Application of Directionlets in Video Coding
Vladan Velisavljevic
*
, Marta Mrak
†
, Ahmet Kondoz
†
*
Deutsche Telekom Laboratories, Germany; email: vladan.velisavljevic@telekom.de
†I-Lab, Centre for Communication Systems Research, University of Surrey, UK; email: {m.mrak, a.kondoz}@surrey.ac.uk
Keywords: Directional transforms, wavelets, Dirac video
codec.
Abstract
The goal of the presented work is to apply a directionally-
adaptive two-dimensional separable wavelet transform, called
directionlets, in common wavelet-based video coders. So far,
directionlets have been used only in still image coding, where
they have provided a sparser representation of images in the
transform domain and a better compression performance. We
apply directionlets to frames of video sequences in the video
coding scheme based on the Dirac video codec. The novel
method outperforms the traditional non-adaptive method
based on the standard wavelet transform in terms of both the
rate-distortion performance and the visual quality of
reconstructed sequences.
1 Introduction
All modern video communication systems require some form
of video compression. Early video coding standards, like
MPEG-2, do not provide sufficient compression for growing
demands to fit a large number of TV channels in the existing
broadcasting systems. New solutions based on the
H.264 / AVC standard enable a better utilisation of the
available bandwidth using higher compression rates [1]. With
a goal to develop innovative compression solutions that meet
their specific needs, BBC has developed an advanced video
coding system called Dirac [2]. The Dirac coder is based on
the technology that has been used in the state-of-the-art image
and video compression systems. In contrast to other popular
video codecs, Dirac is based on an open-source technology
and, thus, can become influential in a wide research
community.
Similar to the widespread MPEG and H.26x video coding
technologies, Dirac uses motion compensation to enhance the
compression efficiency. The produced frames (intra-coded
and motion-compensated) are spatially transformed using the
wavelet transform (WT), which is also adopted in the still
image coding standards, like JPEG-2000 [12].
One of the challenges in application of the WT in video
coding is how to exploit efficiently the high spatio-temporal
coherence of frames. Recently, an adaptation of the spatial
transform to the motion information has been proposed in [7]
and [11] with a goal to improve the sparsity of the signal
representation and, thus, to achieve a better compression
performance and decoding quality. This, so-called, Motion-
Driven Adaptive Transform (MDAT) is beneficial when intra
blocks exist in the motion-compensated frames.
A similar motivation to provide a sparser signal
representation in still image coding has led to a construction
of directionlets [13], which have been built on top of the
standard WT as an adaptive directional asymmetric transform.
Directionlets are capable of characterizing efficiently oriented
elongated features in images, like edges or contours, by
adaptation of transform directions to locally dominant
directions across the image domain. At the same time, they
preserve separability and the conceptual and computational
simplicity of the standard WT. Moreover, directionlets have
been shown to outperform the standard WT in still image
compression algorithms [14].
This achievement motivates us to propose a novel video
coding method obtained as a combination of the Dirac coder
and directionlets. The standard WT is replaced by
directionlets and such a transform is applied to both types of
video frames in the Dirac codec. The transform directions are
adapted within frames and this adaptation is encoded as side
information together with the other coding information.
However, in spite of this overhead data, the novel video
coding method outperforms the standard Dirac coder in terms
of both the rate-distortion performance and the visual quality
of decoded video sequences.
The outline of the paper is as follows. Section 2 reviews the
basic concepts of the wavelet-based video coding, describes
the architecture of the Dirac video encoder and addresses the
limitations of the standard WT. The basic principle of the
construction of directionlets is described in Section 3. Section
4 introduces the concept of the novel coding algorithm
obtained as a combination of directionlets and the Dirac video
codec, whereas, in Section 5, the results of encoding using the
new method are compared to the results obtained by the
standard Dirac coder. Finally, Section 6 concludes this paper.
2 Wavelet-based video coding
The architecture of the Dirac codec is similar to the
architecture of other wavelet-based video codecs [3], [10] and
also of the standard MPEG and H.26x video coding systems.
The Dirac codec consists of the three basic blocks: motion
compensation, spatial WT with quantisation and entropy
coding, as shown in Figure 1.
The main goal of motion compensation is to exploit temporal
redundancy between frames, whereas the task of the spatial
WT is to provide a compact representation of the frame
content, which commonly have most of the energy located in
lower frequency subbands.
Authorized licensed use limited to: XIDIAN UNIVERSITY. Downloaded on January 12, 2010 at 03:14 from IEEE Xplore. Restrictions apply.