A Dataflow Description of ACC-JPEG Codec
Khaled Jerbi
1,2
, Tarek Ouni
1
and Mohamed Abid
1
1
CES Lab, National Engineering School of Sfax, Sfax, Tunisia
2
IETR/INSA, UMR CNRS 6164, F-35043 Rennes, France
Keywords:
Video Compression, Accordeon, Dataflow, MPEG RVC.
Abstract:
Video codec standards evolution raises two major problems. The first one is the design complexity which
makes very difficult the video coders implementation. The second is the computing capability demanding
which requires complex and advanced architectures. To decline the first problem, MPEG normalized the Re-
configurable Video Coding (RVC) standard which allows the reutilization of some generic image processing
modules for advanced video coders. However, the second problem still remains unsolved. Actually, technol-
ogy development becomes incapable to answer the current standards algorithmic increasing complexity. In
this paper, we propose an efficient solution for the two problems by applying the RVC methodology and its
associated tools on a new video coding model called Accordion based video coding. The main advantage of
this video coding model consists in its capacity of providing high compression efficiency with low complexity
which is able to resolve the second video coding problem.
1 INTRODUCTION
During most of two decades, MPEG has produced
several video coding standards such as MPEG-2,
MPEG-4, AVC and SVC. However, the past mono-
lithic specification of such standards (usually in the
form of C/C++ programs) lacks flexibility. Such spec-
ification does not allow to use the combination of
coding algorithms from different standards enabling
to achieve specific design or performance trade-offs
and thus fill, case by case, the requirements of spe-
cific applications. So as to overcome the intrinsic
limitations of specifying codecs algorithms by us-
ing monolithic imperative code, Caltrop Actor Lan-
guage (CAL) (Eker and Janneck, 2003) (ISO/IEC
FDIS 23001-4: 2009, 2009), has been chosen by
the ISO/IEC standardization organization in the new
MPEG standard called Reconfigurable Video Cod-
ing (RVC). RVC-CAL standard is developped in the
ptolemy2 project (Brooks et al., 2004). It is supported
with a complete framework called OpenDF (Bhat-
tacharyya et al., 2008) and a recently developed com-
piler called Open RVC-CAL Compiler (Orcc) (Jan-
neck et al., 2010) allowing users to define a multitude
of codecs, by combining together actors (called cod-
ing tools in RVC) from the MPEG standard library
written in CAL (Eker and Janneck, 2003) (ISO/IEC
FDIS 23001-4: 2009, 2009), that contains video tech-
nology from all existing MPEG video past standards
(i.e. MPEG- 2, MPEG- 4, etc. ). CAL is used to pro-
vide the reference software for all coding tools of the
entire library. The originality of this paper is the ap-
plication of the CAL and its associated tools on a new
video coding model called Accordion based video
coding (Ouni et al., 2009) (Ouni et al., 2010) (Ouni
et al., 2011). The main advantage of this video coding
model consists in its capacity of providing high com-
pression efficiency with low complexity. Such advan-
tage comes from the original idea that consists in ap-
plying a particular 3D scan, called Accordion, on tem-
poral frames generated by a temporal video decompo-
sition which is able to transform a set of video frames
sequence to a high correlated spatial representation
called IACC. The high correlation of the IACC repre-
sentation, which is actually originated from the video
temporal correlation, is easy and efficiently exploited
by any still image coder. It was shown in (Ouni et al.,
2011) that such coder could produce a high compres-
sion ratio (close to Inter compression models such as
MPEG) with low computational requirements (close
to Intra compression models such as M-JPEG). Ad-
ditionally, this model offers the possibility of easy
reutilizing different mature image compression com-
ponents for video compression purpose. The ACC-
JPEG coder, as an example of Accordion based video
coding model, consists in associated the so called Ac-
102
Jerbi K., Ouni T. and Abid M..
A Dataflow Description of ACC-JPEG Codec.
DOI: 10.5220/0004076901020107
In Proceedings of the International Conference on Signal Processing and Multimedia Applications and Wireless Information Networks and Systems
(SIGMAP-2012), pages 102-107
ISBN: 978-989-8565-25-9
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
cordion process to the JPEG standard image coder
(Ouni et al., 2009). We try to prove, via this coder,
the efficiencyof such model, its low computational re-
quirements and the ease of its implementation which
become much easier with adopting the RVC frame-
work. Actually, we propose -via this work- an easy
and practical solution for fast designing and imple-
menting an efficient video coder.
In section 2, we present the Accordion based cod-
ing principle and we give an overview on the RVC-
CAL methodology. Section 3 analyses through the
proposed CAL based implementation the ACC-JPEG
coder performances (coding and decoding speed,
memory consumption). Section 4 concludes.
2 BACKGROUND
In this section we review the Accordion based coding
approach. Then we details the data flow implementa-
tion and the RVC CAL methodology.
2.1 Accordion
The idea behind the Accordion approach relies on the
hypothesis saying that the video stream contains more
temporal redundancies than spatial ones (Yun Q . Shi,
1999), (Gokturk and Aaron, 2002). In order to take
advantages from this hypothesis, the idea consists in
trying to put pixels -which have a very high tempo-
ral correlation - in spatial adjacency. Thus, video data
will be presented with high correlated form which ex-
ploits both temporal and spatial redundancies in video
signal with appropriate portion that put in priority the
temporal redundancy exploitation.
2.1.1 Accordion Representation
The input of our encoder is the so called video cube
(GoF), which is made up of a number of frames. This
cube will be decomposed into temporal frames which
will be gathered into one 2D representation. Temporal
frames are formed by gathering the video cube pix-
els which have the same column index. These frames
will be scanned while reversing the direction of odd
frames in order to more exploit the spatial correlation
of the video cube frames extremities. This represen-
tation transforms temporal correlation of the 3D orig-
inal video source into a high spatial correlation in the
2D representation (”IACC”) (Ouni et al., 2009) (Ouni
et al., 2010) (Ouni et al., 2011). Figure 1 illustrates
the principle of this representation.
Figure 1: Accordion principle.
2.1.2 Accordion Algorithm
In the following we present the algorithms corre-
sponding to Accordion representation of a sequence
of video frames. The input of this algorithm, called
ACC, has as input a group of pictures (I
K
, for 0 ≤
K ≤ N − 1) called GOP (Group of Pictures) and as
output the resulting IACC image.
Inputs : I0,, IN-1
Outputs : IACC
For x from 0 to (L x N)-1
For y from 0 to H-1
If (x div N mod 2 !=0) Then
n=(N-1) (x mod N)
Else
n= x mod N
End
IACC (x,y)= In (x div N,y)
End
End
The inverse algorithm, denoted ACC-1, has as in-
put the image IACC and as output the set of images
I
K
, for 0 ≤ K ≤ N − 1.
Inputs : IACC
Outputs : I0,, IN-1
For n from 0 to N -1
For x from 0 to L -1
For y from 0 to H -1
If (x mod 2=0) Then
XACC= (N -1) n +(x*N)
Else
XACC= n+ (x*N)
End
In(x,y)=IACC (XACC,y)
End
End
End
Let us note that :
• L and H are respectively the length and the height
of the video source frames.
ADataflowDescriptionofACC-JPEGCodec
103
• N is the number of frames of a GOF.
• IACC(x, y) is the pixel intensity with the coordi-
nates x, y according to accordion representation
repair.
• In(x, y) is the intensity of pixel situated in the Nth
frame in the original video source.
2.1.3 Video Coding Model
In this part, we present the coding diagram based on
the Accordion representation. First, the video encoder
takes a video sequence and passes it to a frame buffer
in order to construct volumetric images by combining
N frames into a stack. Then, the obtained stack will
be transformed to form the accordion representation
(IACC). Here N is the constructed stack depth (N is 8
in our experiments). Next, each IACC will be divided
into N x N blocks to be processed furthermore by the
eventual used 2D transform. The encoder block dia-
gram of the Accordion based compression algorithm
is presented above.
Figure 2: Accordion based video coding model.
This Accordion based video coding model as it is
illustrated in figure 2 shows a great flexibility with dif-
ferent possibilities of extensions and reutilization of
existing image processing tools leading to designing
various versions. In this paper, as it was introduced
above, we are interested by the JPEG version known
as ACC-JPEG.
2.2 Dataflow Implementation
In the following we present the dataflow program-
ming based on the Caltrop Actor Language and the
implementation generation using the back-ends of
Open RVC CAL compiler (Janneck et al., 2010).
2.2.1 CAL Programming
The execution of an RVC-CAL code is based on the
exchange of data tokens between computational en-
tities called actors. Each actor is independent from
the others since it has its own parameters and finite
state machine if needed. Actors are connected to form
an application or a design, this connection is insured
by FIFO channels. This connection is modeled using
XML based dialects as Xml Dataflow Format (XDF).
These languages also provide the possibility to in-
clude parameters when instantiating an actor. Con-
sequently, the same actor may be instantiated several
times with different parameters. We currently use the
Graphiti
1
tool to manage XDF and NL graphs.
Executing an actor is based on firing elemen-
tary functions called actions. This action firing may
change the state of the actor. An action may be in-
cluded in a finite state machine or untagged. An un-
tagged action is higher priority than FSM actions. An
RVC-CAL dataflow model is shown in the network of
figure 3.
FIFO
Actor
Consume/produce
tokens
FIFO
Consume/produce
tokens
ternal states
FIFO
Actions
State
ActorActor
Actions are implemented
sequentially and they can
be
sequenced
FIFO
Actor
be
sequenced
FIFO
Figure 3: CAL actor model.
When an action is fired, it consumes token streams
from input ports and produces token streams to output
ports.
We consider a simple clip actor that clips the con-
sumed tokens values greater than 255 to 255 and those
less than 0 to 0. This algorithm is used in the IDCT2D
process of MPEG4 decoders and it is applied on the
image 8x8 macro blocks. The associated RVC-CAL
code is shown in figure 4 and thus the input streams
are:
INPUT1: INPUT
0
, INPUT
1
... INPUT
63
the output stream is:
OUTPUT: OUTPUT
1
, OUTPUT
2
... OUTPUT
63
such as, for an integer k, OUTPUT
k
=
clip(INPUT
k
)
In this case, the firing rule is the presence of 64
tokens at least in the input FIFO and the availability
of 64 memory cells at least in the output FIFO.
2.2.2 CAL Compiling
Orcc is a CAL compiler that takes in the front-end a
set of actors and a graph (see example of graphiti gen-
1
http://sourceforge.net/projects/graphiti-editor
SIGMAP2012-InternationalConferenceonSignalProcessingandMultimediaApplications
104
actor clipActor ()
(int size=8) INPUT1 ==> int(size=8) OUTPUT:
clip: action INPUT1:[ x ] repeat 64 ==>
OUTPUT:[[
if x[i] > 255 then
255
else
if x[i] < 0 then
0
else
x[i]
end
end : for int i in 0 .. 63 ] ] repeat 64
end
end
Figure 4: RVC-CAL example algorithm of clip actor.
erated graph in Figure 5) that specifies the connexion
between theses actors. After some modifications in
the middle-end, Orcc applies a string template on the
intermediate representation (IR) to generate a chosen
back-end this principle is detailed in Figure 6.
Figure 5: XDF Graph example.
Figure 6: Orcc compilation principle.
Several existing language generators more or less
mature are developped and maintained. We cite: C,
C++, Xlim, Verilog, VHDL, Promela, Java, LLVM
etc. In this paper we use the C back-end because it is
the most efficient and also because a C implementa-
tion is the easiest to debug and validate.
3 IMPLEMENTATION AND
RESULTS
To implement the ACC-JPEG coder using RVC
methodology, we start from an RVC-CAL design of
the JPEG codec (see http://orc-apps.sourceforge.net/).
The design is shown in Figure 7.
Figure 7: JPEG reference design.
We notice that the design of Figure 7 represents
the highest granularity level of RVC-CAL JPEG de-
sign. For example, Figure 8 represents a finer granu-
larity of the decoder block.
Figure 8: Finer granularity of JPEG decoder.
The next step consists in integrating the accordion
algorithm in the RawYCbCr actor and the inverse ac-
cordion in the YCbCrToMB actor as presented in Fig-
ure 9.
We used Orcc compiler to generate the software
implementation of the JPEG design and we obtained
a set of .C files corresponding to every actor of the de-
sign and a top file that manages the actors scheduling
and the FIFO data exchange.
Table 1 gives information about the proposed de-
sign cost in comparison with M-JPEG and MPEG 4
part 2 simple profile standards. Table 1 shows that the
ACC-JPEG algorithmic decoding complexity is close
to the M-JPEG one and it is much less than MPEG 4
part 2 one.
The design is evaluated with CIF and Q-CIF video
sequences. Originally, the video frames look like Fig-
ure 10. By adding the accordion algorithm, we ob-
tained the frames of Figure 11. Figure 12 shows a
* LOC:Line Of Code
ADataflowDescriptionofACC-JPEGCodec
105
Table 1: Complexity comparison between CAL and C de-
scriptions.
decoders level actors Parser Parser
LOC* CAL LOC* C
M-JPEG 2 6 184 979
ACC-JPEG 2 7 184 979
MPEG 4 SP 3 27 1285 4720
comparison in terms of rate-distortion between the
ACC-JPEG, MJPEG and MPEG-4. The results shown
are obtained with the sequence ’Hall Monitor’ (CIF,
25Hz). We considered a general processor as an im-
plementation target. The main configuration features
of this target are presented in table 2
Table 2: Implementation target features.
processor Intel dual core CPU
frequency 2,5 GHz
RAM 2 Go
operating system Microsoft windows XP
Some results about the architecture performances
(processing speed) and requirements in term of (mem-
ory and resources) are given in tables 3, 4, 5. All re-
sults are recorded in the same experimentalconditions
given in table 2.
Tables 3, 4 shows the coding frames frequencies
recorded for CIF and QCIF videos. We can notice
that the ACC-JPEG has almost the same processing
(encoding and decoding) speed than M-JPEG coder.
Table 3: Throughput frequency ACC-JPEG VS M-JPEG
encoders.
encoding FPS
resolution ACC-JPEG M-JPEG
CIF (288x352) 178 182
QCIF (144x176) 748 760
Table refACCJPEGMem illustrates the ACC-
JPEG memory requirements in comparison with other
standards coders.
The memory requirements can be divided into two
by eliminating intermediate tables reserved for the
construction of the Accordion Presentation and thus it
will match the theoretical memory requirements as it
was estimated in (Ouni et al., 2011). Even processing
speed can also be improved and tends perfectly to the
M-JPEG processing speed by adjusting the state ma-
chine which was initially designed for M-JPEG and
then not perfectly adapted to the own ACC-JPEG con-
strains. Such state machine involves large timeouts
resulting to the need to buffering of frames for the set
up of the Accordion representation.
Table 4: Throughput frequency ACC-JPEG VS M-JPEG
decoders.
decoding FPS
resolution ACC-JPEG84 M-JPEG
CIF (288x352) 152 156
QCIF (144x176) 690 702
Table 5: ACC-JPEG memory consumption (GOF = 4).
component memory consumption in Ko
Image size CIF QCIF
ACC 304,128x4 76,032x4
Encoder 3,614 3,614
Decoder 66,080 33,240
ACC inverse 304,128x4 76,032x4
FIFO 65,536 32,768
Total 2568,254 677,878
Figure 9: ACC-JPEG design.
Figure 10: Normal frame.
4 CONCLUSIONS
In this paper a CAL description of a new video coder
is evaluated. The presented work constitutes a demon-
stration of an efficient video coder fast implemen-
tation. Regarding the proposed video coder trade-
off between performances and complexity of the pro-
posed video coder and the tradeoff between efficiency
and cost of the implementation methodology, the pre-
sented work could be considered as an economic and
efficient solution for many video applications includ-
ing embedded systems, digital multimedia devices,
SIGMAP2012-InternationalConferenceonSignalProcessingandMultimediaApplications
106
Figure 11: ACC-JPEG output frames.
Figure 12: Performance comparison.
networks sensors which requires low computational
processing and low time to market. Further orienta-
tions will include the implementation of the second
Accordion-based-coder called ACC-JPEG 2000- pro-
posed in (Ouni et al., 2010) as a more efficient version
than one presented in this paper.
REFERENCES
Bhattacharyya, S., Brebner, G., Eker, J., Janneck, J., Mat-
tavelli, M., von Platen, C., and Raulet, M. (2008).
OpenDF - A Dataflow Toolset for Reconfigurable
Hardware and Multicore Systems. First Swedish
Workshop on Multi-Core Computing, MCC , Ron-
neby, Sweden, November 27-28, 2008.
Brooks, C., Lee, E., Liu, X., Neuendorffer, S., Zhao, Y.,
and (eds.), H. Z. (2004). PtolemyII - heterogeneous
concurrent modeling and design in java (volume 1:
Introduction to ptolemyII). Technical Memorandum
UCB/ERL M04/27, University of California, Berke-
ley, CA USA 94720.
Eker, J. and Janneck, J. (2003). CAL Language Report.
Technical Report ERL Technical Memo UCB/ERL
M03/48, University of California at Berkeley.
Gokturk, S. B. and Aaron, A. M. (2002). Applying 3d meth-
ods to video for compression. In Digital Video Pro-
cessing (EE392J) Projects Winter Quarter.
ISO/IEC FDIS 23001-4: 2009 (2009). Information Tech-
nology - MPEG systems technologies - Part 4: Codec
Configuration Representation.
Janneck, J. W., Mattavelli, M., Raulet, M., and Wipliez, M.
(2010). Reconfigurable video coding a stream pro-
gramming approach to the specification of new video
coding standards. In MMSys ’10: Proceedings of the
first annual ACM SIGMM conference on Multimedia
systems, pages 223–234, New York, NY, USA. ACM.
Ouni, T., Ayedi, W., and Abid, M. (2009). New low
complexity dct based video compression method. In
Proceedings of the 16th international conference on
Telecommunications, ICT’09, pages 202–207, Piscat-
away, NJ, USA. IEEE Press.
Ouni, T., Ayedi, W., and Abid, M. (2010). New non predic-
tive wavelet based video coder: Performances analy-
sis. In Campilho, A. and Kamel, M., editors, Image
Analysis and Recognition, volume 6111 of Lecture
Notes in Computer Science, pages 344–353. Springer
Berlin / Heidelberg.
Ouni, T., Ayedi, W., and Abid, M. (2011). A complete non-
predictive video compression scheme based on a 3d
to 2d geometric transform. International Journal of
Signal and Imaging Systems Engineering, 4(3):164–
180.
Yun Q . Shi, H. S. (1999). Image and video compression
for multimedia engineering fundamentals, algorithms,
and standards. In CRC Press.
ADataflowDescriptionofACC-JPEGCodec
107
