Comparison between JPEG2000 and H.264 for digital cinema.

Full text

Page 1

COMPARISON BETWEEN JPEG2000 AND H.264 FOR DIGITAL CINEMA

Boxin Shi, Lin Liu, Chao Xu

State Key Laboratory of Machine Perception
Peking University
Beijing, China
Email: {shiboxin, liulin, xuchao}@cis.pku.edu.cn


ABSTRACT

JPEG2000 and H.264 are the latest image and video coding
standards respectively. Digital cinema is a new kind of
application for super high definition video. The DCI
(Digital Cinema Initiative) specification published in 2005
chose JPEG2000 instead of H.264 as the video coding
standard for digital cinema, which showed that JPEG2000
had a better performance in the field of super high definition
video coding. Until now, only a few basic tests have been
done to compare between JPEG2000 and H.264, moreover
the tested resolutions and video contents were very limited.
In this paper, based on JPEG2000, H.264 intra-frame and
inter-frame coding, we compared the coding efficiency and
subjective image quality using multiple series of test
sequences from low to super high definition. The
experiment results demonstrate some regularity for
JPEG2000 and H.264 video coding, and reveal JPEG2000
is more suitable for digital cinema.


Index Terms—JPEG2000, H.264, Digital Cinema

1. INTRODUCTION

In recent years, digital multimedia devices are more and
more popular in people’s daily life. With the development
of digital technology, digital cinema has been putting into
practice. To establish an necessary industrial standard for
digital cinema, a joint venture of 7 major Hollywood studios,
Digital Cinema Initiatives, LLC（DCI） founded in March
2002, putted forward a DCI specification in July 2005.
DCI specified JEPG2000 with Daubechies 9/7 wavelet
transform as the image compression scheme in digital
cinema. Every frame in a movie must have a high color
depth (12bits or more), a 4: 4: 4 color space and super high
definition of 2K (actually 2K×1K) or 4K (4K×2K) [1].
In the motion picture compression field, H.264
published in 2003 is a well-known excellent video standard.
It not only performs well in low bit rate video coding, but
also in the higher ones, such as HDTV coding. JPEG2000
published in 2001 is mainly a superior still image coding
standard. Digital cinema is usually classified as a kind of
motion picture. JPEG2000 but not H.264 was adopted by
DCI for the movie image compression, which makes it a
motivation for us to figure out the reason.
As we known, some researches have been done to
compare their performances in still image and super high
definition video coding. Aravind et al. [2] give a conclusion
after testing five standard still images that H.264 intra-frame
coding (without motion compensation) generates better
peak signal-to-noise ratio (PNSR, all the PSNR in this paper
is computed from luminance component only) than
JPEG2000 at various bit rates especially at the higher ones.
Michael Smith [3] found that the performance of H.264
intra-frame, inter-frame coding and JPEG2000 is dissimilar
when experimenting on two different DCI evaluation
materials at the resolution of 512×208（272）and 4K.
One result is that the three performances are similar and the
other result is that H.264 inter-frame performs best and
follows the JPEG2000, then H.264 intra-frame at bit rates
from 0 to 3 bit/pixel.
In this paper, we compared the objective and subjective
image quality of H.264 intra-frame, inter-frame and
JPEG2000. The test sequences involve a wider range, from
176×144, 352×240 and 720×576 standard sequences with
different contents to 512×256, 1024×512, 2048×1024 and
4096×2048 DCI evaluation clips with the same content. For
DCI clips, another two different scenes are tested specially
on super high definition at lower bit rate as a further
discussion. Compared to existed works, more miscellaneous
clips at various resolutions are adopted to search the
regularity in these three algorithms, and the subjective
image quality comparison for super high definition is also
processed. The experiment results within our tested videos
show that JPEG2000 perform best at 4K lower bit rate video
coding. Its advantage on PSNR, subjective quality and
computation for super high definition video makes it
suitable for digital cinema of moderate bit rate.

2. EXPERIMENT DESCRIPTION

2.1. Overview

From the technological point of view, JPEG2000 uses
discrete wavelet transform (DWT) and entropy coding,

Page 2

Size, Name Picture Content Features
Back: common
People: 1
Movement: less
Scene: stable
Back: simple
People: many
Movement: violent
Scene: stable
Back: complicated
People: 2
Movement: less
Scene: shake
Frame
176×144
carphone


96
352×240
football


125
720×576
walk


100
Tab. 1. Standard test sequences with lower resolutions

while H.264 uses motion compensation, discrete cosine
transform (DCT) and entropy coding. In JPEG2000, two
kinds of DWT using 9/7 wavelet filter for lossy coding and
5/3 wavelet filter for lossless coding are existed. In H.264,
the part similar to JPEG2000 is the intra-frame coding, and
the entirety is called the inter-frame coding including the
intra-frame coding algorithm and motion compensation. It is
well-known that motion compensation brings great
enhancement for video coding at the cost of intensive
computation.
In our experiment, the comparison is among JPEG2000,
H.264 intra-frame, inter-frame coding with the criterion of
PSNR for objective test and visual effect for subjective test.

2.2. Evaluation Tools

Reference [2] uses JM v7.3 for H.264 and Jasper v1.700.5
for JPEG2000, while in [3] the corresponding tools are
Kakadu v4.2.1 and JM FRext v2.0. In our test, the latest
versions of some of these tools are used, and another tool is
introduced to improve the experiment.
(1) H.264 intra-frame
In consideration of the compatibility with super high
definition video, JM FRext v2.2 [4] is selected as test tool.
JM supports all the latest features of H.264 with high
coding efficiency, except for its boring executing time. The
QP parameter in JM is very important in controlling the bit
rate of H.264.
(2) H.264 inter-frame
For super high definition inter-frame video coding, the
executing time of JM, which is often used in scientific
research, is unacceptable. Commonly used H.264 software
in practice is X264 [5]. Both of X264 and JM perform
perfectly on coding efficiency, but X264 executes 200-500
times faster than JM. Some front-end software such as
megui [6] is used to configure X264. Through the
configuration in megui’s GUI, the inter-frame structure is:
IBBPBBPBBPBBPBBPBBPBBPBBI. And all H.264 tests
are done with High-Profile at Level 4.1.
(3) JPEG2000
JPEG2000 is usually used for still images. For videos,
it keeps processing every single frame independently, and
we have to compute the average PSNR value. Jasper [7] and
Kakadu [8] can be used for JPEG2000 compression. We
choose Kakadu v5.2.5 because of its higher coding
efficiency than Jasper. We set the parameter Creversible =
no in Kakadu to use nonreversible 9/7 wavelet transform.
Bit rate can be directly set in Kakadu. However in H.264,
we need to adjust the QP value to get the corresponding bit
rate. QP is related to the total bits of compressed data, then
to the rate with the formula below:
totalbits
rate
width height

2.3. Test Sequences

Our test uses two groups of sequences. The former is
standard sequences with lower resolution and different
contents; the latter is a DCI sequence with super high
definition up to 4K.
(1) Lower resolution standard test sequence
The resolutions and typical images of standard
sequences are listed in Tab. 1.
(2) DCI evaluation sequence
This sequence is from DCI StEM MMR3 (Standard
Evaluation Material) with a total length of 17239 frames.
Here shown in Fig. 1 is a 60-frame clip. It has complicated
background and many people, but people’s movement and
scene shake is not violent. In our experiment, this sequence
will be scaled to four resolutions: 512×256, 1024×512,
2048×1024 and 4096×2048.
=
×


Fig. 2. Test results of lower resolution sequences

Fig. 1. One frame from DCI

Page 3

3. RESULTS AND ANALYSIS

3.1. Lower Resolution Standard Test Sequence

The objective test compares the values of PSNR (dB, Y-axis)
according to different bit rates (Bit/Pixel, X-axis). The
results are shown in Fig. 2.
For the carphone sequence, the H.264 inter-frame
coding outperforms JPEG2000 and H.264 intra-frame for 10
dB on average. For the football sequence, H.264 mainly
performs better at higher bit rates, the higher the better. This
conclusion is consistent with [2]. For the walk sequence, the
three curves are near to each other, that is to say, the effect
of inter-frame coding is null, even worse at higher bit rate.
Although the divergence brought by the content difference
is significant, some regularity can still be told. For example,
motion compensation does little contribution to coding
efficiency in videos with
furthermore, the advantage of H.264 inter-frame coding is
decreasing with the increasing image size.
complicated movement;
3.2. DCI Evaluation Sequence

To avoid the divergence resulted from different video
contents, we use a DCI clip (Fig. 1) with four ascending
resolutions to calculate the bit rate-PSNR curve. DCI
requires the 9/7 wavelet transform, but the lossless wavelet
transform needs the 5/3 wavelet filter. So a PSNR curve
with 5/3 wavelet is added when testing the super high
definition sequence. The result is shown in Fig. 3.
For the DCI clip, the advantage of inter-frame is about
5dB on average at 512×256, and begins to decrease with the
increase of resolution. This advantage is nearly lost at 4096
×2048 especially at lower bit rate. In comparison with
JPEG2000, H.264 intra-frame coding shows better PSNR at
higher bit rate, but at lower bit rate, JPEG2000 performs
superiorly with the increase of resolution. This superiority
begins to show up when bit rate is lower than 1 bit/pixel at
2048×1024, while 2 at 4096×2048. The trend can also be
found in [3], but further tests for different contents still need
to be done. As to 9/7 and 5/3 wavelet JPEG2000 coding in
digital cinema, 9/7 wavelet has a higher PSNR of 2 dB on
average in our tested bit rate range.

3.3. Subjective Test

The subjective test is still processed on these three coding
algorithms with all the sequences above. Still image and the
corresponding single frame from video are compared for
reconstructed visual quality. As the higher bit rate
corresponds to less compression, the difference of three
images cannot be distinguished by human sense. Thus, the
subjective test only focuses on lower bit rate.
For lower resolution standard sequences, the images
are compared at the bit rate of 0.5bit/pixel. The
reconstructed image of JPEG2000 is blurred, but H.264
presents clearer image with a better consistency among
pixels, even though there are blocking artifacts.
The same test has been done with the DCI sequence of
four different resolutions. For 512×256 and 1024×512,
the results are very similar to the lower resolution test.
While 2048×1024 and 4096×2048 images have too large
size to view the entire range, we crop two parts from each
whole one to demonstrate the difference, which are shown
in Fig. 4 and 5. The picture on the left side is reconstructed
with H.264 inter-frame coding, the middle one with H.264
intra-frame and the right one with JPEG2000.
Fig. 3. DCI evaluation sequence test result

Fig. 4. Part of 2048×1024 image (0.66bit/pixel)



Fig. 5. Part of 4096×2048 image (0.39bit/pixel)

Page 4

When the sequence is at 2048×1024, H.264 inter-
frame has the subtlest and clearest image, H.264 intra-frame
loses some details and performs poor in texture due to
blocking artifacts, and JPEG2000 also gets some blurring.
The result is different when the resolution goes higher to
4096×2048. H.264 intra shows bad detail, while both
H.264 inter-frame and JPEG2000 give good visual effect.

4. FURTHER TESTS FOR DIGITAL CINEMA

The test results above show that JPEG2000 may perform
well when used in super high definition video coding at
lower bit rates. To give a more comprehensive regularity,
another two clips from DCI StEM MMR3 is tested again at
lower bit rates for super high definition.
As shown in Fig. 6, clip A and clip B are from
different scenes and have different content features. Clip A
is a simple one with only two people and less movement,
while clip B has fast moving stuffs. The same tests of PSNR
comparison has been done to these two clips and the results
are shown in Fig.7.In 2048×1024 test, H.264 inter-frame
remains a little advantage, and this advantage is less
significant when testing the fast-move clip. The result is
totally different when the resolution is high up to
4096×2048, where JPEG2000 has the highest PSNR at
these lower bit rates, especially in the fast-move clip.

5. CONCLUSION

A total number of 11 test sequences have been calculated of
PSNR for objective test and compared of visual effect for
subjective test among JPEG2000, H.264 intra and inter-
frame coding. The results show that at lower resolutions,
H.264 inter do great contributions to the gain of PSNR
during the entire bit rate, but this is getting weak with the
increase of resolution. Especially for the super high
definition of digital cinema, inter-frame coding rarely
improves PSNR at the cost of very intensive computation.
For different resolutions, at higher bit rate, H.264 inter and
intra-frame is better than JPEG2000, but at lower bit rate
JPEG2000 is better, even performs best on various 4K
videos at lower bit rates, especially on fast-move scene.
In the subjective test at lower bit rate, H.264 intra and
inter-frame coding perform better on lower resolution
sequences, and the inter-frame is much better. But on super
high definitions, JPEG2000 and H.264 inter-frame
maintains more details of images than H.264 intra-frame.
The compression efficiency of inter-frame coding is
indeed better at the low to moderate bit rate for general
videos. But digital cinema is different, for the resolution is
extremely high, that is up to 4K and the quality
requirements are more stringent when the viewer is in a
cinema. To take a balance between visual lossless and an
acceptable compression, the bit rate is usually 100-300Mbps
(100M is approximate to 0.47bit/pixel) for 4K resolution
and 75Mbps (1.41bit/piexel) for 2K [3]. When the
resolution and bit rate are at this level, if PSNR, visual
performance and computation cost are taken into account
together, JPEG2000 is suitable for digital cinema.

6. REFRENCES

[1] A. Bilgin, M. W. Marcellin, “JPEG2000 for Digital Cinema,”
ISCAS 2006, Island of Kos, Greece, pp. 3878~3881, May 2006.
[2] Aravind AL et al., “Quality and Complexity Comparison of
H.264 Intra Mode with JPEG2000 and JPEG”, ICIP 2004,
Singapore, pp.525~528, Oct 2004.
[3] M. Smith, J. Villasenor, “Intra-frame JPEG2000 vs. Inter-frame
Compression Comparison: The benefits and trade-offs for very
high quality, high resolution sequences,” SMPTE Technical
Conference and Exhibition 2004, Pasadena, California, USA, pp.
1~9, Oct 2004.
[4] http://iphome.hhi.de/suehring/tml/download.
[5] http://x264.nl.
[6] http://megui.sourceforge.net.
[7] http://www.ece.uvic.ca/~mdadams/jasper.
[8] http://www.kakadusoftware.com.

Fig. 6. Clip A (above) and B (below) from DCI


Fig. 7. Results of further tests