of a procedure [24]. Thus, for an
i i image, the computational complexity of IPP structure
with conventional DWT and with pipeline-based DWT are
Oi
3
and Oi
2
, respectively. Additionally, the computa-
tional complexity of OPP structure shown in Fig. 4 is
HSU ET AL.: DESIGN OF AN ERROR-TOLERANCE SCHEME FOR DISCRETE WAVELET TRANSFORM IN JPEG 2000 ENCODER 635
Fig. 10. The overall working flow of the proposed ETS design.
Fig. 11. Detection threshold of the proposed ETS.
www.cognitionlabs.in
evaluated by the data after two lifting steps and the
weighting matrix W. Thus, the computational complexity
of OPP structure is about 2i 1 for one row input samples
(see Fig. 12b). Fig. 12b shows that the operations of DWT are
independent of OPP execution, thus the computational
complexity of OPP is about Oi
2
for an i i image even
if different types of DWT are applied. According to the
above-mentioned discussion, the proposed ETS with pipe-
line-based DWT has better performance in error detection
time than the previous work in [17].
4.2 Error Tolerance Discussion
The quality observation and objective evaluation of bench-
mark images are presented here to discuss the capability of
error tolerance when the proposed ETS design applied to the
JPEG 2000 compression standard. Six 512 512 standard
images, Lena, Peppers, Baboon, Barbara, Gold-
hill, and Cameraman, are selected to experiment. The
experiments of the standard image Lena shown in Fig. 13
act as an example to demonstrate the results of quality
observation. Fig. 13a illustrates the error-free case. Fig. 13b
shows an error case with the value of 1\ 1
11`
is equal to
3.85 percent. Although the quality of image in Fig. 13b is
affectedby the errors, the performance of quality observation
is still good. In other words, the errors occurred in Fig. 13b
can be acceptable by human visual system. However, the
image quality is getting worse when the value of 1\ 1
11`
changes to 6.69 percent. Fig. 13c shows the error case.
Additionally, if the 1\ 1
11`
value increases to 16.51 percent
(see Fig. 13d), the distortion of the image is serious and hard
to observe. Fromthe quality observation in Fig. 13, the values
of 1\ 1
11`
and 1\ 1
11`
of the standard image Lena are
defined to four percent for evaluating the capability of error
tolerance. In other words, if the value of 1\ 1
11`
1\ 1
11`

is less than four percent then the errors occurred in the

standard image Lena can be acceptable, whereas the errors
are unacceptable when the 1\ 1
11`
1\ 1
11`
value is
greater than four percent.
The objective evaluation of a specific image presented
here depends on the relation between BVR and acceptable
error rate (AER). Error rate (ER) is the percentage of vectors
for which values at a set of outputs deviate from error free
response, during normal operation [25]. Thus, the AER in
this paper is defined as (20). Since AER is a percentage of
acceptable errors for all injection errors, AER represents the
capability of error tolerance with injection errors in an image.
11
cccjto/|c 1iioi:
Toto| 1i,cctioi 1iioi:
100%. 20
636 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
Fig. 12. Computation complexity estimation for one row data with
traditional DWT. (a) IPP design. (b) OPP design.
Fig. 13. Quality observation for the benchmark image Lena. (a) Error
free. (b) 1\ 1
11`
3.85%. (c) 1\ 1
11`
6.69%. (d) 1\ 1
11`

16.51%.
www.cognitionlabs.in
Two standard images Baboon and Cameraman are
discussed here (see Fig. 14) to describe the capability of
error tolerance since image Baboon is more complex and
image Cameraman is rather monotone than the other
benchmark images. From the experimental results shown
in Fig. 14, the AER for the images Baboon and Camera-
man are about 12.6 and 5.8 percent, respectively, if the
value of 1\ 1
11`
is set at 4 percent. Thus, if the number of
pixel errors are 262.144 512 512, which are caused by
numerical errors injecting in the lifting section, then about
33,030 and 15,204 pixel errors in the images Baboon and
Cameraman can be acceptable under four percent
brightness variations. Table 1 shows the experimental
results of AERs for six benchmark images with different
brightness variations. The value of 1\ 1
11`
1\ 1
11`
has
to be appropriately defined in accordance with quality
observation for accurately evaluating the AER of a specific
test image.
4.3 Comparisons
The proposed ETS design shown in Fig. 4 can be implemen-
ted by using a comparator and some multipliers, adders,
buffers, and registers. Based on the circuit design shown in
[13], [26], [27], the number of logic gates of the proposed ETS
design is about 6,972. However, about 180 k logic gates are
needed for a VLSI architecture design of JPEG 2000 encoder
[28]. Thus, the area overhead of the proposed ETS design is
only about 3.87 percent, which is a reasonable design for
circuit testing. The comparisons between the proposed
method and previous work [17] are shown in Table 2, which
clearly indicates that the proposed ETS design has good
performance in computational complexity (error detection
time) and error tolerance capability with little area overhead.
5 CONCLUSIONS
An effectively ETS design for lifting DWT error detection
and error tolerance evaluation in JPEG 2000 encoder is
presentedin this paper. The paper first developeda pipeline-
basedDWTstructure to support the proposed ETS design for
speeding up the error detection time. Then, an IPP, an OPP,
and a parity analyzer based on the weighting sum technique
are built of the proposed ETS design to detect the errors. The
error detection performance depends on the detection
thresholds, which are determined by the brightness varia-
tions. Experimental results showthat the proposed ETS with
pipeline-based DWT design can significantly improve the
error detection time compared with the previous work with
conventional DWT structure. Additionally, according the
quality observation and objective evaluation for the test
images, the proposed ETS design also demonstrates the good
performance in error tolerance capability.
ACKNOWLEDGMENTS
The authors wouldlike to thank the National Science Council
of the Republic of China, Taiwan, for financially supporting
this research under Contract No. NSC: 97-2221-E-259-032.
HSU ET AL.: DESIGN OF AN ERROR-TOLERANCE SCHEME FOR DISCRETE WAVELET TRANSFORM IN JPEG 2000 ENCODER 637
TABLE 1
Acceptable Error Rate (AER) for Test Images
Fig. 14. Objective evaluation results.
TABLE 2
Comparison Results
www.cognitionlabs.in
REFERENCES
[1] ITU-T Recommend. T.800-ISO/IEC FCD15444-1: JPEG 2000
Image Coding System, Intl Organization for Standardization,
ISO/IEC JTC1 SC29/WG1, 2000.
[2] C. Christopoulos, A. Skodras, and T. Ebrahimi, The JPEG2000
Still Image Coding System: An Overview, IEEE Trans. Consumer
Electronics Design, vol. 46, no. 4, pp. 1103-1127, Nov. 2000.
[3] C. Constantinescu, Trends and Challenges in VLSI Circuit
Reliability, IEEE Micro, vol. 23, no. 4, pp. 14-19, July/Aug. 2003.
[4] J. Seifert, Z. Xiaowei, and L.W. Massengill, Impact of Scaling on
Soft-Error Rates in Commercial Microprocessors, IEEE Trans.
Nuclear Science, vol. 49, no. 6, pp. 3100-3106, Dec. 2002.
[5] M. Breuer, S. Gupta, and T. Mark, Defect and Error Tolerance in
the Presence of Massive Numbers of Defects, IEEE Trans. Design
Test Computers, vol. 21, no. 3, pp. 216-227, May/June 2004.
[6] H. Chung and A. Ortega, Analysis and Testing for Error Tolerant
Motion Estimation, Proc. 20th IEEE Intl Symp. Defect and Fault
Tolerance in VLSI Systems, pp. 514-522, 2005.
[7] M. Srikar and S. Srinivasan, A Novel Architecture for Lifting-
Based Discrete Wavelet Transform for JPEG 2000 Standard
Suitable for VLSI Implementation, Proc. Intl Conf. VLSI Design,
pp. 202-207, Jan. 2003.
[8] C.J. Lian, K.F. Chen, H.H. Chen, and L.G. Chen, Analysis and
Architecture Design of Lifting Based DWT and EBCOT for JPEG
2000, Proc. Intl Symp. VLSI Technology, Systems, and Applications,
pp. 180-183, 2001.
[9] S.M. Aroutchelvamw and K. Raahemifar, An Efficient Architec-
ture for Lifting-Based Forward and Inverse Discrete Wavelet
Transform, Proc. IEEE Intl Conf. Multimedia and Expo, pp. 816-
819, 2005.
[10] Y.H. Seo and D.W. Kim, VLSI Architecture of Line-Based Lifting
Wavelet Transform for Motion JPEG 2000, IEEE J. Solid-State
Circuits, vol. 42, no. 2, pp. 431-440, Feb. 2007.
[11] I. Daubechies and W. Sweldens, Factoring Wavelet Transforms
into Lifting Steps, J. Fourier Analysis and Application, vol. 4,
pp. 247-269, 1998.
[12] T. Acharya and C. Chakrabarti, A Survey on Lifting-Based
Discrete Wavelet Transform Architectures, J. VLSI Signal Proces-
sing Systems, vol. 42, no. 3, pp. 321-339, Mar. 2006.
[13] B.F. Wu and C.F. Lin, AHigh-Performance and Memory-Efficient
Pipeline Architecture for the 5/3 and 9/7 Discrete Wavelet
Transform of JPEG 2000 Codec, IEEE Trans. Circuits and Systems
for Video Technology, vol. 15, no. 12, pp. 1615-1628, Dec. 2005.
[14] D. Nowroth, I. Polian, and B. Becker, A Study of Cognitive
Resilience in a JPEG Compressor, Proc. Intl Conf. Dependable
Systems and Networks, pp. 32-41, 2008.
[15] N. Seifert, Z. Xiaowei, and L.W. Massengill, Impact of Scaling on
Soft-Error Rates in Commercial Microprocessors, IEEE Trans.
Nuclear Science, vol. 49, no. 6, pp. 3100-3106, Dec. 2002.
[16] P. Shivakumar, M. Kister, S.W. Keckler, D. Burger, and L. Alvisi,
Malertodeling the Effect of Technology Trends on the Soft Error
Rate of Combinational Logic, Proc. Intl Conf. Dependable Systems
and Networks, pp. 389-398, 2002.
[17] G.R. Redinbo and C. Nguyen, Concurrent Error Detection in
Wavelet Lifting Transforms, IEEE Trans. Computers, vol. 53,
no. 10, pp. 1291-1302, Oct. 2004.
[18] A.B. Watson, G.Y. Yang, J.A. Solomon, and J. Villasenor,
Visibility of Wavelet Quantization Noise, IEEE Trans. Image
Processing, vol. 6, no. 8, pp. 1164-1175, Aug. 1997.
[19] J.P. Hornak, The Encyclopedia of Imaging Science and Technology.
John Wiley & Sons, Inc., 2002.
[20] C. Nguyen and G.R. Redinbo, Fault Tolerance Design in JPEG
2000 Image Compression System, IEEE Trans. Dependable and
Secure Computing, vol. 2, no. 1, pp. 57-75, Jan. 2005.
[21] H.W. Haussecker, Simulation Estimation of Optical Flow and
Heat Transport in Infrared Image Sequences, Proc. IEEE Workshop
Computer Vision, pp. 85-93, June 2000.
[22] Y. Shen, D. Zhang, C. Huang, and J. Li, Adaptive Weighted
Prediction in Video Coding, Proc. IEEE Intl Conf. Multimedia and
Expo, vol. 1, pp. 427-430, June 2004.
[23] A. Papoulis, Probability, Random Variable, and Stochastic Process.
McGraw-Hill, 1991.
[24] M. Sipser, Measuring Complexity, Introduction to the Theory of
Computation, pp. 226-228, PWS Publishing, 1997.
[25] D. Shin and S.K. Gupta, A Re-Design Technique for Datapath
Modules in Error Tolerant Applications, Proc. Asian Test Symp.,
pp. 431-437, Nov. 2008.
[26] K.J. Cho, K.C. Lee, J.G. Chung, and K.K. Parhi, Design of Low-
Error Fixed-Width Modified Booth Multiplier, IEEE Trans. Very
Large Scale Integration Systems, vol. 12, no. 5, pp. 522-531, May
2004.
[27] S.W. Cheng, A High-Speed Magnitude Comparator with Small
Transistor Count, Proc. IEEE Intl Conf. Electronics, Circuits and
Systems, pp. 1168-1171, Dec. 2003.
[28] L. Liu, N. Chen, H. Meng, L. Zhang, Z. Wang, and H. Chen, A
VLSI Architecture of JPEG 2000 Encoder, IEEE J. Solid-State
Circuits, vol. 39, no. 11, pp. 2032-2040, Nov. 2004.
Chun-Lung Hsu received the PhD degree in
electrical engineering from National Taiwan
University, Taipei, Taiwan, ROC, in 2000. He is
interested in the relationship of low-power circuit
design, consumer electronics development, sys-
tem on a chip design, and VLSI testing. He is
currently working as an associate professor in
the Department of Electrical Engineering, Na-
tional Dong Hwa University, Hualien, Taiwan,
ROC. The major researches of his laboratory are
3D IC design, built-in self-detect/correct for 3D ICs, and tolerance
scheme development for image system. He is now a member of the
IEICE, IET, and the IEEE.
Yu-Sheng Huang received the bachelor of
engineering degree in computer science and
information engineering from Chaoyang Univer-
sity of Technology, Taichung, Taiwan, ROC., in
2004, and the MS degree in engineering from
National Dong Hwa University, Hualien, Taiwan,
ROC., in 2006. He is currently working toward
the PhD degree at National Dong Hwa Univer-
sity. His research is focused on video IC design,
built-in self-test, and fault/error tolerance.
Ming-Da Chang received the bachelor of
engineering degree in electronic engineering
from Fu Jen Catholic University, Taipei, Taiwan,
ROC., in 2006, and the MS degree in engineer-
ing from National Dong Hwa University, Hua-
lien, Taiwan, ROC., in 2008. He is currently a
hardware engineer at Accton Technology Cor-
poration. His research is focused on built-in
self-test and still image fault/error tolerance.
Hung-Yen Huang received the bachelor of
engineering degree in electronic engineering
from Fu Jen Catholic University, Taipei, Taiwan,
ROC., in 2008. He is currently working toward
the MS degree from National Dong Hwa Uni-
versity. His research is focused on still image
fault/error tolerance.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
638 IEEE TRANSACTIONS ON COMPUTERS, VOL. 60, NO. 5, MAY 2011
www.cognitionlabs.in
www.cognitionlabs.in